[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// Copyright 2021 The Go Authors. All rights reserved.
|
|
|
|
|
// Use of this source code is governed by a BSD-style
|
|
|
|
|
// license that can be found in the LICENSE file.
|
|
|
|
|
|
|
|
|
|
package noder
|
|
|
|
|
|
|
|
|
|
import (
|
2024-01-09 15:06:15 +00:00
|
|
|
"encoding/hex"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"fmt"
|
|
|
|
|
"go/constant"
|
2021-09-08 15:59:11 -07:00
|
|
|
"internal/buildcfg"
|
2022-02-14 09:41:19 -08:00
|
|
|
"internal/pkgbits"
|
cmd/dist: make toolchain build reproducible
- Build cmd with CGO_ENABLED=0. Doing so removes the C compiler
toolchain from the reproducibility perimeter and also results in
cmd/go and cmd/pprof binaries that are statically linked,
so that they will run on a wider variety of systems.
In particular the Linux versions will run on Alpine and NixOS
without needing a simulation of libc.so.6.
The potential downside of disabling cgo is that cmd/go and cmd/pprof
use the pure Go network resolver instead of the host resolver on
Unix systems. This means they will not be able to use non-DNS
resolver mechanisms that may be specified in /etc/resolv.conf,
such as mDNS. Neither program seems likely to need non-DNS names
like those, however.
macOS and Windows systems still use the host resolver, which they
access without cgo.
- Build cmd with -trimpath when building a release.
Doing so removes $GOPATH from the file name prefixes stored in the
binary, so that the build directory does not leak into the final artifacts.
- When CC and CXX are empty, do not pick values to hard-code into
the source tree and binaries. Instead, emit code that makes the
right decision at runtime. In addition to reproducibility, this
makes cross-compiled toolchains work better. A macOS toolchain
cross-compiled on Linux will now correctly look for clang,
instead of looking for gcc because it was built on Linux.
- Convert \ to / in file names stored in .a files.
These are converted to / in the final binaries, but the hashes of
the .a files affect the final build ID of the binaries. Without this
change, builds of a Windows toolchain on Windows and non-Windows
machines produce identical binaries except for the input hash part
of the build ID.
- Due to the conversion of \ to / in .a files, convert back when
reading inline bodies on Windows to preserve output file names
in error messages.
Combined, these four changes (along with Go 1.20's removal of
installed pkg/**.a files and conversion of macOS net away from cgo)
make the output of make.bash fully reproducible, even when
cross-compiling: a released macOS toolchain built on Linux or Windows
will contain exactly the same bits as a released macOS toolchain
built on macOS.
The word "released" in the previous sentence is important.
For the build IDs in the binaries to work out the same on
both systems, a VERSION file must exist to provide a consistent
compiler build ID (instead of using a content hash of the binary).
For #24904.
Fixes #57007.
Change-Id: I665e1ef4ff207d6ff469452347dca5bfc81050e6
Reviewed-on: https://go-review.googlesource.com/c/go/+/454836
Reviewed-by: Bryan Mills <bcmills@google.com>
Run-TryBot: Russ Cox <rsc@golang.org>
Auto-Submit: Russ Cox <rsc@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-12-02 10:39:30 -05:00
|
|
|
"path/filepath"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"strings"
|
|
|
|
|
|
|
|
|
|
"cmd/compile/internal/base"
|
|
|
|
|
"cmd/compile/internal/dwarfgen"
|
2021-07-03 04:53:25 -07:00
|
|
|
"cmd/compile/internal/inline"
|
2023-11-18 20:16:47 -08:00
|
|
|
"cmd/compile/internal/inline/interleaved"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"cmd/compile/internal/ir"
|
2022-07-31 18:48:16 -07:00
|
|
|
"cmd/compile/internal/objw"
|
2021-06-23 21:33:24 -07:00
|
|
|
"cmd/compile/internal/reflectdata"
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
"cmd/compile/internal/staticinit"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"cmd/compile/internal/typecheck"
|
|
|
|
|
"cmd/compile/internal/types"
|
2024-09-04 18:30:35 +07:00
|
|
|
"cmd/internal/hash"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"cmd/internal/obj"
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
"cmd/internal/objabi"
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
"cmd/internal/src"
|
|
|
|
|
)
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// This file implements cmd/compile backend's reader for the Unified
|
|
|
|
|
// IR export data.
|
|
|
|
|
|
|
|
|
|
// A pkgReader reads Unified IR export data.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
type pkgReader struct {
|
2022-02-14 09:41:19 -08:00
|
|
|
pkgbits.PkgDecoder
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// Indices for encoded things; lazily populated as needed.
|
|
|
|
|
//
|
|
|
|
|
// Note: Objects (i.e., ir.Names) are lazily instantiated by
|
|
|
|
|
// populating their types.Sym.Def; see objReader below.
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
posBases []*src.PosBase
|
|
|
|
|
pkgs []*types.Pkg
|
|
|
|
|
typs []*types.Type
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// offset for rewriting the given (absolute!) index into the output,
|
|
|
|
|
// but bitwise inverted so we can detect if we're missing the entry
|
|
|
|
|
// or not.
|
2024-08-20 10:35:23 -07:00
|
|
|
newindex []index
|
2025-01-10 12:49:59 +07:00
|
|
|
|
|
|
|
|
// indicates whether the data is reading during reshaping.
|
|
|
|
|
reshaping bool
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
func newPkgReader(pr pkgbits.PkgDecoder) *pkgReader {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return &pkgReader{
|
2022-02-14 09:41:19 -08:00
|
|
|
PkgDecoder: pr,
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
posBases: make([]*src.PosBase, pr.NumElems(pkgbits.SectionPosBase)),
|
|
|
|
|
pkgs: make([]*types.Pkg, pr.NumElems(pkgbits.SectionPkg)),
|
|
|
|
|
typs: make([]*types.Type, pr.NumElems(pkgbits.SectionType)),
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2024-08-20 10:35:23 -07:00
|
|
|
newindex: make([]index, pr.TotalElems()),
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// A pkgReaderIndex compactly identifies an index (and its
|
|
|
|
|
// corresponding dictionary) within a package's export data.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
type pkgReaderIndex struct {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
pr *pkgReader
|
2024-08-20 10:35:23 -07:00
|
|
|
idx index
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
dict *readerDict
|
|
|
|
|
methodSym *types.Sym
|
|
|
|
|
|
|
|
|
|
synthetic func(pos src.XPos, r *reader)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2025-05-05 13:40:55 -04:00
|
|
|
func (pri pkgReaderIndex) asReader(k pkgbits.SectionKind, marker pkgbits.SyncMarker) *reader {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if pri.synthetic != nil {
|
|
|
|
|
return &reader{synthetic: pri.synthetic}
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r := pri.pr.newReader(k, pri.idx, marker)
|
2021-06-28 22:41:50 -07:00
|
|
|
r.dict = pri.dict
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
r.methodSym = pri.methodSym
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return r
|
|
|
|
|
}
|
|
|
|
|
|
2025-05-05 13:40:55 -04:00
|
|
|
func (pr *pkgReader) newReader(k pkgbits.SectionKind, idx index, marker pkgbits.SyncMarker) *reader {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return &reader{
|
2022-02-14 09:41:19 -08:00
|
|
|
Decoder: pr.NewDecoder(k, idx, marker),
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
p: pr,
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-08-18 20:59:26 -07:00
|
|
|
// A reader provides APIs for reading an individual element.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
type reader struct {
|
2022-02-14 09:41:19 -08:00
|
|
|
pkgbits.Decoder
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
p *pkgReader
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
dict *readerDict
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// TODO(mdempsky): The state below is all specific to reading
|
|
|
|
|
// function bodies. It probably makes sense to split it out
|
|
|
|
|
// separately so that it doesn't take up space in every reader
|
|
|
|
|
// instance.
|
|
|
|
|
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
curfn *ir.Func
|
|
|
|
|
locals []*ir.Name
|
|
|
|
|
closureVars []*ir.Name
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
// funarghack is used during inlining to suppress setting
|
|
|
|
|
// Field.Nname to the inlined copies of the parameters. This is
|
|
|
|
|
// necessary because we reuse the same types.Type as the original
|
|
|
|
|
// function, and most of the compiler still relies on field.Nname to
|
|
|
|
|
// find parameters/results.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
funarghack bool
|
|
|
|
|
|
2025-01-10 12:49:59 +07:00
|
|
|
// reshaping is used during reading exprReshape code, preventing
|
|
|
|
|
// the reader from shapifying the re-shaped type.
|
|
|
|
|
reshaping bool
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// methodSym is the name of method's name, if reading a method.
|
|
|
|
|
// It's nil if reading a normal function or closure body.
|
|
|
|
|
methodSym *types.Sym
|
2022-07-31 18:48:16 -07:00
|
|
|
|
|
|
|
|
// dictParam is the .dict param, if any.
|
|
|
|
|
dictParam *ir.Name
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// synthetic is a callback function to construct a synthetic
|
|
|
|
|
// function body. It's used for creating the bodies of function
|
|
|
|
|
// literals used to curry arguments to shaped functions.
|
|
|
|
|
synthetic func(pos src.XPos, r *reader)
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// scopeVars is a stack tracking the number of variables declared in
|
|
|
|
|
// the current function at the moment each open scope was opened.
|
|
|
|
|
scopeVars []int
|
|
|
|
|
marker dwarfgen.ScopeMarker
|
|
|
|
|
lastCloseScopePos src.XPos
|
|
|
|
|
|
|
|
|
|
// === details for handling inline body expansion ===
|
|
|
|
|
|
|
|
|
|
// If we're reading in a function body because of inlining, this is
|
|
|
|
|
// the call that we're inlining for.
|
|
|
|
|
inlCaller *ir.Func
|
|
|
|
|
inlCall *ir.CallExpr
|
|
|
|
|
inlFunc *ir.Func
|
|
|
|
|
inlTreeIndex int
|
|
|
|
|
inlPosBases map[*src.PosBase]*src.PosBase
|
|
|
|
|
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
// suppressInlPos tracks whether position base rewriting for
|
|
|
|
|
// inlining should be suppressed. See funcLit.
|
|
|
|
|
suppressInlPos int
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
delayResults bool
|
|
|
|
|
|
|
|
|
|
// Label to return to.
|
|
|
|
|
retlabel *types.Sym
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// A readerDict represents an instantiated "compile-time dictionary,"
|
|
|
|
|
// used for resolving any derived types needed for instantiating a
|
|
|
|
|
// generic object.
|
|
|
|
|
//
|
|
|
|
|
// A compile-time dictionary can either be "shaped" or "non-shaped."
|
|
|
|
|
// Shaped compile-time dictionaries are only used for instantiating
|
|
|
|
|
// shaped type definitions and function bodies, while non-shaped
|
|
|
|
|
// compile-time dictionaries are used for instantiating runtime
|
|
|
|
|
// dictionaries.
|
2021-06-28 22:41:50 -07:00
|
|
|
type readerDict struct {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
shaped bool // whether this is a shaped dictionary
|
|
|
|
|
|
|
|
|
|
// baseSym is the symbol for the object this dictionary belongs to.
|
|
|
|
|
// If the object is an instantiated function or defined type, then
|
|
|
|
|
// baseSym is the mangled symbol, including any type arguments.
|
|
|
|
|
baseSym *types.Sym
|
|
|
|
|
|
|
|
|
|
// For non-shaped dictionaries, shapedObj is a reference to the
|
|
|
|
|
// corresponding shaped object (always a function or defined type).
|
|
|
|
|
shapedObj *ir.Name
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
// targs holds the implicit and explicit type arguments in use for
|
|
|
|
|
// reading the current object. For example:
|
|
|
|
|
//
|
|
|
|
|
// func F[T any]() {
|
|
|
|
|
// type X[U any] struct { t T; u U }
|
|
|
|
|
// var _ X[string]
|
|
|
|
|
// }
|
|
|
|
|
//
|
|
|
|
|
// var _ = F[int]
|
|
|
|
|
//
|
|
|
|
|
// While instantiating F[int], we need to in turn instantiate
|
|
|
|
|
// X[string]. [int] and [string] are explicit type arguments for F
|
|
|
|
|
// and X, respectively; but [int] is also the implicit type
|
|
|
|
|
// arguments for X.
|
|
|
|
|
//
|
|
|
|
|
// (As an analogy to function literals, explicits are the function
|
|
|
|
|
// literal's formal parameters, while implicits are variables
|
|
|
|
|
// captured by the function literal.)
|
|
|
|
|
targs []*types.Type
|
|
|
|
|
|
|
|
|
|
// implicits counts how many of types within targs are implicit type
|
|
|
|
|
// arguments; the rest are explicit.
|
|
|
|
|
implicits int
|
|
|
|
|
|
2021-07-13 09:09:32 -07:00
|
|
|
derived []derivedInfo // reloc index of the derived type's descriptor
|
|
|
|
|
derivedTypes []*types.Type // slice of previously computed derived types
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// These slices correspond to entries in the runtime dictionary.
|
|
|
|
|
typeParamMethodExprs []readerMethodExprInfo
|
|
|
|
|
subdicts []objInfo
|
|
|
|
|
rtypes []typeInfo
|
|
|
|
|
itabs []itabInfo
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type readerMethodExprInfo struct {
|
|
|
|
|
typeParamIdx int
|
|
|
|
|
method *types.Sym
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func setType(n ir.Node, typ *types.Type) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
n.SetType(typ)
|
|
|
|
|
n.SetTypecheck(1)
|
|
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func setValue(name *ir.Name, val constant.Value) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name.SetVal(val)
|
|
|
|
|
name.Defn = nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Positions
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// pos reads a position from the bitstream.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) pos() src.XPos {
|
|
|
|
|
return base.Ctxt.PosTable.XPos(r.pos0())
|
|
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
// origPos reads a position from the bitstream, and returns both the
|
|
|
|
|
// original raw position and an inlining-adjusted position.
|
|
|
|
|
func (r *reader) origPos() (origPos, inlPos src.XPos) {
|
|
|
|
|
r.suppressInlPos++
|
|
|
|
|
origPos = r.pos()
|
|
|
|
|
r.suppressInlPos--
|
|
|
|
|
inlPos = r.inlPos(origPos)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) pos0() src.Pos {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncPos)
|
|
|
|
|
if !r.Bool() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return src.NoPos
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
posBase := r.posBase()
|
2022-02-14 09:41:19 -08:00
|
|
|
line := r.Uint()
|
|
|
|
|
col := r.Uint()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return src.MakePos(posBase, line, col)
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// posBase reads a position base from the bitstream.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) posBase() *src.PosBase {
|
2025-05-12 12:21:03 -04:00
|
|
|
return r.inlPosBase(r.p.posBaseIdx(r.Reloc(pkgbits.SectionPosBase)))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// posBaseIdx returns the specified position base, reading it first if
|
|
|
|
|
// needed.
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) posBaseIdx(idx index) *src.PosBase {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if b := pr.posBases[idx]; b != nil {
|
|
|
|
|
return b
|
|
|
|
|
}
|
|
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pr.newReader(pkgbits.SectionPosBase, idx, pkgbits.SyncPosBase)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var b *src.PosBase
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
absFilename := r.String()
|
2021-09-08 15:59:11 -07:00
|
|
|
filename := absFilename
|
|
|
|
|
|
|
|
|
|
// For build artifact stability, the export data format only
|
|
|
|
|
// contains the "absolute" filename as returned by objabi.AbsFile.
|
|
|
|
|
// However, some tests (e.g., test/run.go's asmcheck tests) expect
|
|
|
|
|
// to see the full, original filename printed out. Re-expanding
|
|
|
|
|
// "$GOROOT" to buildcfg.GOROOT is a close-enough approximation to
|
|
|
|
|
// satisfy this.
|
|
|
|
|
//
|
cmd/dist: make toolchain build reproducible
- Build cmd with CGO_ENABLED=0. Doing so removes the C compiler
toolchain from the reproducibility perimeter and also results in
cmd/go and cmd/pprof binaries that are statically linked,
so that they will run on a wider variety of systems.
In particular the Linux versions will run on Alpine and NixOS
without needing a simulation of libc.so.6.
The potential downside of disabling cgo is that cmd/go and cmd/pprof
use the pure Go network resolver instead of the host resolver on
Unix systems. This means they will not be able to use non-DNS
resolver mechanisms that may be specified in /etc/resolv.conf,
such as mDNS. Neither program seems likely to need non-DNS names
like those, however.
macOS and Windows systems still use the host resolver, which they
access without cgo.
- Build cmd with -trimpath when building a release.
Doing so removes $GOPATH from the file name prefixes stored in the
binary, so that the build directory does not leak into the final artifacts.
- When CC and CXX are empty, do not pick values to hard-code into
the source tree and binaries. Instead, emit code that makes the
right decision at runtime. In addition to reproducibility, this
makes cross-compiled toolchains work better. A macOS toolchain
cross-compiled on Linux will now correctly look for clang,
instead of looking for gcc because it was built on Linux.
- Convert \ to / in file names stored in .a files.
These are converted to / in the final binaries, but the hashes of
the .a files affect the final build ID of the binaries. Without this
change, builds of a Windows toolchain on Windows and non-Windows
machines produce identical binaries except for the input hash part
of the build ID.
- Due to the conversion of \ to / in .a files, convert back when
reading inline bodies on Windows to preserve output file names
in error messages.
Combined, these four changes (along with Go 1.20's removal of
installed pkg/**.a files and conversion of macOS net away from cgo)
make the output of make.bash fully reproducible, even when
cross-compiling: a released macOS toolchain built on Linux or Windows
will contain exactly the same bits as a released macOS toolchain
built on macOS.
The word "released" in the previous sentence is important.
For the build IDs in the binaries to work out the same on
both systems, a VERSION file must exist to provide a consistent
compiler build ID (instead of using a content hash of the binary).
For #24904.
Fixes #57007.
Change-Id: I665e1ef4ff207d6ff469452347dca5bfc81050e6
Reviewed-on: https://go-review.googlesource.com/c/go/+/454836
Reviewed-by: Bryan Mills <bcmills@google.com>
Run-TryBot: Russ Cox <rsc@golang.org>
Auto-Submit: Russ Cox <rsc@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-12-02 10:39:30 -05:00
|
|
|
// The export data format only ever uses slash paths
|
|
|
|
|
// (for cross-operating-system reproducible builds),
|
|
|
|
|
// but error messages need to use native paths (backslash on Windows)
|
|
|
|
|
// as if they had been specified on the command line.
|
|
|
|
|
// (The go command always passes native paths to the compiler.)
|
2021-09-08 15:59:11 -07:00
|
|
|
const dollarGOROOT = "$GOROOT"
|
2022-03-15 16:31:02 -04:00
|
|
|
if buildcfg.GOROOT != "" && strings.HasPrefix(filename, dollarGOROOT) {
|
cmd/dist: make toolchain build reproducible
- Build cmd with CGO_ENABLED=0. Doing so removes the C compiler
toolchain from the reproducibility perimeter and also results in
cmd/go and cmd/pprof binaries that are statically linked,
so that they will run on a wider variety of systems.
In particular the Linux versions will run on Alpine and NixOS
without needing a simulation of libc.so.6.
The potential downside of disabling cgo is that cmd/go and cmd/pprof
use the pure Go network resolver instead of the host resolver on
Unix systems. This means they will not be able to use non-DNS
resolver mechanisms that may be specified in /etc/resolv.conf,
such as mDNS. Neither program seems likely to need non-DNS names
like those, however.
macOS and Windows systems still use the host resolver, which they
access without cgo.
- Build cmd with -trimpath when building a release.
Doing so removes $GOPATH from the file name prefixes stored in the
binary, so that the build directory does not leak into the final artifacts.
- When CC and CXX are empty, do not pick values to hard-code into
the source tree and binaries. Instead, emit code that makes the
right decision at runtime. In addition to reproducibility, this
makes cross-compiled toolchains work better. A macOS toolchain
cross-compiled on Linux will now correctly look for clang,
instead of looking for gcc because it was built on Linux.
- Convert \ to / in file names stored in .a files.
These are converted to / in the final binaries, but the hashes of
the .a files affect the final build ID of the binaries. Without this
change, builds of a Windows toolchain on Windows and non-Windows
machines produce identical binaries except for the input hash part
of the build ID.
- Due to the conversion of \ to / in .a files, convert back when
reading inline bodies on Windows to preserve output file names
in error messages.
Combined, these four changes (along with Go 1.20's removal of
installed pkg/**.a files and conversion of macOS net away from cgo)
make the output of make.bash fully reproducible, even when
cross-compiling: a released macOS toolchain built on Linux or Windows
will contain exactly the same bits as a released macOS toolchain
built on macOS.
The word "released" in the previous sentence is important.
For the build IDs in the binaries to work out the same on
both systems, a VERSION file must exist to provide a consistent
compiler build ID (instead of using a content hash of the binary).
For #24904.
Fixes #57007.
Change-Id: I665e1ef4ff207d6ff469452347dca5bfc81050e6
Reviewed-on: https://go-review.googlesource.com/c/go/+/454836
Reviewed-by: Bryan Mills <bcmills@google.com>
Run-TryBot: Russ Cox <rsc@golang.org>
Auto-Submit: Russ Cox <rsc@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-12-02 10:39:30 -05:00
|
|
|
filename = filepath.FromSlash(buildcfg.GOROOT + filename[len(dollarGOROOT):])
|
2021-09-08 15:59:11 -07:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
2021-09-08 15:59:11 -07:00
|
|
|
b = src.NewFileBase(filename, absFilename)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
} else {
|
|
|
|
|
pos := r.pos0()
|
2022-02-14 09:41:19 -08:00
|
|
|
line := r.Uint()
|
|
|
|
|
col := r.Uint()
|
2021-09-08 15:59:11 -07:00
|
|
|
b = src.NewLinePragmaBase(pos, filename, absFilename, line, col)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pr.posBases[idx] = b
|
|
|
|
|
return b
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
// inlPosBase returns the inlining-adjusted src.PosBase corresponding
|
|
|
|
|
// to oldBase, which must be a non-inlined position. When not
|
|
|
|
|
// inlining, this is just oldBase.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) inlPosBase(oldBase *src.PosBase) *src.PosBase {
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
if index := oldBase.InliningIndex(); index >= 0 {
|
|
|
|
|
base.Fatalf("oldBase %v already has inlining index %v", oldBase, index)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if r.inlCall == nil || r.suppressInlPos != 0 {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return oldBase
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if newBase, ok := r.inlPosBases[oldBase]; ok {
|
|
|
|
|
return newBase
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
newBase := src.NewInliningBase(oldBase, r.inlTreeIndex)
|
|
|
|
|
r.inlPosBases[oldBase] = newBase
|
|
|
|
|
return newBase
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
// inlPos returns the inlining-adjusted src.XPos corresponding to
|
|
|
|
|
// xpos, which must be a non-inlined position. When not inlining, this
|
|
|
|
|
// is just xpos.
|
|
|
|
|
func (r *reader) inlPos(xpos src.XPos) src.XPos {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := base.Ctxt.PosTable.Pos(xpos)
|
|
|
|
|
pos.SetBase(r.inlPosBase(pos.Base()))
|
|
|
|
|
return base.Ctxt.PosTable.XPos(pos)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Packages
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// pkg reads a package reference from the bitstream.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) pkg() *types.Pkg {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncPkg)
|
2025-05-12 12:21:03 -04:00
|
|
|
return r.p.pkgIdx(r.Reloc(pkgbits.SectionPkg))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// pkgIdx returns the specified package from the export data, reading
|
|
|
|
|
// it first if needed.
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) pkgIdx(idx index) *types.Pkg {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if pkg := pr.pkgs[idx]; pkg != nil {
|
|
|
|
|
return pkg
|
|
|
|
|
}
|
|
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
pkg := pr.newReader(pkgbits.SectionPkg, idx, pkgbits.SyncPkgDef).doPkg()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pr.pkgs[idx] = pkg
|
|
|
|
|
return pkg
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// doPkg reads a package definition from the bitstream.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) doPkg() *types.Pkg {
|
2022-02-14 09:41:19 -08:00
|
|
|
path := r.String()
|
2022-04-29 10:47:57 -07:00
|
|
|
switch path {
|
|
|
|
|
case "":
|
2022-02-14 09:41:19 -08:00
|
|
|
path = r.p.PkgPath()
|
2022-04-29 10:47:57 -07:00
|
|
|
case "builtin":
|
|
|
|
|
return types.BuiltinPkg
|
|
|
|
|
case "unsafe":
|
|
|
|
|
return types.UnsafePkg
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
name := r.String()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
pkg := types.NewPkg(path, "")
|
|
|
|
|
|
|
|
|
|
if pkg.Name == "" {
|
|
|
|
|
pkg.Name = name
|
|
|
|
|
} else {
|
2022-05-12 15:45:13 -07:00
|
|
|
base.Assertf(pkg.Name == name, "package %q has name %q, but want %q", pkg.Path, pkg.Name, name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return pkg
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Types
|
|
|
|
|
|
|
|
|
|
func (r *reader) typ() *types.Type {
|
2021-08-30 01:01:20 -07:00
|
|
|
return r.typWrapped(true)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// typWrapped is like typ, but allows suppressing generation of
|
|
|
|
|
// unnecessary wrappers as a compile-time optimization.
|
|
|
|
|
func (r *reader) typWrapped(wrapped bool) *types.Type {
|
|
|
|
|
return r.p.typIdx(r.typInfo(), r.dict, wrapped)
|
2021-07-13 09:09:32 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) typInfo() typeInfo {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncType)
|
|
|
|
|
if r.Bool() {
|
2024-08-20 10:35:23 -07:00
|
|
|
return typeInfo{idx: index(r.Len()), derived: true}
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
2025-05-12 12:21:03 -04:00
|
|
|
return typeInfo{idx: r.Reloc(pkgbits.SectionType), derived: false}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// typListIdx returns a list of the specified types, resolving derived
|
|
|
|
|
// types within the given dictionary.
|
|
|
|
|
func (pr *pkgReader) typListIdx(infos []typeInfo, dict *readerDict) []*types.Type {
|
|
|
|
|
typs := make([]*types.Type, len(infos))
|
|
|
|
|
for i, info := range infos {
|
|
|
|
|
typs[i] = pr.typIdx(info, dict, true)
|
|
|
|
|
}
|
|
|
|
|
return typs
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// typIdx returns the specified type. If info specifies a derived
|
|
|
|
|
// type, it's resolved within the given dictionary. If wrapped is
|
|
|
|
|
// true, then method wrappers will be generated, if appropriate.
|
2021-08-30 01:01:20 -07:00
|
|
|
func (pr *pkgReader) typIdx(info typeInfo, dict *readerDict, wrapped bool) *types.Type {
|
2021-07-13 09:09:32 -07:00
|
|
|
idx := info.idx
|
2021-06-28 22:41:50 -07:00
|
|
|
var where **types.Type
|
2021-07-13 09:09:32 -07:00
|
|
|
if info.derived {
|
|
|
|
|
where = &dict.derivedTypes[idx]
|
|
|
|
|
idx = dict.derived[idx].idx
|
2021-06-28 22:41:50 -07:00
|
|
|
} else {
|
|
|
|
|
where = &pr.typs[idx]
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if typ := *where; typ != nil {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return typ
|
|
|
|
|
}
|
|
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pr.newReader(pkgbits.SectionType, idx, pkgbits.SyncTypeIdx)
|
2021-06-28 22:41:50 -07:00
|
|
|
r.dict = dict
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
typ := r.doTyp()
|
2024-05-14 22:22:34 -07:00
|
|
|
if typ == nil {
|
|
|
|
|
base.Fatalf("doTyp returned nil for info=%v", info)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-20 02:06:45 +07:00
|
|
|
// For recursive type declarations involving interfaces and aliases,
|
|
|
|
|
// above r.doTyp() call may have already set pr.typs[idx], so just
|
|
|
|
|
// double check and return the type.
|
|
|
|
|
//
|
|
|
|
|
// Example:
|
|
|
|
|
//
|
|
|
|
|
// type F = func(I)
|
|
|
|
|
//
|
|
|
|
|
// type I interface {
|
|
|
|
|
// m(F)
|
|
|
|
|
// }
|
|
|
|
|
//
|
|
|
|
|
// The writer writes data types in following index order:
|
|
|
|
|
//
|
|
|
|
|
// 0: func(I)
|
|
|
|
|
// 1: I
|
|
|
|
|
// 2: interface{m(func(I))}
|
|
|
|
|
//
|
|
|
|
|
// The reader resolves it in following index order:
|
|
|
|
|
//
|
|
|
|
|
// 0 -> 1 -> 2 -> 0 -> 1
|
|
|
|
|
//
|
|
|
|
|
// and can divide in logically 2 steps:
|
|
|
|
|
//
|
|
|
|
|
// - 0 -> 1 : first time the reader reach type I,
|
|
|
|
|
// it creates new named type with symbol I.
|
|
|
|
|
//
|
|
|
|
|
// - 2 -> 0 -> 1: the reader ends up reaching symbol I again,
|
|
|
|
|
// now the symbol I was setup in above step, so
|
|
|
|
|
// the reader just return the named type.
|
|
|
|
|
//
|
|
|
|
|
// Now, the functions called return, the pr.typs looks like below:
|
|
|
|
|
//
|
|
|
|
|
// - 0 -> 1 -> 2 -> 0 : [<T> I <T>]
|
|
|
|
|
// - 0 -> 1 -> 2 : [func(I) I <T>]
|
|
|
|
|
// - 0 -> 1 : [func(I) I interface { "".m(func("".I)) }]
|
|
|
|
|
//
|
|
|
|
|
// The idx 1, corresponding with type I was resolved successfully
|
|
|
|
|
// after r.doTyp() call.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
if prev := *where; prev != nil {
|
|
|
|
|
return prev
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2021-08-30 01:01:20 -07:00
|
|
|
if wrapped {
|
|
|
|
|
// Only cache if we're adding wrappers, so that other callers that
|
|
|
|
|
// find a cached type know it was wrapped.
|
|
|
|
|
*where = typ
|
|
|
|
|
|
|
|
|
|
r.needWrapper(typ)
|
|
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if !typ.IsUntyped() {
|
|
|
|
|
types.CheckSize(typ)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return typ
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) doTyp() *types.Type {
|
2022-02-14 09:41:19 -08:00
|
|
|
switch tag := pkgbits.CodeType(r.Code(pkgbits.SyncType)); tag {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
default:
|
|
|
|
|
panic(fmt.Sprintf("unexpected type: %v", tag))
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeBasic:
|
|
|
|
|
return *basics[r.Len()]
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeNamed:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
obj := r.obj()
|
|
|
|
|
assert(obj.Op() == ir.OTYPE)
|
|
|
|
|
return obj.Type()
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeTypeParam:
|
|
|
|
|
return r.dict.targs[r.Len()]
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeArray:
|
|
|
|
|
len := int64(r.Uint64())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return types.NewArray(r.typ(), len)
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeChan:
|
|
|
|
|
dir := dirs[r.Len()]
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return types.NewChan(r.typ(), dir)
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeMap:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return types.NewMap(r.typ(), r.typ())
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypePointer:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return types.NewPtr(r.typ())
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeSignature:
|
2022-12-01 17:42:02 -08:00
|
|
|
return r.signature(nil)
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeSlice:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return types.NewSlice(r.typ())
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeStruct:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return r.structType()
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.TypeInterface:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return r.interfaceType()
|
2022-04-04 14:07:36 +07:00
|
|
|
case pkgbits.TypeUnion:
|
|
|
|
|
return r.unionType()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-04-04 14:07:36 +07:00
|
|
|
func (r *reader) unionType() *types.Type {
|
2022-12-01 17:24:23 -08:00
|
|
|
// In the types1 universe, we only need to handle value types.
|
|
|
|
|
// Impure interfaces (i.e., interfaces with non-trivial type sets
|
|
|
|
|
// like "int | string") can only appear as type parameter bounds,
|
|
|
|
|
// and this is enforced by the types2 type checker.
|
|
|
|
|
//
|
|
|
|
|
// However, type unions can still appear in pure interfaces if the
|
|
|
|
|
// type union is equivalent to "any". E.g., typeparam/issue52124.go
|
|
|
|
|
// declares variables with the type "interface { any | int }".
|
|
|
|
|
//
|
|
|
|
|
// To avoid needing to represent type unions in types1 (since we
|
|
|
|
|
// don't have any uses for that today anyway), we simply fold them
|
2023-05-16 12:46:33 -07:00
|
|
|
// to "any".
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Restore consistency check to make sure folding to
|
|
|
|
|
// "any" is safe. This is unfortunately tricky, because a pure
|
|
|
|
|
// interface can reference impure interfaces too, including
|
|
|
|
|
// cyclically (#60117).
|
|
|
|
|
if false {
|
|
|
|
|
pure := false
|
|
|
|
|
for i, n := 0, r.Len(); i < n; i++ {
|
|
|
|
|
_ = r.Bool() // tilde
|
|
|
|
|
term := r.typ()
|
|
|
|
|
if term.IsEmptyInterface() {
|
|
|
|
|
pure = true
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if !pure {
|
|
|
|
|
base.Fatalf("impure type set used in value type")
|
2022-12-01 17:24:23 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return types.Types[types.TINTER]
|
2022-04-04 14:07:36 +07:00
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) interfaceType() *types.Type {
|
2022-02-14 09:41:19 -08:00
|
|
|
nmethods, nembeddeds := r.Len(), r.Len()
|
2022-02-01 17:41:46 -08:00
|
|
|
implicit := nmethods == 0 && nembeddeds == 1 && r.Bool()
|
|
|
|
|
assert(!implicit) // implicit interfaces only appear in constraints
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
fields := make([]*types.Field, nmethods+nembeddeds)
|
|
|
|
|
methods, embeddeds := fields[:nmethods], fields[nmethods:]
|
|
|
|
|
|
|
|
|
|
for i := range methods {
|
2023-09-13 20:01:56 -07:00
|
|
|
methods[i] = types.NewField(r.pos(), r.selector(), r.signature(types.FakeRecv()))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
for i := range embeddeds {
|
|
|
|
|
embeddeds[i] = types.NewField(src.NoXPos, nil, r.typ())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if len(fields) == 0 {
|
|
|
|
|
return types.Types[types.TINTER] // empty interface
|
|
|
|
|
}
|
2022-12-01 17:42:02 -08:00
|
|
|
return types.NewInterface(fields)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) structType() *types.Type {
|
2022-02-14 09:41:19 -08:00
|
|
|
fields := make([]*types.Field, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range fields {
|
2023-09-13 20:01:56 -07:00
|
|
|
field := types.NewField(r.pos(), r.selector(), r.typ())
|
|
|
|
|
field.Note = r.String()
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
field.Embedded = 1
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-09-13 20:01:56 -07:00
|
|
|
fields[i] = field
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2022-12-01 17:42:02 -08:00
|
|
|
return types.NewStruct(fields)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
func (r *reader) signature(recv *types.Field) *types.Type {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncSignature)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
params := r.params()
|
|
|
|
|
results := r.params()
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() { // variadic
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
params[len(params)-1].SetIsDDD(true)
|
|
|
|
|
}
|
|
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
return types.NewSignature(recv, params, results)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
func (r *reader) params() []*types.Field {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncParams)
|
2023-09-13 20:01:56 -07:00
|
|
|
params := make([]*types.Field, r.Len())
|
|
|
|
|
for i := range params {
|
|
|
|
|
params[i] = r.param()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-09-13 20:01:56 -07:00
|
|
|
return params
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-09-13 20:01:56 -07:00
|
|
|
func (r *reader) param() *types.Field {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncParam)
|
2023-09-13 20:01:56 -07:00
|
|
|
return types.NewField(r.pos(), r.localIdent(), r.typ())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Objects
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// objReader maps qualified identifiers (represented as *types.Sym) to
|
|
|
|
|
// a pkgReader and corresponding index that can be used for reading
|
|
|
|
|
// that object's definition.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var objReader = map[*types.Sym]pkgReaderIndex{}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// obj reads an instantiated object reference from the bitstream.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) obj() ir.Node {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
return r.p.objInstIdx(r.objInfo(), r.dict, false)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// objInfo reads an instantiated object reference from the bitstream
|
|
|
|
|
// and returns the encoded reference to it, without instantiating it.
|
|
|
|
|
func (r *reader) objInfo() objInfo {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncObject)
|
2024-08-15 13:57:53 -07:00
|
|
|
if r.Version().Has(pkgbits.DerivedFuncInstance) {
|
|
|
|
|
assert(!r.Bool())
|
|
|
|
|
}
|
2025-05-12 12:21:03 -04:00
|
|
|
idx := r.Reloc(pkgbits.SectionObj)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
explicits := make([]typeInfo, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range explicits {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
explicits[i] = r.typInfo()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
return objInfo{idx, explicits}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// objInstIdx returns the encoded, instantiated object. If shaped is
|
|
|
|
|
// true, then the shaped variant of the object is returned instead.
|
|
|
|
|
func (pr *pkgReader) objInstIdx(info objInfo, dict *readerDict, shaped bool) ir.Node {
|
|
|
|
|
explicits := pr.typListIdx(info.explicits, dict)
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
var implicits []*types.Type
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if dict != nil {
|
|
|
|
|
implicits = dict.targs
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
return pr.objIdx(info.idx, implicits, explicits, shaped)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// objIdx returns the specified object, instantiated with the given
|
2024-02-08 13:07:33 -05:00
|
|
|
// type arguments, if any.
|
|
|
|
|
// If shaped is true, then the shaped variant of the object is returned
|
|
|
|
|
// instead.
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) objIdx(idx index, implicits, explicits []*types.Type, shaped bool) ir.Node {
|
2024-02-08 13:07:33 -05:00
|
|
|
n, err := pr.objIdxMayFail(idx, implicits, explicits, shaped)
|
|
|
|
|
if err != nil {
|
|
|
|
|
base.Fatalf("%v", err)
|
|
|
|
|
}
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// objIdxMayFail is equivalent to objIdx, but returns an error rather than
|
|
|
|
|
// failing the build if this object requires type arguments and the incorrect
|
|
|
|
|
// number of type arguments were passed.
|
|
|
|
|
//
|
|
|
|
|
// Other sources of internal failure (such as duplicate definitions) still fail
|
|
|
|
|
// the build.
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) objIdxMayFail(idx index, implicits, explicits []*types.Type, shaped bool) (ir.Node, error) {
|
2025-05-12 12:21:03 -04:00
|
|
|
rname := pr.newReader(pkgbits.SectionName, idx, pkgbits.SyncObject1)
|
2021-07-20 16:52:37 -07:00
|
|
|
_, sym := rname.qualifiedIdent()
|
2022-02-14 09:41:19 -08:00
|
|
|
tag := pkgbits.CodeObj(rname.Code(pkgbits.SyncCodeObj))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if tag == pkgbits.ObjStub {
|
2021-07-13 13:25:16 -07:00
|
|
|
assert(!sym.IsBlank())
|
|
|
|
|
switch sym.Pkg {
|
2021-08-26 13:17:56 -07:00
|
|
|
case types.BuiltinPkg, types.UnsafePkg:
|
2024-02-08 13:07:33 -05:00
|
|
|
return sym.Def.(ir.Node), nil
|
2021-07-13 13:25:16 -07:00
|
|
|
}
|
|
|
|
|
if pri, ok := objReader[sym]; ok {
|
2024-02-08 13:07:33 -05:00
|
|
|
return pri.pr.objIdxMayFail(pri.idx, nil, explicits, shaped)
|
2021-07-13 13:25:16 -07:00
|
|
|
}
|
2023-06-14 10:56:49 -04:00
|
|
|
if sym.Pkg.Path == "runtime" {
|
2024-02-08 13:07:33 -05:00
|
|
|
return typecheck.LookupRuntime(sym.Name), nil
|
2023-06-14 10:56:49 -04:00
|
|
|
}
|
2021-07-13 13:25:16 -07:00
|
|
|
base.Fatalf("unresolved stub: %v", sym)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2024-02-08 13:07:33 -05:00
|
|
|
dict, err := pr.objDictIdx(sym, idx, implicits, explicits, shaped)
|
|
|
|
|
if err != nil {
|
|
|
|
|
return nil, err
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
sym = dict.baseSym
|
|
|
|
|
if !sym.IsBlank() && sym.Def != nil {
|
2024-02-08 13:07:33 -05:00
|
|
|
return sym.Def.(*ir.Name), nil
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
2021-07-20 16:52:37 -07:00
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pr.newReader(pkgbits.SectionObj, idx, pkgbits.SyncObject1)
|
|
|
|
|
rext := pr.newReader(pkgbits.SectionObjExt, idx, pkgbits.SyncObject1)
|
2021-07-20 16:52:37 -07:00
|
|
|
|
|
|
|
|
r.dict = dict
|
2021-09-07 17:24:50 -07:00
|
|
|
rext.dict = dict
|
2021-06-28 22:41:50 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
do := func(op ir.Op, hasTParams bool) *ir.Name {
|
|
|
|
|
pos := r.pos()
|
2022-05-25 23:27:45 +07:00
|
|
|
setBasePos(pos)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if hasTParams {
|
|
|
|
|
r.typeParamNames()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
name := ir.NewDeclNameAt(pos, op, sym)
|
|
|
|
|
name.Class = ir.PEXTERN // may be overridden later
|
|
|
|
|
if !sym.IsBlank() {
|
|
|
|
|
if sym.Def != nil {
|
|
|
|
|
base.FatalfAt(name.Pos(), "already have a definition for %v", name)
|
|
|
|
|
}
|
|
|
|
|
assert(sym.Def == nil)
|
|
|
|
|
sym.Def = name
|
|
|
|
|
}
|
|
|
|
|
return name
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch tag {
|
|
|
|
|
default:
|
|
|
|
|
panic("unexpected object")
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.ObjAlias:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name := do(ir.OTYPE, false)
|
2024-05-14 22:22:34 -07:00
|
|
|
|
2024-08-20 12:43:59 -07:00
|
|
|
if r.Version().Has(pkgbits.AliasTypeParamNames) {
|
|
|
|
|
r.typeParamNames()
|
|
|
|
|
}
|
|
|
|
|
|
2024-05-14 22:22:34 -07:00
|
|
|
// Clumsy dance: the r.typ() call here might recursively find this
|
|
|
|
|
// type alias name, before we've set its type (#66873). So we
|
|
|
|
|
// temporarily clear sym.Def and then restore it later, if still
|
|
|
|
|
// unset.
|
|
|
|
|
hack := sym.Def == name
|
|
|
|
|
if hack {
|
|
|
|
|
sym.Def = nil
|
|
|
|
|
}
|
|
|
|
|
typ := r.typ()
|
|
|
|
|
if hack {
|
|
|
|
|
if sym.Def != nil {
|
|
|
|
|
name = sym.Def.(*ir.Name)
|
2025-04-25 19:13:52 +07:00
|
|
|
assert(types.IdenticalStrict(name.Type(), typ))
|
2024-05-14 22:22:34 -07:00
|
|
|
return name, nil
|
|
|
|
|
}
|
|
|
|
|
sym.Def = name
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
setType(name, typ)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name.SetAlias(true)
|
2024-02-08 13:07:33 -05:00
|
|
|
return name, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.ObjConst:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name := do(ir.OLITERAL, false)
|
2021-09-07 13:23:08 -07:00
|
|
|
typ := r.typ()
|
2022-02-14 09:41:19 -08:00
|
|
|
val := FixValue(typ, r.Value())
|
2021-09-01 11:53:36 -07:00
|
|
|
setType(name, typ)
|
|
|
|
|
setValue(name, val)
|
2024-02-08 13:07:33 -05:00
|
|
|
return name, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.ObjFunc:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if sym.Name == "init" {
|
2021-04-29 17:02:53 +08:00
|
|
|
sym = Renameinit()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-16 19:45:12 -07:00
|
|
|
npos := r.pos()
|
|
|
|
|
setBasePos(npos)
|
|
|
|
|
r.typeParamNames()
|
|
|
|
|
typ := r.signature(nil)
|
|
|
|
|
fpos := r.pos()
|
|
|
|
|
|
|
|
|
|
fn := ir.NewFunc(fpos, npos, sym, typ)
|
|
|
|
|
name := fn.Nname
|
|
|
|
|
if !sym.IsBlank() {
|
|
|
|
|
if sym.Def != nil {
|
|
|
|
|
base.FatalfAt(name.Pos(), "already have a definition for %v", name)
|
|
|
|
|
}
|
|
|
|
|
assert(sym.Def == nil)
|
|
|
|
|
sym.Def = name
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-03-08 14:59:32 -08:00
|
|
|
if r.hasTypeParams() {
|
|
|
|
|
name.Func.SetDupok(true)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if r.dict.shaped {
|
|
|
|
|
setType(name, shapeSig(name.Func, r.dict))
|
|
|
|
|
} else {
|
|
|
|
|
todoDicts = append(todoDicts, func() {
|
|
|
|
|
r.dict.shapedObj = pr.objIdx(idx, implicits, explicits, true).(*ir.Name)
|
|
|
|
|
})
|
|
|
|
|
}
|
2022-03-08 14:59:32 -08:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
rext.funcExt(name, nil)
|
2024-02-08 13:07:33 -05:00
|
|
|
return name, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.ObjType:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name := do(ir.OTYPE, true)
|
|
|
|
|
typ := types.NewNamed(name)
|
2021-09-01 11:53:36 -07:00
|
|
|
setType(name, typ)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if r.hasTypeParams() && r.dict.shaped {
|
|
|
|
|
typ.SetHasShape(true)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
// Important: We need to do this before SetUnderlying.
|
2021-09-07 17:24:50 -07:00
|
|
|
rext.typeExt(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
// We need to defer CheckSize until we've called SetUnderlying to
|
|
|
|
|
// handle recursive types.
|
|
|
|
|
types.DeferCheckSize()
|
2021-08-30 01:01:20 -07:00
|
|
|
typ.SetUnderlying(r.typWrapped(false))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
types.ResumeCheckSize()
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if r.hasTypeParams() && !r.dict.shaped {
|
|
|
|
|
todoDicts = append(todoDicts, func() {
|
|
|
|
|
r.dict.shapedObj = pr.objIdx(idx, implicits, explicits, true).(*ir.Name)
|
|
|
|
|
})
|
|
|
|
|
}
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
methods := make([]*types.Field, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range methods {
|
2021-09-07 17:24:50 -07:00
|
|
|
methods[i] = r.method(rext)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
if len(methods) != 0 {
|
2023-08-20 10:05:29 -07:00
|
|
|
typ.SetMethods(methods)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if !r.dict.shaped {
|
|
|
|
|
r.needWrapper(typ)
|
|
|
|
|
}
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2024-02-08 13:07:33 -05:00
|
|
|
return name, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
case pkgbits.ObjVar:
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name := do(ir.ONAME, false)
|
2021-09-01 11:53:36 -07:00
|
|
|
setType(name, r.typ())
|
2021-09-07 17:24:50 -07:00
|
|
|
rext.varExt(name)
|
2024-02-08 13:07:33 -05:00
|
|
|
return name, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (dict *readerDict) mangle(sym *types.Sym) *types.Sym {
|
|
|
|
|
if !dict.hasTypeParams() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return sym
|
|
|
|
|
}
|
|
|
|
|
|
2022-08-18 15:16:23 -07:00
|
|
|
// If sym is a locally defined generic type, we need the suffix to
|
|
|
|
|
// stay at the end after mangling so that types/fmt.go can strip it
|
|
|
|
|
// out again when writing the type's runtime descriptor (#54456).
|
|
|
|
|
base, suffix := types.SplitVargenSuffix(sym.Name)
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var buf strings.Builder
|
2022-08-18 15:16:23 -07:00
|
|
|
buf.WriteString(base)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
buf.WriteByte('[')
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
for i, targ := range dict.targs {
|
2021-06-28 22:41:50 -07:00
|
|
|
if i > 0 {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if i == dict.implicits {
|
2021-06-28 22:41:50 -07:00
|
|
|
buf.WriteByte(';')
|
|
|
|
|
} else {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
buf.WriteByte(',')
|
|
|
|
|
}
|
|
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
buf.WriteString(targ.LinkString())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
buf.WriteByte(']')
|
2022-08-18 15:16:23 -07:00
|
|
|
buf.WriteString(suffix)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return sym.Pkg.Lookup(buf.String())
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// shapify returns the shape type for targ.
|
2022-08-18 03:14:23 -07:00
|
|
|
//
|
|
|
|
|
// If basic is true, then the type argument is used to instantiate a
|
|
|
|
|
// type parameter whose constraint is a basic interface.
|
|
|
|
|
func shapify(targ *types.Type, basic bool) *types.Type {
|
2022-08-29 17:25:32 +07:00
|
|
|
if targ.Kind() == types.TFORW {
|
|
|
|
|
if targ.IsFullyInstantiated() {
|
|
|
|
|
// For recursive instantiated type argument, it may still be a TFORW
|
|
|
|
|
// when shapifying happens. If we don't have targ's underlying type,
|
|
|
|
|
// shapify won't work. The worst case is we end up not reusing code
|
|
|
|
|
// optimally in some tricky cases.
|
|
|
|
|
if base.Debug.Shapify != 0 {
|
|
|
|
|
base.Warn("skipping shaping of recursive type %v", targ)
|
|
|
|
|
}
|
|
|
|
|
if targ.HasShape() {
|
|
|
|
|
return targ
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
base.Fatalf("%v is missing its underlying type", targ)
|
|
|
|
|
}
|
|
|
|
|
}
|
2024-01-31 14:47:40 +07:00
|
|
|
// For fully instantiated shape interface type, use it as-is. Otherwise, the instantiation
|
|
|
|
|
// involved recursive generic interface may cause mismatching in function signature, see issue #65362.
|
|
|
|
|
if targ.Kind() == types.TINTER && targ.IsFullyInstantiated() && targ.HasShape() {
|
|
|
|
|
return targ
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
2022-08-18 03:14:23 -07:00
|
|
|
// When a pointer type is used to instantiate a type parameter
|
|
|
|
|
// constrained by a basic interface, we know the pointer's element
|
|
|
|
|
// type can't matter to the generated code. In this case, we can use
|
|
|
|
|
// an arbitrary pointer type as the shape type. (To match the
|
|
|
|
|
// non-unified frontend, we use `*byte`.)
|
|
|
|
|
//
|
|
|
|
|
// Otherwise, we simply use the type's underlying type as its shape.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): It should be possible to do much more aggressive
|
|
|
|
|
// shaping still; e.g., collapsing all pointer-shaped types into a
|
|
|
|
|
// common type, collapsing scalars of the same size/alignment into a
|
|
|
|
|
// common type, recursively shaping the element types of composite
|
|
|
|
|
// types, and discarding struct field names and tags. However, we'll
|
|
|
|
|
// need to start tracking how type parameters are actually used to
|
|
|
|
|
// implement some of these optimizations.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
under := targ.Underlying()
|
2022-08-18 03:14:23 -07:00
|
|
|
if basic && targ.IsPtr() && !targ.Elem().NotInHeap() {
|
|
|
|
|
under = types.NewPtr(types.Types[types.TUINT8])
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
2024-01-09 15:06:15 +00:00
|
|
|
// Hash long type names to bound symbol name length seen by users,
|
|
|
|
|
// particularly for large protobuf structs (#65030).
|
|
|
|
|
uls := under.LinkString()
|
|
|
|
|
if base.Debug.MaxShapeLen != 0 &&
|
|
|
|
|
len(uls) > base.Debug.MaxShapeLen {
|
2024-09-04 18:30:35 +07:00
|
|
|
h := hash.Sum32([]byte(uls))
|
2024-01-09 15:06:15 +00:00
|
|
|
uls = hex.EncodeToString(h[:])
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
sym := types.ShapePkg.Lookup(uls)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if sym.Def == nil {
|
|
|
|
|
name := ir.NewDeclNameAt(under.Pos(), ir.OTYPE, sym)
|
|
|
|
|
typ := types.NewNamed(name)
|
|
|
|
|
typ.SetUnderlying(under)
|
|
|
|
|
sym.Def = typed(typ, name)
|
|
|
|
|
}
|
|
|
|
|
res := sym.Def.Type()
|
|
|
|
|
assert(res.IsShape())
|
|
|
|
|
assert(res.HasShape())
|
|
|
|
|
return res
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// objDictIdx reads and returns the specified object dictionary.
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) objDictIdx(sym *types.Sym, idx index, implicits, explicits []*types.Type, shaped bool) (*readerDict, error) {
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pr.newReader(pkgbits.SectionObjDict, idx, pkgbits.SyncObject1)
|
2021-07-20 16:52:37 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
dict := readerDict{
|
|
|
|
|
shaped: shaped,
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
nimplicits := r.Len()
|
|
|
|
|
nexplicits := r.Len()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
if nimplicits > len(implicits) || nexplicits != len(explicits) {
|
2024-02-08 13:07:33 -05:00
|
|
|
return nil, fmt.Errorf("%v has %v+%v params, but instantiated with %v+%v args", sym, nimplicits, nexplicits, len(implicits), len(explicits))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2021-07-20 16:52:37 -07:00
|
|
|
dict.targs = append(implicits[:nimplicits:nimplicits], explicits...)
|
|
|
|
|
dict.implicits = nimplicits
|
2021-06-28 22:41:50 -07:00
|
|
|
|
2022-08-18 03:14:23 -07:00
|
|
|
// Within the compiler, we can just skip over the type parameters.
|
|
|
|
|
for range dict.targs[dict.implicits:] {
|
|
|
|
|
// Skip past bounds without actually evaluating them.
|
|
|
|
|
r.typInfo()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dict.derived = make([]derivedInfo, r.Len())
|
|
|
|
|
dict.derivedTypes = make([]*types.Type, len(dict.derived))
|
|
|
|
|
for i := range dict.derived {
|
2025-05-12 12:21:03 -04:00
|
|
|
dict.derived[i] = derivedInfo{idx: r.Reloc(pkgbits.SectionType)}
|
2024-08-25 21:06:49 +07:00
|
|
|
if r.Version().Has(pkgbits.DerivedInfoNeeded) {
|
|
|
|
|
assert(!r.Bool())
|
|
|
|
|
}
|
2022-08-18 03:14:23 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Runtime dictionary information; private to the compiler.
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// If any type argument is already shaped, then we're constructing a
|
|
|
|
|
// shaped object, even if not explicitly requested (i.e., calling
|
|
|
|
|
// objIdx with shaped==true). This can happen with instantiating
|
|
|
|
|
// types that are referenced within a function body.
|
|
|
|
|
for _, targ := range dict.targs {
|
|
|
|
|
if targ.HasShape() {
|
|
|
|
|
dict.shaped = true
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// And if we're constructing a shaped object, then shapify all type
|
|
|
|
|
// arguments.
|
2022-08-18 03:14:23 -07:00
|
|
|
for i, targ := range dict.targs {
|
|
|
|
|
basic := r.Bool()
|
2025-06-03 21:35:20 +07:00
|
|
|
isPointerShape := basic && targ.IsPtr() && !targ.Elem().NotInHeap()
|
|
|
|
|
// We should not do shapify during the reshaping process, see #71184.
|
|
|
|
|
// However, this only matters for shapify a pointer type, which will
|
|
|
|
|
// lose the original underlying type.
|
|
|
|
|
//
|
|
|
|
|
// Example with a pointer type:
|
|
|
|
|
//
|
|
|
|
|
// - First, shapifying *[]T -> *uint8
|
|
|
|
|
// - During the reshaping process, *uint8 is shapified to *go.shape.uint8
|
|
|
|
|
// - This ends up with a different type with the original *[]T
|
|
|
|
|
//
|
|
|
|
|
// For a non-pointer type:
|
|
|
|
|
//
|
|
|
|
|
// - int -> go.shape.int
|
|
|
|
|
// - go.shape.int -> go.shape.int
|
|
|
|
|
//
|
|
|
|
|
// We always end up with the identical type.
|
|
|
|
|
canShapify := !pr.reshaping || !isPointerShape
|
|
|
|
|
if dict.shaped && canShapify {
|
2022-08-18 03:14:23 -07:00
|
|
|
dict.targs[i] = shapify(targ, basic)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dict.baseSym = dict.mangle(sym)
|
|
|
|
|
|
|
|
|
|
dict.typeParamMethodExprs = make([]readerMethodExprInfo, r.Len())
|
|
|
|
|
for i := range dict.typeParamMethodExprs {
|
|
|
|
|
typeParamIdx := r.Len()
|
2023-09-13 20:01:56 -07:00
|
|
|
method := r.selector()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
dict.typeParamMethodExprs[i] = readerMethodExprInfo{typeParamIdx, method}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dict.subdicts = make([]objInfo, r.Len())
|
|
|
|
|
for i := range dict.subdicts {
|
|
|
|
|
dict.subdicts[i] = r.objInfo()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dict.rtypes = make([]typeInfo, r.Len())
|
|
|
|
|
for i := range dict.rtypes {
|
|
|
|
|
dict.rtypes[i] = r.typInfo()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dict.itabs = make([]itabInfo, r.Len())
|
|
|
|
|
for i := range dict.itabs {
|
|
|
|
|
dict.itabs[i] = itabInfo{typ: r.typInfo(), iface: r.typInfo()}
|
|
|
|
|
}
|
|
|
|
|
|
2024-02-08 13:07:33 -05:00
|
|
|
return &dict, nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) typeParamNames() {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncTypeParamNames)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
for range r.dict.targs[r.dict.implicits:] {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.pos()
|
|
|
|
|
r.localIdent()
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2021-09-07 17:24:50 -07:00
|
|
|
func (r *reader) method(rext *reader) *types.Field {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncMethod)
|
2023-08-16 19:45:12 -07:00
|
|
|
npos := r.pos()
|
2023-09-13 20:01:56 -07:00
|
|
|
sym := r.selector()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.typeParamNames()
|
2023-09-13 20:01:56 -07:00
|
|
|
recv := r.param()
|
2022-12-01 17:42:02 -08:00
|
|
|
typ := r.signature(recv)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-08-16 19:45:12 -07:00
|
|
|
fpos := r.pos()
|
|
|
|
|
fn := ir.NewFunc(fpos, npos, ir.MethodSym(recv.Type, sym), typ)
|
|
|
|
|
name := fn.Nname
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-03-08 14:59:32 -08:00
|
|
|
if r.hasTypeParams() {
|
|
|
|
|
name.Func.SetDupok(true)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if r.dict.shaped {
|
|
|
|
|
typ = shapeSig(name.Func, r.dict)
|
|
|
|
|
setType(name, typ)
|
|
|
|
|
}
|
2022-03-08 14:59:32 -08:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
rext.funcExt(name, sym)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
meth := types.NewField(name.Func.Pos(), sym, typ)
|
|
|
|
|
meth.Nname = name
|
2021-07-02 16:59:01 -07:00
|
|
|
meth.SetNointerface(name.Func.Pragma&ir.Nointerface != 0)
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return meth
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) qualifiedIdent() (pkg *types.Pkg, sym *types.Sym) {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncSym)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pkg = r.pkg()
|
2022-02-14 09:41:19 -08:00
|
|
|
if name := r.String(); name != "" {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
sym = pkg.Lookup(name)
|
|
|
|
|
}
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2023-09-13 20:01:56 -07:00
|
|
|
func (r *reader) localIdent() *types.Sym {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncLocalIdent)
|
2023-09-13 20:01:56 -07:00
|
|
|
pkg := r.pkg()
|
2022-02-14 09:41:19 -08:00
|
|
|
if name := r.String(); name != "" {
|
2023-09-13 20:01:56 -07:00
|
|
|
return pkg.Lookup(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-09-13 20:01:56 -07:00
|
|
|
return nil
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-09-13 20:01:56 -07:00
|
|
|
func (r *reader) selector() *types.Sym {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncSelector)
|
2023-09-13 20:01:56 -07:00
|
|
|
pkg := r.pkg()
|
2022-02-14 09:41:19 -08:00
|
|
|
name := r.String()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if types.IsExported(name) {
|
|
|
|
|
pkg = types.LocalPkg
|
|
|
|
|
}
|
2023-09-13 20:01:56 -07:00
|
|
|
return pkg.Lookup(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2021-06-20 14:33:51 +07:00
|
|
|
func (r *reader) hasTypeParams() bool {
|
2021-07-20 16:52:37 -07:00
|
|
|
return r.dict.hasTypeParams()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (dict *readerDict) hasTypeParams() bool {
|
|
|
|
|
return dict != nil && len(dict.targs) != 0
|
2021-06-20 14:33:51 +07:00
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// @@@ Compiler extensions
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (r *reader) funcExt(name *ir.Name, method *types.Sym) {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncFuncExt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
fn := name.Func
|
|
|
|
|
|
|
|
|
|
// XXX: Workaround because linker doesn't know how to copy Pos.
|
|
|
|
|
if !fn.Pos().IsKnown() {
|
|
|
|
|
fn.SetPos(name.Pos())
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-20 22:17:19 +07:00
|
|
|
// Normally, we only compile local functions, which saves redundant compilation work.
|
|
|
|
|
// n.Defn is not nil for local functions, and is nil for imported function. But for
|
|
|
|
|
// generic functions, we might have an instantiation that no other package has seen before.
|
|
|
|
|
// So we need to be conservative and compile it again.
|
|
|
|
|
//
|
|
|
|
|
// That's why name.Defn is set here, so ir.VisitFuncsBottomUp can analyze function.
|
|
|
|
|
// TODO(mdempsky,cuonglm): find a cleaner way to handle this.
|
2021-06-20 14:33:51 +07:00
|
|
|
if name.Sym().Pkg == types.LocalPkg || r.hasTypeParams() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name.Defn = fn
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fn.Pragma = r.pragmaFlag()
|
|
|
|
|
r.linkname(name)
|
|
|
|
|
|
2023-01-22 15:30:59 -08:00
|
|
|
if buildcfg.GOARCH == "wasm" {
|
cmd/compile: add basic wasmexport support
This CL adds a compiler directive go:wasmexport, which applies to
a Go function and makes it an exported function of the Wasm module
being built, so it can be called directly from the host. As
proposed in #65199, parameter and result types are limited to
32-bit and 64-bit integers and floats, and there can be at most
one result.
As the Go and Wasm calling conventions are different, for a
wasmexport function we generate a wrapper function does the ABI
conversion at compile time.
Currently this CL only adds basic support. In particular,
- it only supports executable mode, i.e. the Go wasm module calls
into the host via wasmimport, which then calls back to Go via
wasmexport. Library (c-shared) mode is not implemented yet.
- only supports wasip1, not js.
- if the exported function unwinds stacks (goroutine switch, stack
growth, etc.), it probably doesn't work.
TODO: support stack unwinding, c-shared mode, js.
For #65199.
Change-Id: Id1777c2d44f7d51942c1caed3173c0a82f120cc4
Reviewed-on: https://go-review.googlesource.com/c/go/+/603055
Reviewed-by: Than McIntosh <thanm@golang.org>
Reviewed-by: Randy Reddig <randy.reddig@fastly.com>
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
2024-08-03 14:20:58 -04:00
|
|
|
importmod := r.String()
|
|
|
|
|
importname := r.String()
|
|
|
|
|
exportname := r.String()
|
2023-01-22 15:30:59 -08:00
|
|
|
|
cmd/compile: add basic wasmexport support
This CL adds a compiler directive go:wasmexport, which applies to
a Go function and makes it an exported function of the Wasm module
being built, so it can be called directly from the host. As
proposed in #65199, parameter and result types are limited to
32-bit and 64-bit integers and floats, and there can be at most
one result.
As the Go and Wasm calling conventions are different, for a
wasmexport function we generate a wrapper function does the ABI
conversion at compile time.
Currently this CL only adds basic support. In particular,
- it only supports executable mode, i.e. the Go wasm module calls
into the host via wasmimport, which then calls back to Go via
wasmexport. Library (c-shared) mode is not implemented yet.
- only supports wasip1, not js.
- if the exported function unwinds stacks (goroutine switch, stack
growth, etc.), it probably doesn't work.
TODO: support stack unwinding, c-shared mode, js.
For #65199.
Change-Id: Id1777c2d44f7d51942c1caed3173c0a82f120cc4
Reviewed-on: https://go-review.googlesource.com/c/go/+/603055
Reviewed-by: Than McIntosh <thanm@golang.org>
Reviewed-by: Randy Reddig <randy.reddig@fastly.com>
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
2024-08-03 14:20:58 -04:00
|
|
|
if importmod != "" && importname != "" {
|
2023-01-22 15:30:59 -08:00
|
|
|
fn.WasmImport = &ir.WasmImport{
|
cmd/compile: add basic wasmexport support
This CL adds a compiler directive go:wasmexport, which applies to
a Go function and makes it an exported function of the Wasm module
being built, so it can be called directly from the host. As
proposed in #65199, parameter and result types are limited to
32-bit and 64-bit integers and floats, and there can be at most
one result.
As the Go and Wasm calling conventions are different, for a
wasmexport function we generate a wrapper function does the ABI
conversion at compile time.
Currently this CL only adds basic support. In particular,
- it only supports executable mode, i.e. the Go wasm module calls
into the host via wasmimport, which then calls back to Go via
wasmexport. Library (c-shared) mode is not implemented yet.
- only supports wasip1, not js.
- if the exported function unwinds stacks (goroutine switch, stack
growth, etc.), it probably doesn't work.
TODO: support stack unwinding, c-shared mode, js.
For #65199.
Change-Id: Id1777c2d44f7d51942c1caed3173c0a82f120cc4
Reviewed-on: https://go-review.googlesource.com/c/go/+/603055
Reviewed-by: Than McIntosh <thanm@golang.org>
Reviewed-by: Randy Reddig <randy.reddig@fastly.com>
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
2024-08-03 14:20:58 -04:00
|
|
|
Module: importmod,
|
|
|
|
|
Name: importname,
|
2023-01-22 15:30:59 -08:00
|
|
|
}
|
|
|
|
|
}
|
cmd/compile: add basic wasmexport support
This CL adds a compiler directive go:wasmexport, which applies to
a Go function and makes it an exported function of the Wasm module
being built, so it can be called directly from the host. As
proposed in #65199, parameter and result types are limited to
32-bit and 64-bit integers and floats, and there can be at most
one result.
As the Go and Wasm calling conventions are different, for a
wasmexport function we generate a wrapper function does the ABI
conversion at compile time.
Currently this CL only adds basic support. In particular,
- it only supports executable mode, i.e. the Go wasm module calls
into the host via wasmimport, which then calls back to Go via
wasmexport. Library (c-shared) mode is not implemented yet.
- only supports wasip1, not js.
- if the exported function unwinds stacks (goroutine switch, stack
growth, etc.), it probably doesn't work.
TODO: support stack unwinding, c-shared mode, js.
For #65199.
Change-Id: Id1777c2d44f7d51942c1caed3173c0a82f120cc4
Reviewed-on: https://go-review.googlesource.com/c/go/+/603055
Reviewed-by: Than McIntosh <thanm@golang.org>
Reviewed-by: Randy Reddig <randy.reddig@fastly.com>
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
2024-08-03 14:20:58 -04:00
|
|
|
if exportname != "" {
|
|
|
|
|
if method != nil {
|
|
|
|
|
base.ErrorfAt(fn.Pos(), 0, "cannot use //go:wasmexport on a method")
|
|
|
|
|
}
|
|
|
|
|
fn.WasmExport = &ir.WasmExport{Name: exportname}
|
|
|
|
|
}
|
2023-01-22 15:30:59 -08:00
|
|
|
}
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
2022-07-26 21:52:42 -07:00
|
|
|
assert(name.Defn == nil)
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
fn.ABI = obj.ABI(r.Uint64())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
// Escape analysis.
|
2023-08-20 15:07:00 -07:00
|
|
|
for _, f := range name.Type().RecvParams() {
|
|
|
|
|
f.Note = r.String()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
fn.Inl = &ir.Inline{
|
2022-02-14 09:41:19 -08:00
|
|
|
Cost: int32(r.Len()),
|
|
|
|
|
CanDelayResults: r.Bool(),
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-11-16 22:19:37 +07:00
|
|
|
if buildcfg.Experiment.NewInliner {
|
2023-06-29 13:22:26 -04:00
|
|
|
fn.Inl.Properties = r.String()
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
} else {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
r.addBody(name.Func, method)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncEOF)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) typeExt(name *ir.Name) {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncTypeExt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
typ := name.Type()
|
|
|
|
|
|
2021-06-20 14:33:51 +07:00
|
|
|
if r.hasTypeParams() {
|
2024-09-05 23:09:08 +07:00
|
|
|
// Mark type as fully instantiated to ensure the type descriptor is written
|
|
|
|
|
// out as DUPOK and method wrappers are generated even for imported types.
|
|
|
|
|
typ.SetIsFullyInstantiated(true)
|
|
|
|
|
// HasShape should be set if any type argument is or has a shape type.
|
|
|
|
|
for _, targ := range r.dict.targs {
|
|
|
|
|
if targ.HasShape() {
|
|
|
|
|
typ.SetHasShape(true)
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
name.SetPragma(r.pragmaFlag())
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
typecheck.SetBaseTypeIndex(typ, r.Int64(), r.Int64())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) varExt(name *ir.Name) {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncVarExt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.linkname(name)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) linkname(name *ir.Name) {
|
|
|
|
|
assert(name.Op() == ir.ONAME)
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncLinkname)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if idx := r.Int64(); idx >= 0 {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
lsym := name.Linksym()
|
|
|
|
|
lsym.SymIdx = int32(idx)
|
|
|
|
|
lsym.Set(obj.AttrIndexed, true)
|
|
|
|
|
} else {
|
cmd/compile, cmd/link: disallow linkname of some newly added internal functions
Go API is defined through exported symbols. When a package is
imported, the compiler ensures that only exported symbols can be
accessed, and the go command ensures that internal packages cannot
be imported. This ensures API integrity. But there is a hole:
using linkname, one can access internal or non-exported symbols.
Linkname is a mechanism to give access of a symbol to a package
without adding it to the public API. It is intended for coupled
packages to share some implementation details, or to break
circular dependencies, and both "push" (definition) and "pull"
(reference) sides are controlled, so they can be updated in sync.
Nevertheless, it is abused as a mechanism to reach into internal
details of other packages uncontrolled by the user, notably the
runtime. As the other package evolves, the code often breaks,
because the linknamed symbol may no longer exist, or change its
signature or semantics.
This CL adds a mechanism to enforce the integrity of linknames.
Generally, "push" linkname is allowed, as the package defining
the symbol explicitly opt in for access outside of the package.
"Pull" linkname is checked and only allowed in some circumstances.
Given that there are existing code that use "pull"-only linkname
to access other package's internals, disallowing it completely is
too much a change at this point in the release cycle. For a start,
implement a hard-coded blocklist, which contains some newly added
internal functions that, if used inappropriately, may break memory
safety or runtime integrity. All blocked symbols are newly added
in Go 1.23. So existing code that builds with Go 1.22 will
continue to build.
For the implementation, when compiling a package, we mark
linknamed symbols in the current package with an attribute. At
link time, marked linknamed symbols are checked against the
blocklist. Care is taken so it distinguishes a linkname reference
in the current package vs. a reference of a linkname from another
package and propagated to the current package (e.g. through
inlining or instantiation).
Symbol references in assembly code are similar to linknames, and
are treated similarly.
Change-Id: I8067efe29c122740cd4f1effd2dec2d839147d5d
Reviewed-on: https://go-review.googlesource.com/c/go/+/584598
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2024-05-09 17:07:43 -04:00
|
|
|
linkname := r.String()
|
|
|
|
|
sym := name.Sym()
|
|
|
|
|
sym.Linkname = linkname
|
|
|
|
|
if sym.Pkg == types.LocalPkg && linkname != "" {
|
|
|
|
|
// Mark linkname in the current package. We don't mark the
|
|
|
|
|
// ones that are imported and propagated (e.g. through
|
|
|
|
|
// inlining or instantiation, which are marked in their
|
|
|
|
|
// corresponding packages). So we can tell in which package
|
|
|
|
|
// the linkname is used (pulled), and the linker can
|
|
|
|
|
// make a decision for allowing or disallowing it.
|
|
|
|
|
sym.Linksym().Set(obj.AttrLinkname, true)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) pragmaFlag() ir.PragmaFlag {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncPragma)
|
|
|
|
|
return ir.PragmaFlag(r.Int())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Function bodies
|
|
|
|
|
|
2022-07-26 21:52:42 -07:00
|
|
|
// bodyReader tracks where the serialized IR for a local or imported,
|
|
|
|
|
// generic function's body can be found.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var bodyReader = map[*ir.Func]pkgReaderIndex{}
|
|
|
|
|
|
2022-07-26 21:52:42 -07:00
|
|
|
// importBodyReader tracks where the serialized IR for an imported,
|
|
|
|
|
// static (i.e., non-generic) function body can be read.
|
|
|
|
|
var importBodyReader = map[*types.Sym]pkgReaderIndex{}
|
|
|
|
|
|
|
|
|
|
// bodyReaderFor returns the pkgReaderIndex for reading fn's
|
|
|
|
|
// serialized IR, and whether one was found.
|
|
|
|
|
func bodyReaderFor(fn *ir.Func) (pri pkgReaderIndex, ok bool) {
|
|
|
|
|
if fn.Nname.Defn != nil {
|
|
|
|
|
pri, ok = bodyReader[fn]
|
2022-07-31 18:48:16 -07:00
|
|
|
base.AssertfAt(ok, base.Pos, "must have bodyReader for %v", fn) // must always be available
|
2022-07-26 21:52:42 -07:00
|
|
|
} else {
|
|
|
|
|
pri, ok = importBodyReader[fn.Sym()]
|
|
|
|
|
}
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// todoDicts holds the list of dictionaries that still need their
|
|
|
|
|
// runtime dictionary objects constructed.
|
|
|
|
|
var todoDicts []func()
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// todoBodies holds the list of function bodies that still need to be
|
|
|
|
|
// constructed.
|
|
|
|
|
var todoBodies []*ir.Func
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// addBody reads a function body reference from the element bitstream,
|
|
|
|
|
// and associates it with fn.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (r *reader) addBody(fn *ir.Func, method *types.Sym) {
|
2022-07-26 21:52:42 -07:00
|
|
|
// addBody should only be called for local functions or imported
|
|
|
|
|
// generic functions; see comment in funcExt.
|
|
|
|
|
assert(fn.Nname.Defn != nil)
|
|
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
idx := r.Reloc(pkgbits.SectionBody)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
pri := pkgReaderIndex{r.p, idx, r.dict, method, nil}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
bodyReader[fn] = pri
|
|
|
|
|
|
2022-03-06 20:11:21 -08:00
|
|
|
if r.curfn == nil {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
todoBodies = append(todoBodies, fn)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pri.funcBody(fn)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pri pkgReaderIndex) funcBody(fn *ir.Func) {
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pri.asReader(pkgbits.SectionBody, pkgbits.SyncFuncBody)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.funcBody(fn)
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// funcBody reads a function body definition from the element
|
|
|
|
|
// bitstream, and populates fn with it.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) funcBody(fn *ir.Func) {
|
|
|
|
|
r.curfn = fn
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
r.closureVars = fn.ClosureVars
|
2022-07-31 18:48:16 -07:00
|
|
|
if len(r.closureVars) != 0 && r.hasTypeParams() {
|
|
|
|
|
r.dictParam = r.closureVars[len(r.closureVars)-1] // dictParam is last; see reader.funcLit
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-07-01 15:23:41 -07:00
|
|
|
ir.WithFunc(fn, func() {
|
2023-09-13 19:26:32 -07:00
|
|
|
r.declareParams()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if r.syntheticBody(fn.Pos()) {
|
2021-07-01 15:23:41 -07:00
|
|
|
return
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if !r.Bool() {
|
2022-07-31 18:48:16 -07:00
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
body := r.stmts()
|
|
|
|
|
if body == nil {
|
2022-02-11 11:55:27 -08:00
|
|
|
body = []ir.Node{typecheck.Stmt(ir.NewBlockStmt(src.NoXPos, nil))}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
fn.Body = body
|
|
|
|
|
fn.Endlineno = r.pos()
|
2021-07-01 15:23:41 -07:00
|
|
|
})
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
r.marker.WriteTo(fn)
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// syntheticBody adds a synthetic body to r.curfn if appropriate, and
|
|
|
|
|
// reports whether it did.
|
|
|
|
|
func (r *reader) syntheticBody(pos src.XPos) bool {
|
|
|
|
|
if r.synthetic != nil {
|
|
|
|
|
r.synthetic(pos, r)
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// If this function has type parameters and isn't shaped, then we
|
|
|
|
|
// just tail call its corresponding shaped variant.
|
|
|
|
|
if r.hasTypeParams() && !r.dict.shaped {
|
|
|
|
|
r.callShaped(pos)
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-31 18:48:16 -07:00
|
|
|
// callShaped emits a tail call to r.shapedFn, passing along the
|
|
|
|
|
// arguments to the current function.
|
|
|
|
|
func (r *reader) callShaped(pos src.XPos) {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
shapedObj := r.dict.shapedObj
|
|
|
|
|
assert(shapedObj != nil)
|
|
|
|
|
|
|
|
|
|
var shapedFn ir.Node
|
|
|
|
|
if r.methodSym == nil {
|
|
|
|
|
// Instantiating a generic function; shapedObj is the shaped
|
|
|
|
|
// function itself.
|
|
|
|
|
assert(shapedObj.Op() == ir.ONAME && shapedObj.Class == ir.PFUNC)
|
|
|
|
|
shapedFn = shapedObj
|
|
|
|
|
} else {
|
|
|
|
|
// Instantiating a generic type's method; shapedObj is the shaped
|
|
|
|
|
// type, so we need to select it's corresponding method.
|
|
|
|
|
shapedFn = shapedMethodExpr(pos, shapedObj, r.methodSym)
|
|
|
|
|
}
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
params := r.syntheticArgs()
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// Construct the arguments list: receiver (if any), then runtime
|
|
|
|
|
// dictionary, and finally normal parameters.
|
|
|
|
|
//
|
|
|
|
|
// Note: For simplicity, shaped methods are added as normal methods
|
|
|
|
|
// on their shaped types. So existing code (e.g., packages ir and
|
|
|
|
|
// typecheck) expects the shaped type to appear as the receiver
|
|
|
|
|
// parameter (or first parameter, as a method expression). Hence
|
|
|
|
|
// putting the dictionary parameter after that is the least invasive
|
|
|
|
|
// solution at the moment.
|
2022-07-31 18:48:16 -07:00
|
|
|
var args ir.Nodes
|
2023-09-13 19:26:32 -07:00
|
|
|
if r.methodSym != nil {
|
|
|
|
|
args.Append(params[0])
|
|
|
|
|
params = params[1:]
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
args.Append(typecheck.Expr(ir.NewAddrExpr(pos, r.p.dictNameOf(r.dict))))
|
|
|
|
|
args.Append(params...)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
r.syntheticTailCall(pos, shapedFn, args)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// syntheticArgs returns the recvs and params arguments passed to the
|
|
|
|
|
// current function.
|
2023-09-13 19:26:32 -07:00
|
|
|
func (r *reader) syntheticArgs() ir.Nodes {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
sig := r.curfn.Nname.Type()
|
2023-09-13 19:26:32 -07:00
|
|
|
return ir.ToNodes(r.curfn.Dcl[:sig.NumRecvs()+sig.NumParams()])
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// syntheticTailCall emits a tail call to fn, passing the given
|
|
|
|
|
// arguments list.
|
|
|
|
|
func (r *reader) syntheticTailCall(pos src.XPos, fn ir.Node, args ir.Nodes) {
|
2022-07-31 18:48:16 -07:00
|
|
|
// Mark the function as a wrapper so it doesn't show up in stack
|
|
|
|
|
// traces.
|
|
|
|
|
r.curfn.SetWrapper(true)
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
call := typecheck.Call(pos, fn, args, fn.Type().IsVariadic()).(*ir.CallExpr)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
|
|
|
|
var stmt ir.Node
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if fn.Type().NumResults() != 0 {
|
2022-07-31 18:48:16 -07:00
|
|
|
stmt = typecheck.Stmt(ir.NewReturnStmt(pos, []ir.Node{call}))
|
|
|
|
|
} else {
|
|
|
|
|
stmt = call
|
|
|
|
|
}
|
|
|
|
|
r.curfn.Body.Append(stmt)
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// dictNameOf returns the runtime dictionary corresponding to dict.
|
|
|
|
|
func (pr *pkgReader) dictNameOf(dict *readerDict) *ir.Name {
|
|
|
|
|
pos := base.AutogeneratedPos
|
|
|
|
|
|
|
|
|
|
// Check that we only instantiate runtime dictionaries with real types.
|
|
|
|
|
base.AssertfAt(!dict.shaped, pos, "runtime dictionary of shaped object %v", dict.baseSym)
|
|
|
|
|
|
|
|
|
|
sym := dict.baseSym.Pkg.Lookup(objabi.GlobalDictPrefix + "." + dict.baseSym.Name)
|
|
|
|
|
if sym.Def != nil {
|
|
|
|
|
return sym.Def.(*ir.Name)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-16 17:23:52 -07:00
|
|
|
name := ir.NewNameAt(pos, sym, dict.varType())
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
name.Class = ir.PEXTERN
|
|
|
|
|
sym.Def = name // break cycles with mutual subdictionaries
|
|
|
|
|
|
|
|
|
|
lsym := name.Linksym()
|
|
|
|
|
ot := 0
|
|
|
|
|
|
|
|
|
|
assertOffset := func(section string, offset int) {
|
|
|
|
|
base.AssertfAt(ot == offset*types.PtrSize, pos, "writing section %v at offset %v, but it should be at %v*%v", section, ot, offset, types.PtrSize)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
assertOffset("type param method exprs", dict.typeParamMethodExprsOffset())
|
|
|
|
|
for _, info := range dict.typeParamMethodExprs {
|
|
|
|
|
typeParam := dict.targs[info.typeParamIdx]
|
2023-08-18 18:03:24 -07:00
|
|
|
method := typecheck.NewMethodExpr(pos, typeParam, info.method)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
rsym := method.FuncName().Linksym()
|
|
|
|
|
assert(rsym.ABI() == obj.ABIInternal) // must be ABIInternal; see ir.OCFUNC in ssagen/ssa.go
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
ot = objw.SymPtr(lsym, ot, rsym, 0)
|
|
|
|
|
}
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
assertOffset("subdictionaries", dict.subdictsOffset())
|
|
|
|
|
for _, info := range dict.subdicts {
|
|
|
|
|
explicits := pr.typListIdx(info.explicits, dict)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// Careful: Due to subdictionary cycles, name may not be fully
|
|
|
|
|
// initialized yet.
|
|
|
|
|
name := pr.objDictName(info.idx, dict.targs, explicits)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
ot = objw.SymPtr(lsym, ot, name.Linksym(), 0)
|
|
|
|
|
}
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
assertOffset("rtypes", dict.rtypesOffset())
|
|
|
|
|
for _, info := range dict.rtypes {
|
|
|
|
|
typ := pr.typIdx(info, dict, true)
|
|
|
|
|
ot = objw.SymPtr(lsym, ot, reflectdata.TypeLinksym(typ), 0)
|
2022-07-31 18:48:16 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// TODO(mdempsky): Double check this.
|
|
|
|
|
reflectdata.MarkTypeUsedInInterface(typ, lsym)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
2022-09-01 16:58:15 -07:00
|
|
|
// For each (typ, iface) pair, we write the *runtime.itab pointer
|
|
|
|
|
// for the pair. For pairs that don't actually require an itab
|
|
|
|
|
// (i.e., typ is an interface, or iface is an empty interface), we
|
|
|
|
|
// write a nil pointer instead. This is wasteful, but rare in
|
|
|
|
|
// practice (e.g., instantiating a type parameter with an interface
|
|
|
|
|
// type).
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
assertOffset("itabs", dict.itabsOffset())
|
|
|
|
|
for _, info := range dict.itabs {
|
|
|
|
|
typ := pr.typIdx(info.typ, dict, true)
|
|
|
|
|
iface := pr.typIdx(info.iface, dict, true)
|
|
|
|
|
|
2022-09-01 16:58:15 -07:00
|
|
|
if !typ.IsInterface() && iface.IsInterface() && !iface.IsEmptyInterface() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
ot = objw.SymPtr(lsym, ot, reflectdata.ITabLsym(typ, iface), 0)
|
|
|
|
|
} else {
|
|
|
|
|
ot += types.PtrSize
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Double check this.
|
|
|
|
|
reflectdata.MarkTypeUsedInInterface(typ, lsym)
|
|
|
|
|
reflectdata.MarkTypeUsedInInterface(iface, lsym)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
objw.Global(lsym, int32(ot), obj.DUPOK|obj.RODATA)
|
|
|
|
|
|
|
|
|
|
return name
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// typeParamMethodExprsOffset returns the offset of the runtime
|
|
|
|
|
// dictionary's type parameter method expressions section, in words.
|
|
|
|
|
func (dict *readerDict) typeParamMethodExprsOffset() int {
|
|
|
|
|
return 0
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// subdictsOffset returns the offset of the runtime dictionary's
|
|
|
|
|
// subdictionary section, in words.
|
|
|
|
|
func (dict *readerDict) subdictsOffset() int {
|
|
|
|
|
return dict.typeParamMethodExprsOffset() + len(dict.typeParamMethodExprs)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// rtypesOffset returns the offset of the runtime dictionary's rtypes
|
|
|
|
|
// section, in words.
|
|
|
|
|
func (dict *readerDict) rtypesOffset() int {
|
|
|
|
|
return dict.subdictsOffset() + len(dict.subdicts)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// itabsOffset returns the offset of the runtime dictionary's itabs
|
|
|
|
|
// section, in words.
|
|
|
|
|
func (dict *readerDict) itabsOffset() int {
|
|
|
|
|
return dict.rtypesOffset() + len(dict.rtypes)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// numWords returns the total number of words that comprise dict's
|
|
|
|
|
// runtime dictionary variable.
|
2022-07-31 18:48:16 -07:00
|
|
|
func (dict *readerDict) numWords() int64 {
|
2022-09-01 16:58:15 -07:00
|
|
|
return int64(dict.itabsOffset() + len(dict.itabs))
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// varType returns the type of dict's runtime dictionary variable.
|
|
|
|
|
func (dict *readerDict) varType() *types.Type {
|
|
|
|
|
return types.NewArray(types.Types[types.TUINTPTR], dict.numWords())
|
|
|
|
|
}
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
func (r *reader) declareParams() {
|
|
|
|
|
r.curfn.DeclareParams(!r.funarghack)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
for _, name := range r.curfn.Dcl {
|
|
|
|
|
if name.Sym().Name == dictParamName {
|
|
|
|
|
r.dictParam = name
|
|
|
|
|
continue
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
r.addLocal(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
func (r *reader) addLocal(name *ir.Name) {
|
|
|
|
|
if r.synthetic == nil {
|
|
|
|
|
r.Sync(pkgbits.SyncAddLocal)
|
|
|
|
|
if r.p.SyncMarkers() {
|
|
|
|
|
want := r.Int()
|
|
|
|
|
if have := len(r.locals); have != want {
|
|
|
|
|
base.FatalfAt(name.Pos(), "locals table has desynced")
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-09-13 19:26:32 -07:00
|
|
|
r.varDictIndex(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
r.locals = append(r.locals, name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) useLocal() *ir.Name {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncUseObjLocal)
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
return r.locals[r.Len()]
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
}
|
2022-02-14 09:41:19 -08:00
|
|
|
return r.closureVars[r.Len()]
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) openScope() {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncOpenScope)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.pos()
|
|
|
|
|
|
|
|
|
|
if base.Flag.Dwarf {
|
|
|
|
|
r.scopeVars = append(r.scopeVars, len(r.curfn.Dcl))
|
|
|
|
|
r.marker.Push(pos)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) closeScope() {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncCloseScope)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.lastCloseScopePos = r.pos()
|
|
|
|
|
|
|
|
|
|
r.closeAnotherScope()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// closeAnotherScope is like closeScope, but it reuses the same mark
|
|
|
|
|
// position as the last closeScope call. This is useful for "for" and
|
|
|
|
|
// "if" statements, as their implicit blocks always end at the same
|
|
|
|
|
// position as an explicit block.
|
|
|
|
|
func (r *reader) closeAnotherScope() {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncCloseAnotherScope)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
if base.Flag.Dwarf {
|
|
|
|
|
scopeVars := r.scopeVars[len(r.scopeVars)-1]
|
|
|
|
|
r.scopeVars = r.scopeVars[:len(r.scopeVars)-1]
|
|
|
|
|
|
2021-07-01 15:23:41 -07:00
|
|
|
// Quirkish: noder decides which scopes to keep before
|
|
|
|
|
// typechecking, whereas incremental typechecking during IR
|
|
|
|
|
// construction can result in new autotemps being allocated. To
|
|
|
|
|
// produce identical output, we ignore autotemps here for the
|
|
|
|
|
// purpose of deciding whether to retract the scope.
|
|
|
|
|
//
|
|
|
|
|
// This is important for net/http/fcgi, because it contains:
|
|
|
|
|
//
|
|
|
|
|
// var body io.ReadCloser
|
|
|
|
|
// if len(content) > 0 {
|
|
|
|
|
// body, req.pw = io.Pipe()
|
|
|
|
|
// } else { … }
|
|
|
|
|
//
|
|
|
|
|
// Notably, io.Pipe is inlinable, and inlining it introduces a ~R0
|
|
|
|
|
// variable at the call site.
|
|
|
|
|
//
|
|
|
|
|
// Noder does not preserve the scope where the io.Pipe() call
|
|
|
|
|
// resides, because it doesn't contain any declared variables in
|
|
|
|
|
// source. So the ~R0 variable ends up being assigned to the
|
|
|
|
|
// enclosing scope instead.
|
|
|
|
|
//
|
|
|
|
|
// However, typechecking this assignment also introduces
|
|
|
|
|
// autotemps, because io.Pipe's results need conversion before
|
|
|
|
|
// they can be assigned to their respective destination variables.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): We should probably just keep all scopes, and
|
|
|
|
|
// let dwarfgen take care of pruning them instead.
|
|
|
|
|
retract := true
|
|
|
|
|
for _, n := range r.curfn.Dcl[scopeVars:] {
|
|
|
|
|
if !n.AutoTemp() {
|
|
|
|
|
retract = false
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if retract {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// no variables were declared in this scope, so we can retract it.
|
|
|
|
|
r.marker.Unpush()
|
|
|
|
|
} else {
|
|
|
|
|
r.marker.Pop(r.lastCloseScopePos)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Statements
|
|
|
|
|
|
|
|
|
|
func (r *reader) stmt() ir.Node {
|
2023-08-22 17:44:19 -07:00
|
|
|
return block(r.stmts())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func block(stmts []ir.Node) ir.Node {
|
|
|
|
|
switch len(stmts) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case 0:
|
|
|
|
|
return nil
|
|
|
|
|
case 1:
|
|
|
|
|
return stmts[0]
|
|
|
|
|
default:
|
|
|
|
|
return ir.NewBlockStmt(stmts[0].Pos(), stmts)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-22 17:44:19 -07:00
|
|
|
func (r *reader) stmts() ir.Nodes {
|
2021-07-01 15:23:41 -07:00
|
|
|
assert(ir.CurFunc == r.curfn)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var res ir.Nodes
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncStmts)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for {
|
2022-02-14 09:41:19 -08:00
|
|
|
tag := codeStmt(r.Code(pkgbits.SyncStmt1))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if tag == stmtEnd {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncStmtsEnd)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return res
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if n := r.stmt1(tag, &res); n != nil {
|
2021-07-01 15:23:41 -07:00
|
|
|
res.Append(typecheck.Stmt(n))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) stmt1(tag codeStmt, out *ir.Nodes) ir.Node {
|
|
|
|
|
var label *types.Sym
|
|
|
|
|
if n := len(*out); n > 0 {
|
|
|
|
|
if ls, ok := (*out)[n-1].(*ir.LabelStmt); ok {
|
|
|
|
|
label = ls.Label
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch tag {
|
|
|
|
|
default:
|
|
|
|
|
panic("unexpected statement")
|
|
|
|
|
|
|
|
|
|
case stmtAssign:
|
|
|
|
|
pos := r.pos()
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
names, lhs := r.assignList()
|
2022-06-24 14:19:06 -07:00
|
|
|
rhs := r.multiExpr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
if len(rhs) == 0 {
|
|
|
|
|
for _, name := range names {
|
|
|
|
|
as := ir.NewAssignStmt(pos, name, nil)
|
|
|
|
|
as.PtrInit().Append(ir.NewDecl(pos, ir.ODCL, name))
|
2021-07-01 15:23:41 -07:00
|
|
|
out.Append(typecheck.Stmt(as))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
return nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if len(lhs) == 1 && len(rhs) == 1 {
|
|
|
|
|
n := ir.NewAssignStmt(pos, lhs[0], rhs[0])
|
|
|
|
|
n.Def = r.initDefn(n, names)
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
n := ir.NewAssignListStmt(pos, ir.OAS2, lhs, rhs)
|
|
|
|
|
n.Def = r.initDefn(n, names)
|
|
|
|
|
return n
|
|
|
|
|
|
|
|
|
|
case stmtAssignOp:
|
|
|
|
|
op := r.op()
|
|
|
|
|
lhs := r.expr()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
rhs := r.expr()
|
|
|
|
|
return ir.NewAssignOpStmt(pos, op, lhs, rhs)
|
|
|
|
|
|
|
|
|
|
case stmtIncDec:
|
|
|
|
|
op := r.op()
|
|
|
|
|
lhs := r.expr()
|
|
|
|
|
pos := r.pos()
|
2023-09-08 15:44:57 -07:00
|
|
|
n := ir.NewAssignOpStmt(pos, op, lhs, ir.NewOne(pos, lhs.Type()))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
n.IncDec = true
|
|
|
|
|
return n
|
|
|
|
|
|
|
|
|
|
case stmtBlock:
|
|
|
|
|
out.Append(r.blockStmt()...)
|
|
|
|
|
return nil
|
|
|
|
|
|
|
|
|
|
case stmtBranch:
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
op := r.op()
|
|
|
|
|
sym := r.optLabel()
|
|
|
|
|
return ir.NewBranchStmt(pos, op, sym)
|
|
|
|
|
|
|
|
|
|
case stmtCall:
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
op := r.op()
|
|
|
|
|
call := r.expr()
|
2023-06-14 10:56:49 -04:00
|
|
|
stmt := ir.NewGoDeferStmt(pos, op, call)
|
|
|
|
|
if op == ir.ODEFER {
|
|
|
|
|
x := r.optExpr()
|
|
|
|
|
if x != nil {
|
|
|
|
|
stmt.DeferAt = x.(ir.Expr)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return stmt
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case stmtExpr:
|
|
|
|
|
return r.expr()
|
|
|
|
|
|
|
|
|
|
case stmtFor:
|
|
|
|
|
return r.forStmt(label)
|
|
|
|
|
|
|
|
|
|
case stmtIf:
|
|
|
|
|
return r.ifStmt()
|
|
|
|
|
|
|
|
|
|
case stmtLabel:
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
sym := r.label()
|
|
|
|
|
return ir.NewLabelStmt(pos, sym)
|
|
|
|
|
|
|
|
|
|
case stmtReturn:
|
|
|
|
|
pos := r.pos()
|
2022-06-24 14:19:06 -07:00
|
|
|
results := r.multiExpr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return ir.NewReturnStmt(pos, results)
|
|
|
|
|
|
|
|
|
|
case stmtSelect:
|
|
|
|
|
return r.selectStmt(label)
|
|
|
|
|
|
|
|
|
|
case stmtSend:
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
ch := r.expr()
|
|
|
|
|
value := r.expr()
|
|
|
|
|
return ir.NewSendStmt(pos, ch, value)
|
|
|
|
|
|
|
|
|
|
case stmtSwitch:
|
|
|
|
|
return r.switchStmt(label)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) assignList() ([]*ir.Name, []ir.Node) {
|
2022-02-14 09:41:19 -08:00
|
|
|
lhs := make([]ir.Node, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var names []*ir.Name
|
|
|
|
|
|
|
|
|
|
for i := range lhs {
|
2022-06-03 13:21:09 -07:00
|
|
|
expr, def := r.assign()
|
|
|
|
|
lhs[i] = expr
|
|
|
|
|
if def {
|
|
|
|
|
names = append(names, expr.(*ir.Name))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return names, lhs
|
|
|
|
|
}
|
|
|
|
|
|
2022-06-03 13:21:09 -07:00
|
|
|
// assign returns an assignee expression. It also reports whether the
|
|
|
|
|
// returned expression is a newly declared variable.
|
|
|
|
|
func (r *reader) assign() (ir.Node, bool) {
|
|
|
|
|
switch tag := codeAssign(r.Code(pkgbits.SyncAssign)); tag {
|
|
|
|
|
default:
|
|
|
|
|
panic("unhandled assignee expression")
|
|
|
|
|
|
|
|
|
|
case assignBlank:
|
|
|
|
|
return typecheck.AssignExpr(ir.BlankNode), false
|
|
|
|
|
|
|
|
|
|
case assignDef:
|
|
|
|
|
pos := r.pos()
|
2023-09-13 20:01:56 -07:00
|
|
|
setBasePos(pos) // test/fixedbugs/issue49767.go depends on base.Pos being set for the r.typ() call here, ugh
|
|
|
|
|
name := r.curfn.NewLocal(pos, r.localIdent(), r.typ())
|
2023-09-13 19:26:32 -07:00
|
|
|
r.addLocal(name)
|
2022-06-03 13:21:09 -07:00
|
|
|
return name, true
|
|
|
|
|
|
|
|
|
|
case assignExpr:
|
|
|
|
|
return r.expr(), false
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) blockStmt() []ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncBlockStmt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.openScope()
|
|
|
|
|
stmts := r.stmts()
|
|
|
|
|
r.closeScope()
|
|
|
|
|
return stmts
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) forStmt(label *types.Sym) ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncForStmt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
r.openScope()
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.pos()
|
2023-01-25 17:08:16 -05:00
|
|
|
rang := ir.NewRangeStmt(pos, nil, nil, nil, nil, false)
|
2022-07-22 23:33:30 -07:00
|
|
|
rang.Label = label
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
|
|
|
|
|
names, lhs := r.assignList()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if len(lhs) >= 1 {
|
|
|
|
|
rang.Key = lhs[0]
|
|
|
|
|
if len(lhs) >= 2 {
|
|
|
|
|
rang.Value = lhs[1]
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
rang.Def = r.initDefn(rang, names)
|
2022-06-30 17:58:47 -07:00
|
|
|
|
2022-07-22 23:33:30 -07:00
|
|
|
rang.X = r.expr()
|
|
|
|
|
if rang.X.Type().IsMap() {
|
|
|
|
|
rang.RType = r.rtype(pos)
|
|
|
|
|
}
|
2022-07-23 00:46:10 -07:00
|
|
|
if rang.Key != nil && !ir.IsBlank(rang.Key) {
|
|
|
|
|
rang.KeyTypeWord, rang.KeySrcRType = r.convRTTI(pos)
|
|
|
|
|
}
|
|
|
|
|
if rang.Value != nil && !ir.IsBlank(rang.Value) {
|
|
|
|
|
rang.ValueTypeWord, rang.ValueSrcRType = r.convRTTI(pos)
|
2022-06-30 17:58:47 -07:00
|
|
|
}
|
|
|
|
|
|
2022-07-22 23:33:30 -07:00
|
|
|
rang.Body = r.blockStmt()
|
2023-01-25 17:08:16 -05:00
|
|
|
rang.DistinctVars = r.Bool()
|
2022-07-22 23:33:30 -07:00
|
|
|
r.closeAnotherScope()
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return rang
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
init := r.stmt()
|
2022-06-03 12:49:11 -07:00
|
|
|
cond := r.optExpr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
post := r.stmt()
|
|
|
|
|
body := r.blockStmt()
|
2023-07-05 16:21:19 -04:00
|
|
|
perLoopVars := r.Bool()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.closeAnotherScope()
|
|
|
|
|
|
2023-08-09 04:05:35 -07:00
|
|
|
if ir.IsConst(cond, constant.Bool) && !ir.BoolVal(cond) {
|
|
|
|
|
return init // simplify "for init; false; post { ... }" into "init"
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-05 16:21:19 -04:00
|
|
|
stmt := ir.NewForStmt(pos, init, cond, post, body, perLoopVars)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
stmt.Label = label
|
|
|
|
|
return stmt
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) ifStmt() ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncIfStmt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.openScope()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
init := r.stmts()
|
|
|
|
|
cond := r.expr()
|
2023-08-22 17:44:19 -07:00
|
|
|
staticCond := r.Int()
|
|
|
|
|
var then, els []ir.Node
|
|
|
|
|
if staticCond >= 0 {
|
|
|
|
|
then = r.blockStmt()
|
|
|
|
|
} else {
|
|
|
|
|
r.lastCloseScopePos = r.pos()
|
|
|
|
|
}
|
|
|
|
|
if staticCond <= 0 {
|
|
|
|
|
els = r.stmts()
|
|
|
|
|
}
|
2023-08-09 04:05:35 -07:00
|
|
|
r.closeAnotherScope()
|
|
|
|
|
|
2023-08-22 17:44:19 -07:00
|
|
|
if staticCond != 0 {
|
|
|
|
|
// We may have removed a dead return statement, which can trip up
|
|
|
|
|
// later passes (#62211). To avoid confusion, we instead flatten
|
|
|
|
|
// the if statement into a block.
|
|
|
|
|
|
|
|
|
|
if cond.Op() != ir.OLITERAL {
|
|
|
|
|
init.Append(typecheck.Stmt(ir.NewAssignStmt(pos, ir.BlankNode, cond))) // for side effects
|
|
|
|
|
}
|
|
|
|
|
init.Append(then...)
|
|
|
|
|
init.Append(els...)
|
|
|
|
|
return block(init)
|
2023-08-09 04:05:35 -07:00
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
n := ir.NewIfStmt(pos, cond, then, els)
|
|
|
|
|
n.SetInit(init)
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) selectStmt(label *types.Sym) ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncSelectStmt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
pos := r.pos()
|
2022-02-14 09:41:19 -08:00
|
|
|
clauses := make([]*ir.CommClause, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range clauses {
|
|
|
|
|
if i > 0 {
|
|
|
|
|
r.closeScope()
|
|
|
|
|
}
|
|
|
|
|
r.openScope()
|
|
|
|
|
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
comm := r.stmt()
|
|
|
|
|
body := r.stmts()
|
|
|
|
|
|
2022-07-18 12:41:56 -07:00
|
|
|
// "case i = <-c: ..." may require an implicit conversion (e.g.,
|
|
|
|
|
// see fixedbugs/bug312.go). Currently, typecheck throws away the
|
|
|
|
|
// implicit conversion and relies on it being reinserted later,
|
|
|
|
|
// but that would lose any explicit RTTI operands too. To preserve
|
|
|
|
|
// RTTI, we rewrite this as "case tmp := <-c: i = tmp; ...".
|
|
|
|
|
if as, ok := comm.(*ir.AssignStmt); ok && as.Op() == ir.OAS && !as.Def {
|
|
|
|
|
if conv, ok := as.Y.(*ir.ConvExpr); ok && conv.Op() == ir.OCONVIFACE {
|
|
|
|
|
base.AssertfAt(conv.Implicit(), conv.Pos(), "expected implicit conversion: %v", conv)
|
|
|
|
|
|
|
|
|
|
recv := conv.X
|
|
|
|
|
base.AssertfAt(recv.Op() == ir.ORECV, recv.Pos(), "expected receive expression: %v", recv)
|
|
|
|
|
|
|
|
|
|
tmp := r.temp(pos, recv.Type())
|
|
|
|
|
|
|
|
|
|
// Replace comm with `tmp := <-c`.
|
|
|
|
|
tmpAs := ir.NewAssignStmt(pos, tmp, recv)
|
|
|
|
|
tmpAs.Def = true
|
|
|
|
|
tmpAs.PtrInit().Append(ir.NewDecl(pos, ir.ODCL, tmp))
|
|
|
|
|
comm = tmpAs
|
|
|
|
|
|
|
|
|
|
// Change original assignment to `i = tmp`, and prepend to body.
|
|
|
|
|
conv.X = tmp
|
|
|
|
|
body = append([]ir.Node{as}, body...)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[dev.unified] cmd/compile/internal/noder: implicit conversions for multi-valued expressions
This CL changes GOEXPERIMENT=unified to insert implicit conversions
for multi-valued expressions.
Unfortunately, IR doesn't have strong, first-class support for
multi-valued expressions, so this CL takes the approach of spilling
them to temporary variables, which can then be implicitly converted.
This is the same approach taken by walk, but doing it this early does
introduce some minor complications:
1. For select case clauses with comma-ok assignments (e.g., `case x,
ok := <-ch:`), the compiler middle end wants to see the OAS2RECV
assignment is the CommClause.Comm statement. So when constructing
select statements, we need to massage this around a little.
2. The extra temporary variables and assignments skew the existing
inlining heuristics. As mentioned, the temporaries/assignments will
eventually be added (and often optimized away again) anyway, but now
they're visible to the inliner. So this CL also kludges the inlining
heuristics in this case to keep things comparable.
Change-Id: I3e3ea756ad92472ebe28bae3963be61ed7684a75
Reviewed-on: https://go-review.googlesource.com/c/go/+/415244
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-06-28 16:31:29 -07:00
|
|
|
// multiExpr will have desugared a comma-ok receive expression
|
|
|
|
|
// into a separate statement. However, the rest of the compiler
|
|
|
|
|
// expects comm to be the OAS2RECV statement itself, so we need to
|
|
|
|
|
// shuffle things around to fit that pattern.
|
|
|
|
|
if as2, ok := comm.(*ir.AssignListStmt); ok && as2.Op() == ir.OAS2 {
|
|
|
|
|
init := ir.TakeInit(as2.Rhs[0])
|
|
|
|
|
base.AssertfAt(len(init) == 1 && init[0].Op() == ir.OAS2RECV, as2.Pos(), "unexpected assignment: %+v", as2)
|
|
|
|
|
|
|
|
|
|
comm = init[0]
|
|
|
|
|
body = append([]ir.Node{as2}, body...)
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
clauses[i] = ir.NewCommStmt(pos, comm, body)
|
|
|
|
|
}
|
|
|
|
|
if len(clauses) > 0 {
|
|
|
|
|
r.closeScope()
|
|
|
|
|
}
|
|
|
|
|
n := ir.NewSelectStmt(pos, clauses)
|
|
|
|
|
n.Label = label
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) switchStmt(label *types.Sym) ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncSwitchStmt)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
r.openScope()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
init := r.stmt()
|
2021-07-09 17:47:15 -07:00
|
|
|
|
|
|
|
|
var tag ir.Node
|
2022-03-06 23:47:27 -08:00
|
|
|
var ident *ir.Ident
|
|
|
|
|
var iface *types.Type
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
2021-07-09 17:47:15 -07:00
|
|
|
pos := r.pos()
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() {
|
2023-09-13 20:01:56 -07:00
|
|
|
ident = ir.NewIdent(r.pos(), r.localIdent())
|
2021-07-09 17:47:15 -07:00
|
|
|
}
|
|
|
|
|
x := r.expr()
|
2022-03-06 23:47:27 -08:00
|
|
|
iface = x.Type()
|
2021-07-09 17:47:15 -07:00
|
|
|
tag = ir.NewTypeSwitchGuard(pos, ident, x)
|
|
|
|
|
} else {
|
2022-06-03 12:49:11 -07:00
|
|
|
tag = r.optExpr()
|
2021-07-09 17:47:15 -07:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
clauses := make([]*ir.CaseClause, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range clauses {
|
|
|
|
|
if i > 0 {
|
|
|
|
|
r.closeScope()
|
|
|
|
|
}
|
|
|
|
|
r.openScope()
|
|
|
|
|
|
|
|
|
|
pos := r.pos()
|
2022-06-21 02:30:21 -07:00
|
|
|
var cases, rtypes []ir.Node
|
2022-03-06 23:47:27 -08:00
|
|
|
if iface != nil {
|
|
|
|
|
cases = make([]ir.Node, r.Len())
|
|
|
|
|
if len(cases) == 0 {
|
|
|
|
|
cases = nil // TODO(mdempsky): Unclear if this matters.
|
|
|
|
|
}
|
|
|
|
|
for i := range cases {
|
2022-07-25 12:21:21 -07:00
|
|
|
if r.Bool() { // case nil
|
|
|
|
|
cases[i] = typecheck.Expr(types.BuiltinPkg.Lookup("nil").Def.(*ir.NilExpr))
|
|
|
|
|
} else {
|
|
|
|
|
cases[i] = r.exprType()
|
|
|
|
|
}
|
2022-03-06 23:47:27 -08:00
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
cases = r.exprList()
|
2022-06-21 02:30:21 -07:00
|
|
|
|
2022-07-18 13:45:57 -07:00
|
|
|
// For `switch { case any(true): }` (e.g., issue 3980 in
|
|
|
|
|
// test/switch.go), the backend still creates a mixed bool/any
|
|
|
|
|
// comparison, and we need to explicitly supply the RTTI for the
|
|
|
|
|
// comparison.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): Change writer.go to desugar "switch {" into
|
|
|
|
|
// "switch true {", which we already handle correctly.
|
|
|
|
|
if tag == nil {
|
|
|
|
|
for i, cas := range cases {
|
|
|
|
|
if cas.Type().IsEmptyInterface() {
|
|
|
|
|
for len(rtypes) < i {
|
|
|
|
|
rtypes = append(rtypes, nil)
|
|
|
|
|
}
|
|
|
|
|
rtypes = append(rtypes, reflectdata.TypePtrAt(cas.Pos(), types.Types[types.TBOOL]))
|
2022-06-21 02:30:21 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
2022-03-06 23:47:27 -08:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
clause := ir.NewCaseStmt(pos, cases, nil)
|
2022-06-21 02:30:21 -07:00
|
|
|
clause.RTypes = rtypes
|
2022-03-06 23:47:27 -08:00
|
|
|
|
|
|
|
|
if ident != nil {
|
2023-09-13 19:26:32 -07:00
|
|
|
name := r.curfn.NewLocal(r.pos(), ident.Sym(), r.typ())
|
|
|
|
|
r.addLocal(name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
clause.Var = name
|
2022-03-06 23:47:27 -08:00
|
|
|
name.Defn = tag
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
clause.Body = r.stmts()
|
|
|
|
|
clauses[i] = clause
|
|
|
|
|
}
|
|
|
|
|
if len(clauses) > 0 {
|
|
|
|
|
r.closeScope()
|
|
|
|
|
}
|
|
|
|
|
r.closeScope()
|
|
|
|
|
|
|
|
|
|
n := ir.NewSwitchStmt(pos, tag, clauses)
|
|
|
|
|
n.Label = label
|
|
|
|
|
if init != nil {
|
|
|
|
|
n.SetInit([]ir.Node{init})
|
|
|
|
|
}
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) label() *types.Sym {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncLabel)
|
|
|
|
|
name := r.String()
|
2024-11-03 15:52:29 +07:00
|
|
|
if r.inlCall != nil && name != "_" {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
name = fmt.Sprintf("~%s·%d", name, inlgen)
|
|
|
|
|
}
|
|
|
|
|
return typecheck.Lookup(name)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) optLabel() *types.Sym {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncOptLabel)
|
|
|
|
|
if r.Bool() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return r.label()
|
|
|
|
|
}
|
|
|
|
|
return nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// initDefn marks the given names as declared by defn and populates
|
|
|
|
|
// its Init field with ODCL nodes. It then reports whether any names
|
|
|
|
|
// were so declared, which can be used to initialize defn.Def.
|
|
|
|
|
func (r *reader) initDefn(defn ir.InitNode, names []*ir.Name) bool {
|
|
|
|
|
if len(names) == 0 {
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
init := make([]ir.Node, len(names))
|
|
|
|
|
for i, name := range names {
|
|
|
|
|
name.Defn = defn
|
|
|
|
|
init[i] = ir.NewDecl(name.Pos(), ir.ODCL, name)
|
|
|
|
|
}
|
|
|
|
|
defn.SetInit(init)
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Expressions
|
|
|
|
|
|
2021-07-09 17:47:15 -07:00
|
|
|
// expr reads and returns a typechecked expression.
|
2021-07-09 19:47:45 -07:00
|
|
|
func (r *reader) expr() (res ir.Node) {
|
|
|
|
|
defer func() {
|
|
|
|
|
if res != nil && res.Typecheck() == 0 {
|
|
|
|
|
base.FatalfAt(res.Pos(), "%v missed typecheck", res)
|
|
|
|
|
}
|
|
|
|
|
}()
|
2021-07-09 17:47:15 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
switch tag := codeExpr(r.Code(pkgbits.SyncExpr)); tag {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
default:
|
|
|
|
|
panic("unhandled expression")
|
|
|
|
|
|
|
|
|
|
case exprLocal:
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(r.useLocal())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-06-03 12:38:39 -07:00
|
|
|
case exprGlobal:
|
2021-07-09 19:47:45 -07:00
|
|
|
// Callee instead of Expr allows builtins
|
|
|
|
|
// TODO(mdempsky): Handle builtins directly in exprCall, like method calls?
|
|
|
|
|
return typecheck.Callee(r.obj())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-08-05 20:29:51 -07:00
|
|
|
case exprFuncInst:
|
2023-02-14 15:13:50 -08:00
|
|
|
origPos, pos := r.origPos()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
wrapperFn, baseFn, dictPtr := r.funcInst(pos)
|
|
|
|
|
if wrapperFn != nil {
|
|
|
|
|
return wrapperFn
|
|
|
|
|
}
|
2023-02-14 15:13:50 -08:00
|
|
|
return r.curry(origPos, false, baseFn, dictPtr, nil)
|
2022-08-05 20:29:51 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case exprConst:
|
|
|
|
|
pos := r.pos()
|
2021-09-07 13:23:08 -07:00
|
|
|
typ := r.typ()
|
2022-02-14 09:41:19 -08:00
|
|
|
val := FixValue(typ, r.Value())
|
2023-09-08 15:44:57 -07:00
|
|
|
return ir.NewBasicLit(pos, typ, val)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-12 04:51:16 -07:00
|
|
|
case exprZero:
|
2022-07-18 13:13:46 -07:00
|
|
|
pos := r.pos()
|
|
|
|
|
typ := r.typ()
|
2023-09-12 04:51:16 -07:00
|
|
|
return ir.NewZero(pos, typ)
|
2022-07-18 13:13:46 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case exprCompLit:
|
|
|
|
|
return r.compLit()
|
|
|
|
|
|
|
|
|
|
case exprFuncLit:
|
|
|
|
|
return r.funcLit()
|
|
|
|
|
|
2022-08-05 20:40:45 -07:00
|
|
|
case exprFieldVal:
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
2023-09-13 20:01:56 -07:00
|
|
|
sym := r.selector()
|
2022-07-23 23:54:15 -07:00
|
|
|
|
2023-08-18 19:40:33 -07:00
|
|
|
return typecheck.XDotField(pos, x, sym)
|
2022-08-05 20:40:45 -07:00
|
|
|
|
|
|
|
|
case exprMethodVal:
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
recv := r.expr()
|
2023-02-14 15:13:50 -08:00
|
|
|
origPos, pos := r.origPos()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
wrapperFn, baseFn, dictPtr := r.methodExpr()
|
|
|
|
|
|
|
|
|
|
// For simple wrapperFn values, the existing machinery for creating
|
|
|
|
|
// and deduplicating wrapperFn value wrappers still works fine.
|
|
|
|
|
if wrapperFn, ok := wrapperFn.(*ir.SelectorExpr); ok && wrapperFn.Op() == ir.OMETHEXPR {
|
|
|
|
|
// The receiver expression we constructed may have a shape type.
|
|
|
|
|
// For example, in fixedbugs/issue54343.go, `New[int]()` is
|
|
|
|
|
// constructed as `New[go.shape.int](&.dict.New[int])`, which
|
|
|
|
|
// has type `*T[go.shape.int]`, not `*T[int]`.
|
|
|
|
|
//
|
|
|
|
|
// However, the method we want to select here is `(*T[int]).M`,
|
|
|
|
|
// not `(*T[go.shape.int]).M`, so we need to manually convert
|
|
|
|
|
// the type back so that the OXDOT resolves correctly.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): Logically it might make more sense for
|
|
|
|
|
// exprCall to take responsibility for setting a non-shaped
|
|
|
|
|
// result type, but this is the only place where we care
|
|
|
|
|
// currently. And only because existing ir.OMETHVALUE backend
|
|
|
|
|
// code relies on n.X.Type() instead of n.Selection.Recv().Type
|
|
|
|
|
// (because the latter is types.FakeRecvType() in the case of
|
|
|
|
|
// interface method values).
|
|
|
|
|
//
|
|
|
|
|
if recv.Type().HasShape() {
|
2023-08-20 10:05:29 -07:00
|
|
|
typ := wrapperFn.Type().Param(0).Type
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if !types.Identical(typ, recv.Type()) {
|
|
|
|
|
base.FatalfAt(wrapperFn.Pos(), "receiver %L does not match %L", recv, wrapperFn)
|
|
|
|
|
}
|
|
|
|
|
recv = typecheck.Expr(ir.NewConvExpr(recv.Pos(), ir.OCONVNOP, typ, recv))
|
|
|
|
|
}
|
2022-03-06 23:47:27 -08:00
|
|
|
|
2023-08-18 19:40:33 -07:00
|
|
|
n := typecheck.XDotMethod(pos, recv, wrapperFn.Sel, false)
|
2023-02-21 13:16:22 -08:00
|
|
|
|
|
|
|
|
// As a consistency check here, we make sure "n" selected the
|
|
|
|
|
// same method (represented by a types.Field) that wrapperFn
|
|
|
|
|
// selected. However, for anonymous receiver types, there can be
|
|
|
|
|
// multiple such types.Field instances (#58563). So we may need
|
|
|
|
|
// to fallback to making sure Sym and Type (including the
|
|
|
|
|
// receiver parameter's type) match.
|
|
|
|
|
if n.Selection != wrapperFn.Selection {
|
|
|
|
|
assert(n.Selection.Sym == wrapperFn.Selection.Sym)
|
|
|
|
|
assert(types.Identical(n.Selection.Type, wrapperFn.Selection.Type))
|
|
|
|
|
assert(types.Identical(n.Selection.Type.Recv().Type, wrapperFn.Selection.Type.Recv().Type))
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
wrapper := methodValueWrapper{
|
|
|
|
|
rcvr: n.X.Type(),
|
|
|
|
|
method: n.Selection,
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if r.importedDef() {
|
|
|
|
|
haveMethodValueWrappers = append(haveMethodValueWrappers, wrapper)
|
|
|
|
|
} else {
|
|
|
|
|
needMethodValueWrappers = append(needMethodValueWrappers, wrapper)
|
|
|
|
|
}
|
|
|
|
|
return n
|
2021-08-30 01:01:20 -07:00
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// For more complicated method expressions, we construct a
|
|
|
|
|
// function literal wrapper.
|
2023-02-14 15:13:50 -08:00
|
|
|
return r.curry(origPos, true, baseFn, recv, dictPtr)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-08-05 20:40:45 -07:00
|
|
|
case exprMethodExpr:
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
recv := r.typ()
|
|
|
|
|
|
|
|
|
|
implicits := make([]int, r.Len())
|
|
|
|
|
for i := range implicits {
|
|
|
|
|
implicits[i] = r.Len()
|
|
|
|
|
}
|
|
|
|
|
var deref, addr bool
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
deref = true
|
|
|
|
|
} else if r.Bool() {
|
|
|
|
|
addr = true
|
|
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
origPos, pos := r.origPos()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
wrapperFn, baseFn, dictPtr := r.methodExpr()
|
2022-08-05 20:40:45 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// If we already have a wrapper and don't need to do anything with
|
|
|
|
|
// it, we can just return the wrapper directly.
|
|
|
|
|
//
|
|
|
|
|
// N.B., we use implicits/deref/addr here as the source of truth
|
|
|
|
|
// rather than types.Identical, because the latter can be confused
|
|
|
|
|
// by tricky promoted methods (e.g., typeparam/mdempsky/21.go).
|
|
|
|
|
if wrapperFn != nil && len(implicits) == 0 && !deref && !addr {
|
2023-08-20 10:05:29 -07:00
|
|
|
if !types.Identical(recv, wrapperFn.Type().Param(0).Type) {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
base.FatalfAt(pos, "want receiver type %v, but have method %L", recv, wrapperFn)
|
|
|
|
|
}
|
|
|
|
|
return wrapperFn
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Otherwise, if the wrapper function is a static method
|
|
|
|
|
// expression (OMETHEXPR) and the receiver type is unshaped, then
|
|
|
|
|
// we can rely on a statically generated wrapper being available.
|
|
|
|
|
if method, ok := wrapperFn.(*ir.SelectorExpr); ok && method.Op() == ir.OMETHEXPR && !recv.HasShape() {
|
2023-08-18 18:03:24 -07:00
|
|
|
return typecheck.NewMethodExpr(pos, recv, method.Sel)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
return r.methodExprWrap(origPos, recv, implicits, deref, addr, baseFn, dictPtr)
|
2022-08-05 20:40:45 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case exprIndex:
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
index := r.expr()
|
2022-06-21 00:57:18 -07:00
|
|
|
n := typecheck.Expr(ir.NewIndexExpr(pos, x, index))
|
|
|
|
|
switch n.Op() {
|
|
|
|
|
case ir.OINDEXMAP:
|
|
|
|
|
n := n.(*ir.IndexExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-06-21 00:57:18 -07:00
|
|
|
}
|
|
|
|
|
return n
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case exprSlice:
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
var index [3]ir.Node
|
|
|
|
|
for i := range index {
|
2022-06-03 12:49:11 -07:00
|
|
|
index[i] = r.optExpr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
op := ir.OSLICE
|
|
|
|
|
if index[2] != nil {
|
|
|
|
|
op = ir.OSLICE3
|
|
|
|
|
}
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(ir.NewSliceExpr(pos, op, x, index[0], index[1], index[2]))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case exprAssert:
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
2022-07-25 12:21:21 -07:00
|
|
|
typ := r.exprType()
|
2022-07-22 23:33:30 -07:00
|
|
|
srcRType := r.rtype(pos)
|
2022-03-06 23:47:27 -08:00
|
|
|
|
2022-07-22 23:33:30 -07:00
|
|
|
// TODO(mdempsky): Always emit ODYNAMICDOTTYPE for uniformity?
|
2022-03-06 23:47:27 -08:00
|
|
|
if typ, ok := typ.(*ir.DynamicType); ok && typ.Op() == ir.ODYNAMICTYPE {
|
2022-05-09 17:19:58 -07:00
|
|
|
assert := ir.NewDynamicTypeAssertExpr(pos, ir.ODYNAMICDOTTYPE, x, typ.RType)
|
2022-07-22 23:33:30 -07:00
|
|
|
assert.SrcRType = srcRType
|
2022-05-09 17:19:58 -07:00
|
|
|
assert.ITab = typ.ITab
|
|
|
|
|
return typed(typ.Type(), assert)
|
2022-03-06 23:47:27 -08:00
|
|
|
}
|
2022-05-02 18:39:07 -07:00
|
|
|
return typecheck.Expr(ir.NewTypeAssertExpr(pos, x, typ.Type()))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case exprUnaryOp:
|
|
|
|
|
op := r.op()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
x := r.expr()
|
|
|
|
|
|
|
|
|
|
switch op {
|
|
|
|
|
case ir.OADDR:
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(typecheck.NodAddrAt(pos, x))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case ir.ODEREF:
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(ir.NewStarExpr(pos, x))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(ir.NewUnaryExpr(pos, op, x))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case exprBinaryOp:
|
|
|
|
|
op := r.op()
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
y := r.expr()
|
|
|
|
|
|
|
|
|
|
switch op {
|
|
|
|
|
case ir.OANDAND, ir.OOROR:
|
2021-07-09 19:47:45 -07:00
|
|
|
return typecheck.Expr(ir.NewLogicalExpr(pos, op, x, y))
|
2023-10-05 11:11:51 +07:00
|
|
|
case ir.OLSH, ir.ORSH:
|
|
|
|
|
// Untyped rhs of non-constant shift, e.g. x << 1.0.
|
|
|
|
|
// If we have a constant value, it must be an int >= 0.
|
|
|
|
|
if ir.IsConstNode(y) {
|
|
|
|
|
val := constant.ToInt(y.Val())
|
|
|
|
|
assert(val.Kind() == constant.Int && constant.Sign(val) >= 0)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2022-07-23 00:00:49 -07:00
|
|
|
return typecheck.Expr(ir.NewBinaryExpr(pos, op, x, y))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-08-06 16:08:41 -07:00
|
|
|
case exprRecv:
|
|
|
|
|
x := r.expr()
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
for i, n := 0, r.Len(); i < n; i++ {
|
2023-08-18 19:40:33 -07:00
|
|
|
x = Implicit(typecheck.DotField(pos, x, r.Len()))
|
2022-08-06 16:08:41 -07:00
|
|
|
}
|
|
|
|
|
if r.Bool() { // needs deref
|
|
|
|
|
x = Implicit(Deref(pos, x.Type().Elem(), x))
|
|
|
|
|
} else if r.Bool() { // needs addr
|
|
|
|
|
x = Implicit(Addr(pos, x))
|
|
|
|
|
}
|
|
|
|
|
return x
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
case exprCall:
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var fun ir.Node
|
|
|
|
|
var args ir.Nodes
|
2022-02-14 09:41:19 -08:00
|
|
|
if r.Bool() { // method call
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
recv := r.expr()
|
|
|
|
|
_, method, dictPtr := r.methodExpr()
|
|
|
|
|
|
|
|
|
|
if recv.Type().IsInterface() && method.Op() == ir.OMETHEXPR {
|
|
|
|
|
method := method.(*ir.SelectorExpr)
|
|
|
|
|
|
|
|
|
|
// The compiler backend (e.g., devirtualization) handle
|
|
|
|
|
// OCALLINTER/ODOTINTER better than OCALLFUNC/OMETHEXPR for
|
|
|
|
|
// interface calls, so we prefer to continue constructing
|
|
|
|
|
// calls that way where possible.
|
|
|
|
|
//
|
|
|
|
|
// There are also corner cases where semantically it's perhaps
|
|
|
|
|
// significant; e.g., fixedbugs/issue15975.go, #38634, #52025.
|
|
|
|
|
|
2023-08-18 19:40:33 -07:00
|
|
|
fun = typecheck.XDotMethod(method.Pos(), recv, method.Sel, true)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
} else {
|
|
|
|
|
if recv.Type().IsInterface() {
|
|
|
|
|
// N.B., this happens currently for typeparam/issue51521.go
|
|
|
|
|
// and typeparam/typeswitch3.go.
|
cmd/compile: experimental loop iterator capture semantics change
Adds:
GOEXPERIMENT=loopvar (expected way of invoking)
-d=loopvar={-1,0,1,2,11,12} (for per-package control and/or logging)
-d=loopvarhash=... (for hash debugging)
loopvar=11,12 are for testing, benchmarking, and debugging.
If enabled,for loops of the form `for x,y := range thing`, if x and/or
y are addressed or captured by a closure, are transformed by renaming
x/y to a temporary and prepending an assignment to the body of the
loop x := tmp_x. This changes the loop semantics by making each
iteration's instance of x be distinct from the others (currently they
are all aliased, and when this matters, it is almost always a bug).
3-range with captured iteration variables are also transformed,
though it is a more complex transformation.
"Optimized" to do a simpler transformation for
3-clause for where the increment is empty.
(Prior optimization of address-taking under Return disabled, because
it was incorrect; returns can have loops for children. Restored in
a later CL.)
Includes support for -d=loopvarhash=<binary string> intended for use
with hash search and GOCOMPILEDEBUG=loopvarhash=<binary string>
(use `gossahash -e loopvarhash command-that-fails`).
Minor feature upgrades to hash-triggered features; clients can specify
that file-position hashes use only the most-inline position, and/or that
they use only the basenames of source files (not the full directory path).
Most-inlined is the right choice for debugging loop-iteration change
once the semantics are linked to the package across inlining; basename-only
makes it tractable to write tests (which, otherwise, depend on the full
pathname of the source file and thus vary).
Updates #57969.
Change-Id: I180a51a3f8d4173f6210c861f10de23de8a1b1db
Reviewed-on: https://go-review.googlesource.com/c/go/+/411904
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-06-12 15:33:57 -04:00
|
|
|
if base.Flag.LowerM != 0 {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
base.WarnfAt(method.Pos(), "imprecise interface call")
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fun = method
|
|
|
|
|
args.Append(recv)
|
|
|
|
|
}
|
|
|
|
|
if dictPtr != nil {
|
|
|
|
|
args.Append(dictPtr)
|
|
|
|
|
}
|
|
|
|
|
} else if r.Bool() { // call to instanced function
|
2021-07-09 19:47:45 -07:00
|
|
|
pos := r.pos()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
_, shapedFn, dictPtr := r.funcInst(pos)
|
|
|
|
|
fun = shapedFn
|
|
|
|
|
args.Append(dictPtr)
|
|
|
|
|
} else {
|
|
|
|
|
fun = r.expr()
|
2021-07-09 19:47:45 -07:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.pos()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
args.Append(r.multiExpr()...)
|
2022-06-24 14:19:06 -07:00
|
|
|
dots := r.Bool()
|
2022-06-21 00:57:18 -07:00
|
|
|
n := typecheck.Call(pos, fun, args, dots)
|
|
|
|
|
switch n.Op() {
|
|
|
|
|
case ir.OAPPEND:
|
|
|
|
|
n := n.(*ir.CallExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-08-09 02:55:42 +07:00
|
|
|
// For append(a, b...), we don't need the implicit conversion. The typechecker already
|
|
|
|
|
// ensured that a and b are both slices with the same base type, or []byte and string.
|
|
|
|
|
if n.IsDDD {
|
|
|
|
|
if conv, ok := n.Args[1].(*ir.ConvExpr); ok && conv.Op() == ir.OCONVNOP && conv.Implicit() {
|
|
|
|
|
n.Args[1] = conv.X
|
|
|
|
|
}
|
|
|
|
|
}
|
2022-06-21 00:57:18 -07:00
|
|
|
case ir.OCOPY:
|
|
|
|
|
n := n.(*ir.BinaryExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-06-21 00:57:18 -07:00
|
|
|
case ir.ODELETE:
|
|
|
|
|
n := n.(*ir.CallExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-06-21 00:57:18 -07:00
|
|
|
case ir.OUNSAFESLICE:
|
|
|
|
|
n := n.(*ir.BinaryExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-06-21 00:57:18 -07:00
|
|
|
}
|
|
|
|
|
return n
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-06-03 14:30:04 -07:00
|
|
|
case exprMake:
|
|
|
|
|
pos := r.pos()
|
2022-07-25 12:21:21 -07:00
|
|
|
typ := r.exprType()
|
2022-06-03 14:30:04 -07:00
|
|
|
extra := r.exprs()
|
2022-06-21 00:57:18 -07:00
|
|
|
n := typecheck.Expr(ir.NewCallExpr(pos, ir.OMAKE, nil, append([]ir.Node{typ}, extra...))).(*ir.MakeExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
n.RType = r.rtype(pos)
|
2022-06-21 00:57:18 -07:00
|
|
|
return n
|
2022-06-03 14:30:04 -07:00
|
|
|
|
|
|
|
|
case exprNew:
|
|
|
|
|
pos := r.pos()
|
2022-07-25 12:21:21 -07:00
|
|
|
typ := r.exprType()
|
2022-06-03 14:30:04 -07:00
|
|
|
return typecheck.Expr(ir.NewUnaryExpr(pos, ir.ONEW, typ))
|
|
|
|
|
|
cmd/compile/internal/noder: handle unsafe.Sizeof, etc in unified IR
Previously, the unified frontend implemented unsafe.Sizeof, etc that
involved derived types by constructing a normal OSIZEOF, etc
expression, including fully instantiating their argument. (When
unsafe.Sizeof is applied to a non-generic type, types2 handles
constant folding it.)
This worked, but involves unnecessary work, since all we really need
to track is the argument type (and the field selections, for
unsafe.Offsetof).
Further, the argument expression could generate temporary variables,
which would then go unused after typecheck replaced the OSIZEOF
expression with an OLITERAL. This results in compiler failures after
CL 523315, which made later passes stricter about expecting the
frontend to not construct unused temporaries.
Fixes #62515.
Change-Id: I37baed048fd2e35648c59243f66c97c24413aa94
Reviewed-on: https://go-review.googlesource.com/c/go/+/527097
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
2023-09-08 16:21:17 -07:00
|
|
|
case exprSizeof:
|
|
|
|
|
return ir.NewUintptr(r.pos(), r.typ().Size())
|
|
|
|
|
|
|
|
|
|
case exprAlignof:
|
|
|
|
|
return ir.NewUintptr(r.pos(), r.typ().Alignment())
|
|
|
|
|
|
|
|
|
|
case exprOffsetof:
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
typ := r.typ()
|
|
|
|
|
types.CalcSize(typ)
|
|
|
|
|
|
|
|
|
|
var offset int64
|
|
|
|
|
for i := r.Len(); i >= 0; i-- {
|
|
|
|
|
field := typ.Field(r.Len())
|
|
|
|
|
offset += field.Offset
|
|
|
|
|
typ = field.Type
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return ir.NewUintptr(pos, offset)
|
|
|
|
|
|
2022-08-18 20:59:26 -07:00
|
|
|
case exprReshape:
|
|
|
|
|
typ := r.typ()
|
2025-01-10 12:49:59 +07:00
|
|
|
old := r.reshaping
|
|
|
|
|
r.reshaping = true
|
2022-08-18 20:59:26 -07:00
|
|
|
x := r.expr()
|
2025-01-10 12:49:59 +07:00
|
|
|
r.reshaping = old
|
2022-08-18 20:59:26 -07:00
|
|
|
|
|
|
|
|
if types.IdenticalStrict(x.Type(), typ) {
|
|
|
|
|
return x
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Comparison expressions are constructed as "untyped bool" still.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): It should be safe to reshape them here too, but
|
|
|
|
|
// maybe it's better to construct them with the proper type
|
|
|
|
|
// instead.
|
|
|
|
|
if x.Type() == types.UntypedBool && typ.IsBoolean() {
|
|
|
|
|
return x
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
base.AssertfAt(x.Type().HasShape() || typ.HasShape(), x.Pos(), "%L and %v are not shape types", x, typ)
|
|
|
|
|
base.AssertfAt(types.Identical(x.Type(), typ), x.Pos(), "%L is not shape-identical to %v", x, typ)
|
|
|
|
|
|
|
|
|
|
// We use ir.HasUniquePos here as a check that x only appears once
|
|
|
|
|
// in the AST, so it's okay for us to call SetType without
|
|
|
|
|
// breaking any other uses of it.
|
|
|
|
|
//
|
|
|
|
|
// Notably, any ONAMEs should already have the exactly right shape
|
|
|
|
|
// type and been caught by types.IdenticalStrict above.
|
|
|
|
|
base.AssertfAt(ir.HasUniquePos(x), x.Pos(), "cannot call SetType(%v) on %L", typ, x)
|
|
|
|
|
|
|
|
|
|
if base.Debug.Reshape != 0 {
|
|
|
|
|
base.WarnfAt(x.Pos(), "reshaping %L to %v", x, typ)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
x.SetType(typ)
|
|
|
|
|
return x
|
|
|
|
|
|
2021-07-09 17:47:15 -07:00
|
|
|
case exprConvert:
|
2022-06-21 07:07:41 -07:00
|
|
|
implicit := r.Bool()
|
2021-07-09 17:47:15 -07:00
|
|
|
typ := r.typ()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.pos()
|
2022-07-23 00:46:10 -07:00
|
|
|
typeWord, srcRType := r.convRTTI(pos)
|
2022-08-06 20:56:39 +07:00
|
|
|
dstTypeParam := r.Bool()
|
2022-09-01 16:06:11 -07:00
|
|
|
identical := r.Bool()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
x := r.expr()
|
2022-03-15 22:25:36 -07:00
|
|
|
|
|
|
|
|
// TODO(mdempsky): Stop constructing expressions of untyped type.
|
|
|
|
|
x = typecheck.DefaultLit(x, typ)
|
|
|
|
|
|
2022-08-06 20:56:39 +07:00
|
|
|
ce := ir.NewConvExpr(pos, ir.OCONV, typ, x)
|
|
|
|
|
ce.TypeWord, ce.SrcRType = typeWord, srcRType
|
2022-06-30 17:58:47 -07:00
|
|
|
if implicit {
|
2022-08-06 20:56:39 +07:00
|
|
|
ce.SetImplicit(true)
|
2022-06-21 07:07:41 -07:00
|
|
|
}
|
2022-08-06 20:56:39 +07:00
|
|
|
n := typecheck.Expr(ce)
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// Conversions between non-identical, non-empty interfaces always
|
|
|
|
|
// requires a runtime call, even if they have identical underlying
|
|
|
|
|
// interfaces. This is because we create separate itab instances
|
|
|
|
|
// for each unique interface type, not merely each unique
|
|
|
|
|
// interface shape.
|
|
|
|
|
//
|
|
|
|
|
// However, due to shape types, typecheck.Expr might mistakenly
|
|
|
|
|
// think a conversion between two non-empty interfaces are
|
|
|
|
|
// identical and set ir.OCONVNOP, instead of ir.OCONVIFACE. To
|
|
|
|
|
// ensure we update the itab field appropriately, we force it to
|
|
|
|
|
// ir.OCONVIFACE instead when shape types are involved.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): Are there other places we might get this wrong?
|
|
|
|
|
// Should this be moved down into typecheck.{Assign,Convert}op?
|
|
|
|
|
// This would be a non-issue if itabs were unique for each
|
|
|
|
|
// *underlying* interface type instead.
|
2022-09-01 16:06:11 -07:00
|
|
|
if !identical {
|
|
|
|
|
if n, ok := n.(*ir.ConvExpr); ok && n.Op() == ir.OCONVNOP && n.Type().IsInterface() && !n.Type().IsEmptyInterface() && (n.Type().HasShape() || n.X.Type().HasShape()) {
|
|
|
|
|
n.SetOp(ir.OCONVIFACE)
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
2022-08-06 20:56:39 +07:00
|
|
|
// spec: "If the type is a type parameter, the constant is converted
|
|
|
|
|
// into a non-constant value of the type parameter."
|
|
|
|
|
if dstTypeParam && ir.IsConstNode(n) {
|
|
|
|
|
// Wrap in an OCONVNOP node to ensure result is non-constant.
|
|
|
|
|
n = Implicit(ir.NewConvExpr(pos, ir.OCONVNOP, n.Type(), n))
|
|
|
|
|
n.SetTypecheck(1)
|
|
|
|
|
}
|
|
|
|
|
return n
|
2023-11-07 17:20:35 -05:00
|
|
|
|
|
|
|
|
case exprRuntimeBuiltin:
|
|
|
|
|
builtin := typecheck.LookupRuntime(r.String())
|
|
|
|
|
return builtin
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// funcInst reads an instantiated function reference, and returns
|
|
|
|
|
// three (possibly nil) expressions related to it:
|
|
|
|
|
//
|
|
|
|
|
// baseFn is always non-nil: it's either a function of the appropriate
|
|
|
|
|
// type already, or it has an extra dictionary parameter as the first
|
|
|
|
|
// parameter.
|
|
|
|
|
//
|
|
|
|
|
// If dictPtr is non-nil, then it's a dictionary argument that must be
|
|
|
|
|
// passed as the first argument to baseFn.
|
|
|
|
|
//
|
|
|
|
|
// If wrapperFn is non-nil, then it's either the same as baseFn (if
|
|
|
|
|
// dictPtr is nil), or it's semantically equivalent to currying baseFn
|
|
|
|
|
// to pass dictPtr. (wrapperFn is nil when dictPtr is an expression
|
|
|
|
|
// that needs to be computed dynamically.)
|
|
|
|
|
//
|
|
|
|
|
// For callers that are creating a call to the returned function, it's
|
|
|
|
|
// best to emit a call to baseFn, and include dictPtr in the arguments
|
|
|
|
|
// list as appropriate.
|
|
|
|
|
//
|
|
|
|
|
// For callers that want to return the function without invoking it,
|
|
|
|
|
// they may return wrapperFn if it's non-nil; but otherwise, they need
|
|
|
|
|
// to create their own wrapper.
|
|
|
|
|
func (r *reader) funcInst(pos src.XPos) (wrapperFn, baseFn, dictPtr ir.Node) {
|
|
|
|
|
// Like in methodExpr, I'm pretty sure this isn't needed.
|
|
|
|
|
var implicits []*types.Type
|
|
|
|
|
if r.dict != nil {
|
|
|
|
|
implicits = r.dict.targs
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if r.Bool() { // dynamic subdictionary
|
|
|
|
|
idx := r.Len()
|
|
|
|
|
info := r.dict.subdicts[idx]
|
|
|
|
|
explicits := r.p.typListIdx(info.explicits, r.dict)
|
|
|
|
|
|
2025-01-10 12:49:59 +07:00
|
|
|
old := r.p.reshaping
|
|
|
|
|
r.p.reshaping = r.reshaping
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
baseFn = r.p.objIdx(info.idx, implicits, explicits, true).(*ir.Name)
|
2025-01-10 12:49:59 +07:00
|
|
|
r.p.reshaping = old
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// TODO(mdempsky): Is there a more robust way to get the
|
|
|
|
|
// dictionary pointer type here?
|
2023-08-20 10:05:29 -07:00
|
|
|
dictPtrType := baseFn.Type().Param(0).Type
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
dictPtr = typecheck.Expr(ir.NewConvExpr(pos, ir.OCONVNOP, dictPtrType, r.dictWord(pos, r.dict.subdictsOffset()+idx)))
|
|
|
|
|
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
info := r.objInfo()
|
|
|
|
|
explicits := r.p.typListIdx(info.explicits, r.dict)
|
|
|
|
|
|
|
|
|
|
wrapperFn = r.p.objIdx(info.idx, implicits, explicits, false).(*ir.Name)
|
|
|
|
|
baseFn = r.p.objIdx(info.idx, implicits, explicits, true).(*ir.Name)
|
|
|
|
|
|
|
|
|
|
dictName := r.p.objDictName(info.idx, implicits, explicits)
|
|
|
|
|
dictPtr = typecheck.Expr(ir.NewAddrExpr(pos, dictName))
|
|
|
|
|
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2024-08-20 10:35:23 -07:00
|
|
|
func (pr *pkgReader) objDictName(idx index, implicits, explicits []*types.Type) *ir.Name {
|
2025-05-12 12:21:03 -04:00
|
|
|
rname := pr.newReader(pkgbits.SectionName, idx, pkgbits.SyncObject1)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
_, sym := rname.qualifiedIdent()
|
|
|
|
|
tag := pkgbits.CodeObj(rname.Code(pkgbits.SyncCodeObj))
|
|
|
|
|
|
|
|
|
|
if tag == pkgbits.ObjStub {
|
|
|
|
|
assert(!sym.IsBlank())
|
|
|
|
|
if pri, ok := objReader[sym]; ok {
|
|
|
|
|
return pri.pr.objDictName(pri.idx, nil, explicits)
|
|
|
|
|
}
|
|
|
|
|
base.Fatalf("unresolved stub: %v", sym)
|
|
|
|
|
}
|
|
|
|
|
|
2024-02-08 13:07:33 -05:00
|
|
|
dict, err := pr.objDictIdx(sym, idx, implicits, explicits, false)
|
|
|
|
|
if err != nil {
|
|
|
|
|
base.Fatalf("%v", err)
|
|
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
return pr.dictNameOf(dict)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// curry returns a function literal that calls fun with arg0 and
|
|
|
|
|
// (optionally) arg1, accepting additional arguments to the function
|
|
|
|
|
// literal as necessary to satisfy fun's signature.
|
|
|
|
|
//
|
|
|
|
|
// If nilCheck is true and arg0 is an interface value, then it's
|
|
|
|
|
// checked to be non-nil as an initial step at the point of evaluating
|
|
|
|
|
// the function literal itself.
|
2023-02-14 15:13:50 -08:00
|
|
|
func (r *reader) curry(origPos src.XPos, ifaceHack bool, fun ir.Node, arg0, arg1 ir.Node) ir.Node {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var captured ir.Nodes
|
|
|
|
|
captured.Append(fun, arg0)
|
|
|
|
|
if arg1 != nil {
|
|
|
|
|
captured.Append(arg1)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
params, results := syntheticSig(fun.Type())
|
|
|
|
|
params = params[len(captured)-1:] // skip curried parameters
|
2022-12-01 17:42:02 -08:00
|
|
|
typ := types.NewSignature(nil, params, results)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
addBody := func(pos src.XPos, r *reader, captured []ir.Node) {
|
|
|
|
|
fun := captured[0]
|
|
|
|
|
|
|
|
|
|
var args ir.Nodes
|
|
|
|
|
args.Append(captured[1:]...)
|
2023-09-13 19:26:32 -07:00
|
|
|
args.Append(r.syntheticArgs()...)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
r.syntheticTailCall(pos, fun, args)
|
|
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
return r.syntheticClosure(origPos, typ, ifaceHack, captured, addBody)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// methodExprWrap returns a function literal that changes method's
|
|
|
|
|
// first parameter's type to recv, and uses implicits/deref/addr to
|
|
|
|
|
// select the appropriate receiver parameter to pass to method.
|
2023-02-14 15:13:50 -08:00
|
|
|
func (r *reader) methodExprWrap(origPos src.XPos, recv *types.Type, implicits []int, deref, addr bool, method, dictPtr ir.Node) ir.Node {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var captured ir.Nodes
|
|
|
|
|
captured.Append(method)
|
|
|
|
|
|
|
|
|
|
params, results := syntheticSig(method.Type())
|
|
|
|
|
|
|
|
|
|
// Change first parameter to recv.
|
|
|
|
|
params[0].Type = recv
|
|
|
|
|
|
|
|
|
|
// If we have a dictionary pointer argument to pass, then omit the
|
|
|
|
|
// underlying method expression's dictionary parameter from the
|
|
|
|
|
// returned signature too.
|
|
|
|
|
if dictPtr != nil {
|
|
|
|
|
captured.Append(dictPtr)
|
|
|
|
|
params = append(params[:1], params[2:]...)
|
|
|
|
|
}
|
|
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
typ := types.NewSignature(nil, params, results)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
addBody := func(pos src.XPos, r *reader, captured []ir.Node) {
|
|
|
|
|
fn := captured[0]
|
2023-09-13 19:26:32 -07:00
|
|
|
args := r.syntheticArgs()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// Rewrite first argument based on implicits/deref/addr.
|
|
|
|
|
{
|
|
|
|
|
arg := args[0]
|
|
|
|
|
for _, ix := range implicits {
|
2023-08-18 19:40:33 -07:00
|
|
|
arg = Implicit(typecheck.DotField(pos, arg, ix))
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
if deref {
|
|
|
|
|
arg = Implicit(Deref(pos, arg.Type().Elem(), arg))
|
|
|
|
|
} else if addr {
|
|
|
|
|
arg = Implicit(Addr(pos, arg))
|
|
|
|
|
}
|
|
|
|
|
args[0] = arg
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Insert dictionary argument, if provided.
|
|
|
|
|
if dictPtr != nil {
|
|
|
|
|
newArgs := make([]ir.Node, len(args)+1)
|
|
|
|
|
newArgs[0] = args[0]
|
|
|
|
|
newArgs[1] = captured[1]
|
|
|
|
|
copy(newArgs[2:], args[1:])
|
|
|
|
|
args = newArgs
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
r.syntheticTailCall(pos, fn, args)
|
|
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
return r.syntheticClosure(origPos, typ, false, captured, addBody)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// syntheticClosure constructs a synthetic function literal for
|
2023-02-14 15:13:50 -08:00
|
|
|
// currying dictionary arguments. origPos is the position used for the
|
|
|
|
|
// closure, which must be a non-inlined position. typ is the function
|
|
|
|
|
// literal's signature type.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
//
|
|
|
|
|
// captures is a list of expressions that need to be evaluated at the
|
|
|
|
|
// point of function literal evaluation and captured by the function
|
|
|
|
|
// literal. If ifaceHack is true and captures[1] is an interface type,
|
|
|
|
|
// it's checked to be non-nil after evaluation.
|
|
|
|
|
//
|
|
|
|
|
// addBody is a callback function to populate the function body. The
|
|
|
|
|
// list of captured values passed back has the captured variables for
|
|
|
|
|
// use within the function literal, corresponding to the expressions
|
|
|
|
|
// in captures.
|
2023-02-14 15:13:50 -08:00
|
|
|
func (r *reader) syntheticClosure(origPos src.XPos, typ *types.Type, ifaceHack bool, captures ir.Nodes, addBody func(pos src.XPos, r *reader, captured []ir.Node)) ir.Node {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// isSafe reports whether n is an expression that we can safely
|
|
|
|
|
// defer to evaluating inside the closure instead, to avoid storing
|
|
|
|
|
// them into the closure.
|
|
|
|
|
//
|
|
|
|
|
// In practice this is always (and only) the wrappee function.
|
|
|
|
|
isSafe := func(n ir.Node) bool {
|
|
|
|
|
if n.Op() == ir.ONAME && n.(*ir.Name).Class == ir.PFUNC {
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
if n.Op() == ir.OMETHEXPR {
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
2024-01-26 17:49:33 -05:00
|
|
|
fn := r.inlClosureFunc(origPos, typ, ir.OCLOSURE)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
fn.SetWrapper(true)
|
|
|
|
|
|
2023-08-16 21:16:29 -07:00
|
|
|
clo := fn.OClosure
|
|
|
|
|
inlPos := clo.Pos()
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var init ir.Nodes
|
|
|
|
|
for i, n := range captures {
|
|
|
|
|
if isSafe(n) {
|
|
|
|
|
continue // skip capture; can reference directly
|
|
|
|
|
}
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
tmp := r.tempCopy(inlPos, n, &init)
|
|
|
|
|
ir.NewClosureVar(origPos, fn, tmp)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// We need to nil check interface receivers at the point of method
|
|
|
|
|
// value evaluation, ugh.
|
|
|
|
|
if ifaceHack && i == 1 && n.Type().IsInterface() {
|
2023-02-14 15:13:50 -08:00
|
|
|
check := ir.NewUnaryExpr(inlPos, ir.OCHECKNIL, ir.NewUnaryExpr(inlPos, ir.OITAB, tmp))
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
init.Append(typecheck.Stmt(check))
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pri := pkgReaderIndex{synthetic: func(pos src.XPos, r *reader) {
|
|
|
|
|
captured := make([]ir.Node, len(captures))
|
|
|
|
|
next := 0
|
|
|
|
|
for i, n := range captures {
|
|
|
|
|
if isSafe(n) {
|
|
|
|
|
captured[i] = n
|
|
|
|
|
} else {
|
|
|
|
|
captured[i] = r.closureVars[next]
|
|
|
|
|
next++
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
assert(next == len(r.closureVars))
|
|
|
|
|
|
2023-02-14 15:13:50 -08:00
|
|
|
addBody(origPos, r, captured)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}}
|
|
|
|
|
bodyReader[fn] = pri
|
|
|
|
|
pri.funcBody(fn)
|
|
|
|
|
|
2023-08-16 21:16:29 -07:00
|
|
|
return ir.InitExpr(init, clo)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// syntheticSig duplicates and returns the params and results lists
|
|
|
|
|
// for sig, but renaming anonymous parameters so they can be assigned
|
|
|
|
|
// ir.Names.
|
|
|
|
|
func syntheticSig(sig *types.Type) (params, results []*types.Field) {
|
|
|
|
|
clone := func(params []*types.Field) []*types.Field {
|
|
|
|
|
res := make([]*types.Field, len(params))
|
|
|
|
|
for i, param := range params {
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
// TODO(mdempsky): It would be nice to preserve the original
|
|
|
|
|
// parameter positions here instead, but at least
|
|
|
|
|
// typecheck.NewMethodType replaces them with base.Pos, making
|
|
|
|
|
// them useless. Worse, the positions copied from base.Pos may
|
|
|
|
|
// have inlining contexts, which we definitely don't want here
|
|
|
|
|
// (e.g., #54625).
|
2023-09-13 19:26:32 -07:00
|
|
|
res[i] = types.NewField(base.AutogeneratedPos, param.Sym, param.Type)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
res[i].SetIsDDD(param.IsDDD())
|
|
|
|
|
}
|
|
|
|
|
return res
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-20 10:05:29 -07:00
|
|
|
return clone(sig.Params()), clone(sig.Results())
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
2022-06-03 12:49:11 -07:00
|
|
|
func (r *reader) optExpr() ir.Node {
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
return r.expr()
|
|
|
|
|
}
|
|
|
|
|
return nil
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// methodExpr reads a method expression reference, and returns three
|
|
|
|
|
// (possibly nil) expressions related to it:
|
|
|
|
|
//
|
|
|
|
|
// baseFn is always non-nil: it's either a function of the appropriate
|
|
|
|
|
// type already, or it has an extra dictionary parameter as the second
|
|
|
|
|
// parameter (i.e., immediately after the promoted receiver
|
|
|
|
|
// parameter).
|
|
|
|
|
//
|
|
|
|
|
// If dictPtr is non-nil, then it's a dictionary argument that must be
|
|
|
|
|
// passed as the second argument to baseFn.
|
|
|
|
|
//
|
|
|
|
|
// If wrapperFn is non-nil, then it's either the same as baseFn (if
|
|
|
|
|
// dictPtr is nil), or it's semantically equivalent to currying baseFn
|
|
|
|
|
// to pass dictPtr. (wrapperFn is nil when dictPtr is an expression
|
|
|
|
|
// that needs to be computed dynamically.)
|
|
|
|
|
//
|
|
|
|
|
// For callers that are creating a call to the returned method, it's
|
|
|
|
|
// best to emit a call to baseFn, and include dictPtr in the arguments
|
|
|
|
|
// list as appropriate.
|
|
|
|
|
//
|
|
|
|
|
// For callers that want to return a method expression without
|
|
|
|
|
// invoking it, they may return wrapperFn if it's non-nil; but
|
|
|
|
|
// otherwise, they need to create their own wrapper.
|
|
|
|
|
func (r *reader) methodExpr() (wrapperFn, baseFn, dictPtr ir.Node) {
|
|
|
|
|
recv := r.typ()
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
sig0 := r.typ()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
pos := r.pos()
|
2023-09-13 20:01:56 -07:00
|
|
|
sym := r.selector()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
// Signature type to return (i.e., recv prepended to the method's
|
|
|
|
|
// normal parameters list).
|
|
|
|
|
sig := typecheck.NewMethodType(sig0, recv)
|
|
|
|
|
|
|
|
|
|
if r.Bool() { // type parameter method expression
|
|
|
|
|
idx := r.Len()
|
|
|
|
|
word := r.dictWord(pos, r.dict.typeParamMethodExprsOffset()+idx)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): If the type parameter was instantiated with an
|
|
|
|
|
// interface type (i.e., embed.IsInterface()), then we could
|
|
|
|
|
// return the OMETHEXPR instead and save an indirection.
|
|
|
|
|
|
|
|
|
|
// We wrote the method expression's entry point PC into the
|
|
|
|
|
// dictionary, but for Go `func` values we need to return a
|
|
|
|
|
// closure (i.e., pointer to a structure with the PC as the first
|
|
|
|
|
// field). Because method expressions don't have any closure
|
|
|
|
|
// variables, we pun the dictionary entry as the closure struct.
|
|
|
|
|
fn := typecheck.Expr(ir.NewConvExpr(pos, ir.OCONVNOP, sig, ir.NewAddrExpr(pos, word)))
|
|
|
|
|
return fn, fn, nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): I'm pretty sure this isn't needed: implicits is
|
|
|
|
|
// only relevant to locally defined types, but they can't have
|
|
|
|
|
// (non-promoted) methods.
|
|
|
|
|
var implicits []*types.Type
|
|
|
|
|
if r.dict != nil {
|
|
|
|
|
implicits = r.dict.targs
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if r.Bool() { // dynamic subdictionary
|
|
|
|
|
idx := r.Len()
|
|
|
|
|
info := r.dict.subdicts[idx]
|
|
|
|
|
explicits := r.p.typListIdx(info.explicits, r.dict)
|
|
|
|
|
|
|
|
|
|
shapedObj := r.p.objIdx(info.idx, implicits, explicits, true).(*ir.Name)
|
|
|
|
|
shapedFn := shapedMethodExpr(pos, shapedObj, sym)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Is there a more robust way to get the
|
|
|
|
|
// dictionary pointer type here?
|
2023-08-20 10:05:29 -07:00
|
|
|
dictPtrType := shapedFn.Type().Param(1).Type
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
dictPtr := typecheck.Expr(ir.NewConvExpr(pos, ir.OCONVNOP, dictPtrType, r.dictWord(pos, r.dict.subdictsOffset()+idx)))
|
|
|
|
|
|
|
|
|
|
return nil, shapedFn, dictPtr
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if r.Bool() { // static dictionary
|
|
|
|
|
info := r.objInfo()
|
|
|
|
|
explicits := r.p.typListIdx(info.explicits, r.dict)
|
|
|
|
|
|
|
|
|
|
shapedObj := r.p.objIdx(info.idx, implicits, explicits, true).(*ir.Name)
|
|
|
|
|
shapedFn := shapedMethodExpr(pos, shapedObj, sym)
|
|
|
|
|
|
|
|
|
|
dict := r.p.objDictName(info.idx, implicits, explicits)
|
|
|
|
|
dictPtr := typecheck.Expr(ir.NewAddrExpr(pos, dict))
|
|
|
|
|
|
|
|
|
|
// Check that dictPtr matches shapedFn's dictionary parameter.
|
2023-08-20 10:05:29 -07:00
|
|
|
if !types.Identical(dictPtr.Type(), shapedFn.Type().Param(1).Type) {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
base.FatalfAt(pos, "dict %L, but shaped method %L", dict, shapedFn)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// For statically known instantiations, we can take advantage of
|
|
|
|
|
// the stenciled wrapper.
|
|
|
|
|
base.AssertfAt(!recv.HasShape(), pos, "shaped receiver %v", recv)
|
2023-08-18 18:03:24 -07:00
|
|
|
wrapperFn := typecheck.NewMethodExpr(pos, recv, sym)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
base.AssertfAt(types.Identical(sig, wrapperFn.Type()), pos, "wrapper %L does not have type %v", wrapperFn, sig)
|
|
|
|
|
|
|
|
|
|
return wrapperFn, shapedFn, dictPtr
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Simple method expression; no dictionary needed.
|
|
|
|
|
base.AssertfAt(!recv.HasShape() || recv.IsInterface(), pos, "shaped receiver %v", recv)
|
2023-08-18 18:03:24 -07:00
|
|
|
fn := typecheck.NewMethodExpr(pos, recv, sym)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
return fn, fn, nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// shapedMethodExpr returns the specified method on the given shaped
|
|
|
|
|
// type.
|
|
|
|
|
func shapedMethodExpr(pos src.XPos, obj *ir.Name, sym *types.Sym) *ir.SelectorExpr {
|
|
|
|
|
assert(obj.Op() == ir.OTYPE)
|
|
|
|
|
|
|
|
|
|
typ := obj.Type()
|
|
|
|
|
assert(typ.HasShape())
|
|
|
|
|
|
|
|
|
|
method := func() *types.Field {
|
2023-08-20 10:05:29 -07:00
|
|
|
for _, method := range typ.Methods() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if method.Sym == sym {
|
|
|
|
|
return method
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
base.FatalfAt(pos, "failed to find method %v in shaped type %v", sym, typ)
|
|
|
|
|
panic("unreachable")
|
|
|
|
|
}()
|
|
|
|
|
|
|
|
|
|
// Construct an OMETHEXPR node.
|
|
|
|
|
recv := method.Type.Recv().Type
|
2023-08-18 18:03:24 -07:00
|
|
|
return typecheck.NewMethodExpr(pos, recv, sym)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
2022-06-24 14:19:06 -07:00
|
|
|
func (r *reader) multiExpr() []ir.Node {
|
|
|
|
|
r.Sync(pkgbits.SyncMultiExpr)
|
|
|
|
|
|
[dev.unified] cmd/compile/internal/noder: implicit conversions for multi-valued expressions
This CL changes GOEXPERIMENT=unified to insert implicit conversions
for multi-valued expressions.
Unfortunately, IR doesn't have strong, first-class support for
multi-valued expressions, so this CL takes the approach of spilling
them to temporary variables, which can then be implicitly converted.
This is the same approach taken by walk, but doing it this early does
introduce some minor complications:
1. For select case clauses with comma-ok assignments (e.g., `case x,
ok := <-ch:`), the compiler middle end wants to see the OAS2RECV
assignment is the CommClause.Comm statement. So when constructing
select statements, we need to massage this around a little.
2. The extra temporary variables and assignments skew the existing
inlining heuristics. As mentioned, the temporaries/assignments will
eventually be added (and often optimized away again) anyway, but now
they're visible to the inliner. So this CL also kludges the inlining
heuristics in this case to keep things comparable.
Change-Id: I3e3ea756ad92472ebe28bae3963be61ed7684a75
Reviewed-on: https://go-review.googlesource.com/c/go/+/415244
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-06-28 16:31:29 -07:00
|
|
|
if r.Bool() { // N:1
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
expr := r.expr()
|
|
|
|
|
|
|
|
|
|
results := make([]ir.Node, r.Len())
|
|
|
|
|
as := ir.NewAssignListStmt(pos, ir.OAS2, nil, []ir.Node{expr})
|
|
|
|
|
as.Def = true
|
|
|
|
|
for i := range results {
|
2022-07-18 12:41:56 -07:00
|
|
|
tmp := r.temp(pos, r.typ())
|
[dev.unified] cmd/compile/internal/noder: implicit conversions for multi-valued expressions
This CL changes GOEXPERIMENT=unified to insert implicit conversions
for multi-valued expressions.
Unfortunately, IR doesn't have strong, first-class support for
multi-valued expressions, so this CL takes the approach of spilling
them to temporary variables, which can then be implicitly converted.
This is the same approach taken by walk, but doing it this early does
introduce some minor complications:
1. For select case clauses with comma-ok assignments (e.g., `case x,
ok := <-ch:`), the compiler middle end wants to see the OAS2RECV
assignment is the CommClause.Comm statement. So when constructing
select statements, we need to massage this around a little.
2. The extra temporary variables and assignments skew the existing
inlining heuristics. As mentioned, the temporaries/assignments will
eventually be added (and often optimized away again) anyway, but now
they're visible to the inliner. So this CL also kludges the inlining
heuristics in this case to keep things comparable.
Change-Id: I3e3ea756ad92472ebe28bae3963be61ed7684a75
Reviewed-on: https://go-review.googlesource.com/c/go/+/415244
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-06-28 16:31:29 -07:00
|
|
|
as.PtrInit().Append(ir.NewDecl(pos, ir.ODCL, tmp))
|
|
|
|
|
as.Lhs.Append(tmp)
|
|
|
|
|
|
|
|
|
|
res := ir.Node(tmp)
|
|
|
|
|
if r.Bool() {
|
2022-06-30 17:58:47 -07:00
|
|
|
n := ir.NewConvExpr(pos, ir.OCONV, r.typ(), res)
|
2022-07-23 00:46:10 -07:00
|
|
|
n.TypeWord, n.SrcRType = r.convRTTI(pos)
|
2022-06-30 17:58:47 -07:00
|
|
|
n.SetImplicit(true)
|
|
|
|
|
res = typecheck.Expr(n)
|
[dev.unified] cmd/compile/internal/noder: implicit conversions for multi-valued expressions
This CL changes GOEXPERIMENT=unified to insert implicit conversions
for multi-valued expressions.
Unfortunately, IR doesn't have strong, first-class support for
multi-valued expressions, so this CL takes the approach of spilling
them to temporary variables, which can then be implicitly converted.
This is the same approach taken by walk, but doing it this early does
introduce some minor complications:
1. For select case clauses with comma-ok assignments (e.g., `case x,
ok := <-ch:`), the compiler middle end wants to see the OAS2RECV
assignment is the CommClause.Comm statement. So when constructing
select statements, we need to massage this around a little.
2. The extra temporary variables and assignments skew the existing
inlining heuristics. As mentioned, the temporaries/assignments will
eventually be added (and often optimized away again) anyway, but now
they're visible to the inliner. So this CL also kludges the inlining
heuristics in this case to keep things comparable.
Change-Id: I3e3ea756ad92472ebe28bae3963be61ed7684a75
Reviewed-on: https://go-review.googlesource.com/c/go/+/415244
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-06-28 16:31:29 -07:00
|
|
|
}
|
|
|
|
|
results[i] = res
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Could use ir.InlinedCallExpr instead?
|
|
|
|
|
results[0] = ir.InitExpr([]ir.Node{typecheck.Stmt(as)}, results[0])
|
|
|
|
|
return results
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// N:N
|
2022-06-24 14:19:06 -07:00
|
|
|
exprs := make([]ir.Node, r.Len())
|
|
|
|
|
if len(exprs) == 0 {
|
|
|
|
|
return nil
|
|
|
|
|
}
|
|
|
|
|
for i := range exprs {
|
|
|
|
|
exprs[i] = r.expr()
|
|
|
|
|
}
|
|
|
|
|
return exprs
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-18 12:41:56 -07:00
|
|
|
// temp returns a new autotemp of the specified type.
|
|
|
|
|
func (r *reader) temp(pos src.XPos, typ *types.Type) *ir.Name {
|
2023-08-24 03:04:01 -07:00
|
|
|
return typecheck.TempAt(pos, r.curfn, typ)
|
2022-07-18 12:41:56 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
// tempCopy declares and returns a new autotemp initialized to the
|
|
|
|
|
// value of expr.
|
|
|
|
|
func (r *reader) tempCopy(pos src.XPos, expr ir.Node, init *ir.Nodes) *ir.Name {
|
|
|
|
|
tmp := r.temp(pos, expr.Type())
|
|
|
|
|
|
|
|
|
|
init.Append(typecheck.Stmt(ir.NewDecl(pos, ir.ODCL, tmp)))
|
|
|
|
|
|
|
|
|
|
assign := ir.NewAssignStmt(pos, tmp, expr)
|
|
|
|
|
assign.Def = true
|
|
|
|
|
init.Append(typecheck.Stmt(ir.NewAssignStmt(pos, tmp, expr)))
|
|
|
|
|
|
|
|
|
|
tmp.Defn = assign
|
|
|
|
|
|
|
|
|
|
return tmp
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) compLit() ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncCompLit)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.pos()
|
2021-07-09 19:47:45 -07:00
|
|
|
typ0 := r.typ()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-07-09 19:47:45 -07:00
|
|
|
typ := typ0
|
|
|
|
|
if typ.IsPtr() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
typ = typ.Elem()
|
|
|
|
|
}
|
|
|
|
|
if typ.Kind() == types.TFORW {
|
|
|
|
|
base.FatalfAt(pos, "unresolved composite literal type: %v", typ)
|
|
|
|
|
}
|
2022-07-22 23:33:30 -07:00
|
|
|
var rtype ir.Node
|
|
|
|
|
if typ.IsMap() {
|
|
|
|
|
rtype = r.rtype(pos)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
isStruct := typ.Kind() == types.TSTRUCT
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
elems := make([]ir.Node, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range elems {
|
|
|
|
|
elemp := &elems[i]
|
|
|
|
|
|
|
|
|
|
if isStruct {
|
2022-02-14 09:41:19 -08:00
|
|
|
sk := ir.NewStructKeyExpr(r.pos(), typ.Field(r.Len()), nil)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
*elemp, elemp = sk, &sk.Value
|
2022-02-14 09:41:19 -08:00
|
|
|
} else if r.Bool() {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
kv := ir.NewKeyExpr(r.pos(), r.expr(), nil)
|
|
|
|
|
*elemp, elemp = kv, &kv.Value
|
|
|
|
|
}
|
|
|
|
|
|
2024-03-20 11:25:45 -07:00
|
|
|
*elemp = r.expr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2022-05-02 18:02:22 -07:00
|
|
|
lit := typecheck.Expr(ir.NewCompLitExpr(pos, ir.OCOMPLIT, typ, elems))
|
2022-07-22 23:33:30 -07:00
|
|
|
if rtype != nil {
|
2022-06-21 02:29:49 -07:00
|
|
|
lit := lit.(*ir.CompLitExpr)
|
2022-07-22 23:33:30 -07:00
|
|
|
lit.RType = rtype
|
2022-06-21 02:29:49 -07:00
|
|
|
}
|
2021-07-09 19:47:45 -07:00
|
|
|
if typ0.IsPtr() {
|
|
|
|
|
lit = typecheck.Expr(typecheck.NodAddrAt(pos, lit))
|
|
|
|
|
lit.SetType(typ0)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
return lit
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) funcLit() ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncFuncLit)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
// The underlying function declaration (including its parameters'
|
|
|
|
|
// positions, if any) need to remain the original, uninlined
|
|
|
|
|
// positions. This is because we track inlining-context on nodes so
|
|
|
|
|
// we can synthesize the extra implied stack frames dynamically when
|
|
|
|
|
// generating tracebacks, whereas those stack frames don't make
|
|
|
|
|
// sense *within* the function literal. (Any necessary inlining
|
|
|
|
|
// adjustments will have been applied to the call expression
|
|
|
|
|
// instead.)
|
|
|
|
|
//
|
|
|
|
|
// This is subtle, and getting it wrong leads to cycles in the
|
|
|
|
|
// inlining tree, which lead to infinite loops during stack
|
|
|
|
|
// unwinding (#46234, #54625).
|
|
|
|
|
//
|
|
|
|
|
// Note that we *do* want the inline-adjusted position for the
|
|
|
|
|
// OCLOSURE node, because that position represents where any heap
|
|
|
|
|
// allocation of the closure is credited (#49171).
|
|
|
|
|
r.suppressInlPos++
|
2023-08-16 21:16:29 -07:00
|
|
|
origPos := r.pos()
|
|
|
|
|
sig := r.signature(nil)
|
cmd/compile/internal/noder: fix inlined function literal positions
When inlining function calls, we rewrite the position information on
all of the nodes to keep track of the inlining context. This is
necessary so that at runtime, we can synthesize additional stack
frames so that the inlining is transparent to the user.
However, for function literals, we *don't* want to apply this
rewriting to the underlying function. Because within the function
literal (when it's not itself inlined), the inlining context (if any)
will have already be available at the caller PC instead.
Unified IR was already getting this right in the case of user-written
statements within the function literal, which is what the unit test
for #46234 tested. However, it was still using inline-adjusted
positions for the function declaration and its parameters, which
occasionally end up getting used for generated code (e.g., loading
captured values from the closure record).
I've manually verified that this fixes the hang in
https://go.dev/play/p/avQ0qgRzOgt, and spot-checked the
-d=pctab=pctoinline output for kube-apiserver and kubelet and they
seem better.
However, I'm still working on a more robust test for this (hence
"Updates" not "Fixes") and internal assertions to verify that we're
emitting correct inline trees. In particular, there are still other
cases (even in the non-unified frontend) where we're producing
corrupt (but at least acyclic) inline trees.
Updates #54625.
Change-Id: Iacfd2e1eb06ae8dc299c0679f377461d3d46c15a
Reviewed-on: https://go-review.googlesource.com/c/go/+/425395
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
2022-08-24 15:03:50 -07:00
|
|
|
r.suppressInlPos--
|
2024-01-26 17:49:33 -05:00
|
|
|
why := ir.OCLOSURE
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
why = ir.ORANGE
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2024-01-26 17:49:33 -05:00
|
|
|
fn := r.inlClosureFunc(origPos, sig, why)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
fn.ClosureVars = make([]*ir.Name, 0, r.Len())
|
2021-06-22 18:10:59 -07:00
|
|
|
for len(fn.ClosureVars) < cap(fn.ClosureVars) {
|
2023-08-16 21:16:29 -07:00
|
|
|
// TODO(mdempsky): I think these should be original positions too
|
|
|
|
|
// (i.e., not inline-adjusted).
|
2021-07-02 16:29:42 -07:00
|
|
|
ir.NewClosureVar(r.pos(), fn, r.useLocal())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2022-07-31 18:48:16 -07:00
|
|
|
if param := r.dictParam; param != nil {
|
|
|
|
|
// If we have a dictionary parameter, capture it too. For
|
|
|
|
|
// simplicity, we capture it last and unconditionally.
|
|
|
|
|
ir.NewClosureVar(param.Pos(), fn, param)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
r.addBody(fn, nil)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-08-16 18:56:41 -07:00
|
|
|
return fn.OClosure
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-16 21:16:29 -07:00
|
|
|
// inlClosureFunc constructs a new closure function, but correctly
|
|
|
|
|
// handles inlining.
|
2024-01-26 17:49:33 -05:00
|
|
|
func (r *reader) inlClosureFunc(origPos src.XPos, sig *types.Type, why ir.Op) *ir.Func {
|
2023-08-16 21:16:29 -07:00
|
|
|
curfn := r.inlCaller
|
|
|
|
|
if curfn == nil {
|
|
|
|
|
curfn = r.curfn
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Remove hard-coding of typecheck.Target.
|
2024-01-26 17:49:33 -05:00
|
|
|
return ir.NewClosureFunc(origPos, r.inlPos(origPos), why, sig, curfn, typecheck.Target)
|
2023-08-16 21:16:29 -07:00
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) exprList() []ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncExprList)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
return r.exprs()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) exprs() []ir.Node {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncExprs)
|
|
|
|
|
nodes := make([]ir.Node, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if len(nodes) == 0 {
|
|
|
|
|
return nil // TODO(mdempsky): Unclear if this matters.
|
|
|
|
|
}
|
|
|
|
|
for i := range nodes {
|
|
|
|
|
nodes[i] = r.expr()
|
|
|
|
|
}
|
|
|
|
|
return nodes
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-31 18:48:16 -07:00
|
|
|
// dictWord returns an expression to return the specified
|
|
|
|
|
// uintptr-typed word from the dictionary parameter.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (r *reader) dictWord(pos src.XPos, idx int) ir.Node {
|
2022-07-31 18:48:16 -07:00
|
|
|
base.AssertfAt(r.dictParam != nil, pos, "expected dictParam in %v", r.curfn)
|
2023-09-08 15:44:57 -07:00
|
|
|
return typecheck.Expr(ir.NewIndexExpr(pos, r.dictParam, ir.NewInt(pos, int64(idx))))
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// rttiWord is like dictWord, but converts it to *byte (the type used
|
|
|
|
|
// internally to represent *runtime._type and *runtime.itab).
|
|
|
|
|
func (r *reader) rttiWord(pos src.XPos, idx int) ir.Node {
|
|
|
|
|
return typecheck.Expr(ir.NewConvExpr(pos, ir.OCONVNOP, types.NewPtr(types.Types[types.TUINT8]), r.dictWord(pos, idx)))
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// rtype reads a type reference from the element bitstream, and
|
|
|
|
|
// returns an expression of type *runtime._type representing that
|
|
|
|
|
// type.
|
2022-07-22 23:33:30 -07:00
|
|
|
func (r *reader) rtype(pos src.XPos) ir.Node {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
_, rtype := r.rtype0(pos)
|
|
|
|
|
return rtype
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (r *reader) rtype0(pos src.XPos) (typ *types.Type, rtype ir.Node) {
|
|
|
|
|
r.Sync(pkgbits.SyncRType)
|
|
|
|
|
if r.Bool() { // derived type
|
|
|
|
|
idx := r.Len()
|
|
|
|
|
info := r.dict.rtypes[idx]
|
|
|
|
|
typ = r.p.typIdx(info, r.dict, true)
|
|
|
|
|
rtype = r.rttiWord(pos, r.dict.rtypesOffset()+idx)
|
|
|
|
|
return
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
|
|
|
|
|
typ = r.typ()
|
|
|
|
|
rtype = reflectdata.TypePtrAt(pos, typ)
|
|
|
|
|
return
|
2022-07-22 23:33:30 -07:00
|
|
|
}
|
|
|
|
|
|
2022-08-18 03:53:58 -07:00
|
|
|
// varDictIndex populates name.DictIndex if name is a derived type.
|
|
|
|
|
func (r *reader) varDictIndex(name *ir.Name) {
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
idx := 1 + r.dict.rtypesOffset() + r.Len()
|
|
|
|
|
if int(uint16(idx)) != idx {
|
|
|
|
|
base.FatalfAt(name.Pos(), "DictIndex overflow for %v: %v", name, idx)
|
|
|
|
|
}
|
|
|
|
|
name.DictIndex = uint16(idx)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-09-01 16:58:15 -07:00
|
|
|
// itab returns a (typ, iface) pair of types.
|
|
|
|
|
//
|
|
|
|
|
// typRType and ifaceRType are expressions that evaluate to the
|
|
|
|
|
// *runtime._type for typ and iface, respectively.
|
|
|
|
|
//
|
|
|
|
|
// If typ is a concrete type and iface is a non-empty interface type,
|
|
|
|
|
// then itab is an expression that evaluates to the *runtime.itab for
|
|
|
|
|
// the pair. Otherwise, itab is nil.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
func (r *reader) itab(pos src.XPos) (typ *types.Type, typRType ir.Node, iface *types.Type, ifaceRType ir.Node, itab ir.Node) {
|
2022-09-01 16:58:15 -07:00
|
|
|
typ, typRType = r.rtype0(pos)
|
|
|
|
|
iface, ifaceRType = r.rtype0(pos)
|
|
|
|
|
|
|
|
|
|
idx := -1
|
|
|
|
|
if r.Bool() {
|
|
|
|
|
idx = r.Len()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
}
|
|
|
|
|
|
2022-09-01 16:58:15 -07:00
|
|
|
if !typ.IsInterface() && iface.IsInterface() && !iface.IsEmptyInterface() {
|
|
|
|
|
if idx >= 0 {
|
|
|
|
|
itab = r.rttiWord(pos, r.dict.itabsOffset()+idx)
|
|
|
|
|
} else {
|
|
|
|
|
base.AssertfAt(!typ.HasShape(), pos, "%v is a shape type", typ)
|
|
|
|
|
base.AssertfAt(!iface.HasShape(), pos, "%v is a shape type", iface)
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
lsym := reflectdata.ITabLsym(typ, iface)
|
|
|
|
|
itab = typecheck.LinksymAddr(pos, lsym, types.Types[types.TUINT8])
|
|
|
|
|
}
|
|
|
|
|
}
|
2022-09-01 16:58:15 -07:00
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
return
|
2022-08-06 15:27:30 -07:00
|
|
|
}
|
|
|
|
|
|
2022-07-23 00:46:10 -07:00
|
|
|
// convRTTI returns expressions appropriate for populating an
|
|
|
|
|
// ir.ConvExpr's TypeWord and SrcRType fields, respectively.
|
|
|
|
|
func (r *reader) convRTTI(pos src.XPos) (typeWord, srcRType ir.Node) {
|
|
|
|
|
r.Sync(pkgbits.SyncConvRTTI)
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
src, srcRType0, dst, dstRType, itab := r.itab(pos)
|
2022-07-23 00:46:10 -07:00
|
|
|
if !dst.IsInterface() {
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// See reflectdata.ConvIfaceTypeWord.
|
|
|
|
|
switch {
|
|
|
|
|
case dst.IsEmptyInterface():
|
|
|
|
|
if !src.IsInterface() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
typeWord = srcRType0 // direct eface construction
|
2022-07-23 00:46:10 -07:00
|
|
|
}
|
|
|
|
|
case !src.IsInterface():
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
typeWord = itab // direct iface construction
|
2022-07-23 00:46:10 -07:00
|
|
|
default:
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
typeWord = dstRType // convI2I
|
2022-07-23 00:46:10 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// See reflectdata.ConvIfaceSrcRType.
|
|
|
|
|
if !src.IsInterface() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
srcRType = srcRType0
|
2022-07-23 00:46:10 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-25 12:21:21 -07:00
|
|
|
func (r *reader) exprType() ir.Node {
|
2022-03-06 23:47:27 -08:00
|
|
|
r.Sync(pkgbits.SyncExprType)
|
|
|
|
|
pos := r.pos()
|
|
|
|
|
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
var typ *types.Type
|
2022-05-09 17:19:58 -07:00
|
|
|
var rtype, itab ir.Node
|
2022-03-06 23:47:27 -08:00
|
|
|
|
2022-03-07 03:35:14 -08:00
|
|
|
if r.Bool() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
typ, rtype, _, _, itab = r.itab(pos)
|
2022-09-01 16:58:15 -07:00
|
|
|
if !typ.IsInterface() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
rtype = nil // TODO(mdempsky): Leave set?
|
2022-05-09 17:19:58 -07:00
|
|
|
}
|
2022-03-07 03:35:14 -08:00
|
|
|
} else {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
typ, rtype = r.rtype0(pos)
|
|
|
|
|
|
|
|
|
|
if !r.Bool() { // not derived
|
2023-08-17 23:23:40 -07:00
|
|
|
return ir.TypeNode(typ)
|
2022-03-06 23:47:27 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2022-05-09 17:19:58 -07:00
|
|
|
dt := ir.NewDynamicType(pos, rtype)
|
|
|
|
|
dt.ITab = itab
|
2024-10-01 02:30:39 +07:00
|
|
|
dt = typed(typ, dt).(*ir.DynamicType)
|
|
|
|
|
if st := dt.ToStatic(); st != nil {
|
|
|
|
|
return st
|
|
|
|
|
}
|
|
|
|
|
return dt
|
2022-03-06 23:47:27 -08:00
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func (r *reader) op() ir.Op {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncOp)
|
|
|
|
|
return ir.Op(r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Package initialization
|
|
|
|
|
|
|
|
|
|
func (r *reader) pkgInit(self *types.Pkg, target *ir.Package) {
|
2022-02-14 09:41:19 -08:00
|
|
|
cgoPragmas := make([][]string, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range cgoPragmas {
|
2022-02-14 09:41:19 -08:00
|
|
|
cgoPragmas[i] = r.Strings()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
target.CgoPragmas = cgoPragmas
|
|
|
|
|
|
2023-08-24 03:04:01 -07:00
|
|
|
r.pkgInitOrder(target)
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.pkgDecls(target)
|
|
|
|
|
|
2023-08-24 03:04:01 -07:00
|
|
|
r.Sync(pkgbits.SyncEOF)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// pkgInitOrder creates a synthetic init function to handle any
|
|
|
|
|
// package-scope initialization statements.
|
|
|
|
|
func (r *reader) pkgInitOrder(target *ir.Package) {
|
2023-08-09 01:43:47 -07:00
|
|
|
initOrder := make([]ir.Node, r.Len())
|
2023-08-24 03:04:01 -07:00
|
|
|
if len(initOrder) == 0 {
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Make a function that contains all the initialization statements.
|
|
|
|
|
pos := base.AutogeneratedPos
|
|
|
|
|
base.Pos = pos
|
|
|
|
|
|
|
|
|
|
fn := ir.NewFunc(pos, pos, typecheck.Lookup("init"), types.NewSignature(nil, nil, nil))
|
|
|
|
|
fn.SetIsPackageInit(true)
|
|
|
|
|
fn.SetInlinabilityChecked(true) // suppress useless "can inline" diagnostics
|
|
|
|
|
|
|
|
|
|
typecheck.DeclFunc(fn)
|
|
|
|
|
r.curfn = fn
|
|
|
|
|
|
2023-08-09 01:43:47 -07:00
|
|
|
for i := range initOrder {
|
|
|
|
|
lhs := make([]ir.Node, r.Len())
|
|
|
|
|
for j := range lhs {
|
|
|
|
|
lhs[j] = r.obj()
|
|
|
|
|
}
|
|
|
|
|
rhs := r.expr()
|
|
|
|
|
pos := lhs[0].Pos()
|
|
|
|
|
|
|
|
|
|
var as ir.Node
|
|
|
|
|
if len(lhs) == 1 {
|
|
|
|
|
as = typecheck.Stmt(ir.NewAssignStmt(pos, lhs[0], rhs))
|
|
|
|
|
} else {
|
|
|
|
|
as = typecheck.Stmt(ir.NewAssignListStmt(pos, ir.OAS2, lhs, []ir.Node{rhs}))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for _, v := range lhs {
|
|
|
|
|
v.(*ir.Name).Defn = as
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
initOrder[i] = as
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-24 03:04:01 -07:00
|
|
|
fn.Body = initOrder
|
|
|
|
|
|
|
|
|
|
typecheck.FinishFuncBody()
|
|
|
|
|
r.curfn = nil
|
|
|
|
|
r.locals = nil
|
|
|
|
|
|
|
|
|
|
// Outline (if legal/profitable) global map inits.
|
|
|
|
|
staticinit.OutlineMapInits(fn)
|
|
|
|
|
|
|
|
|
|
target.Inits = append(target.Inits, fn)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) pkgDecls(target *ir.Package) {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncDecls)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for {
|
2022-02-14 09:41:19 -08:00
|
|
|
switch code := codeDecl(r.Code(pkgbits.SyncDecl)); code {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
default:
|
|
|
|
|
panic(fmt.Sprintf("unhandled decl: %v", code))
|
|
|
|
|
|
|
|
|
|
case declEnd:
|
|
|
|
|
return
|
|
|
|
|
|
|
|
|
|
case declFunc:
|
|
|
|
|
names := r.pkgObjs(target)
|
|
|
|
|
assert(len(names) == 1)
|
2023-08-09 08:39:47 -07:00
|
|
|
target.Funcs = append(target.Funcs, names[0].Func)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case declMethod:
|
|
|
|
|
typ := r.typ()
|
2023-09-13 20:01:56 -07:00
|
|
|
sym := r.selector()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
method := typecheck.Lookdot1(nil, sym, typ, typ.Methods(), 0)
|
2023-08-09 08:39:47 -07:00
|
|
|
target.Funcs = append(target.Funcs, method.Nname.(*ir.Name).Func)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
case declVar:
|
|
|
|
|
names := r.pkgObjs(target)
|
|
|
|
|
|
2022-02-14 09:41:19 -08:00
|
|
|
if n := r.Len(); n > 0 {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
assert(len(names) == 1)
|
|
|
|
|
embeds := make([]ir.Embed, n)
|
|
|
|
|
for i := range embeds {
|
2022-02-14 09:41:19 -08:00
|
|
|
embeds[i] = ir.Embed{Pos: r.pos(), Patterns: r.Strings()}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
names[0].Embed = &embeds
|
|
|
|
|
target.Embeds = append(target.Embeds, names[0])
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case declOther:
|
|
|
|
|
r.pkgObjs(target)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (r *reader) pkgObjs(target *ir.Package) []*ir.Name {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncDeclNames)
|
|
|
|
|
nodes := make([]*ir.Name, r.Len())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for i := range nodes {
|
2022-02-14 09:41:19 -08:00
|
|
|
r.Sync(pkgbits.SyncDeclName)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
name := r.obj().(*ir.Name)
|
|
|
|
|
nodes[i] = name
|
|
|
|
|
|
|
|
|
|
sym := name.Sym()
|
|
|
|
|
if sym.IsBlank() {
|
|
|
|
|
continue
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch name.Class {
|
|
|
|
|
default:
|
|
|
|
|
base.FatalfAt(name.Pos(), "unexpected class: %v", name.Class)
|
|
|
|
|
|
|
|
|
|
case ir.PEXTERN:
|
|
|
|
|
target.Externs = append(target.Externs, name)
|
|
|
|
|
|
|
|
|
|
case ir.PFUNC:
|
|
|
|
|
assert(name.Type().Recv() == nil)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Cleaner way to recognize init?
|
|
|
|
|
if strings.HasPrefix(sym.Name, "init.") {
|
|
|
|
|
target.Inits = append(target.Inits, name.Func)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile: simplify asmhdr and plugin exports handling
This CL removes a bunch of obsolete code, which made the overall
possible data flow of the compiler much harder to understand. In
particular, it:
1. Removes typecheck.Declare by inlining its only two remaining uses,
and simplifying them down to just the couple of relevant assignments
for each remaining caller.
2. Renames ir.Package.{Asms,Exports} to {AsmHdrDecls,PluginExports},
respectively, to better describe what they're used for. In particular,
PluginExports now actually holds only the subset of Exports that used
to be confusingly called "ptabs" in package reflectdata.
3. Renames reflectdata.WriteTabs to reflectdata.WritePluginTable, to
make it clearer what it does.
4. Removes the consistency checks on len(Exports) and len(ptabs),
since now it's plainly obvious that only the unified importer ever
appends to PluginExports.
Change-Id: Iedc9d0a4e7648de4e734f7e3e7df302580fed542
Reviewed-on: https://go-review.googlesource.com/c/go/+/518757
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
2023-08-11 12:04:22 -07:00
|
|
|
if base.Ctxt.Flag_dynlink && types.LocalPkg.Name == "main" && types.IsExported(sym.Name) && name.Op() == ir.ONAME {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
assert(!sym.OnExportList())
|
cmd/compile: simplify asmhdr and plugin exports handling
This CL removes a bunch of obsolete code, which made the overall
possible data flow of the compiler much harder to understand. In
particular, it:
1. Removes typecheck.Declare by inlining its only two remaining uses,
and simplifying them down to just the couple of relevant assignments
for each remaining caller.
2. Renames ir.Package.{Asms,Exports} to {AsmHdrDecls,PluginExports},
respectively, to better describe what they're used for. In particular,
PluginExports now actually holds only the subset of Exports that used
to be confusingly called "ptabs" in package reflectdata.
3. Renames reflectdata.WriteTabs to reflectdata.WritePluginTable, to
make it clearer what it does.
4. Removes the consistency checks on len(Exports) and len(ptabs),
since now it's plainly obvious that only the unified importer ever
appends to PluginExports.
Change-Id: Iedc9d0a4e7648de4e734f7e3e7df302580fed542
Reviewed-on: https://go-review.googlesource.com/c/go/+/518757
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
2023-08-11 12:04:22 -07:00
|
|
|
target.PluginExports = append(target.PluginExports, name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
sym.SetOnExportList(true)
|
|
|
|
|
}
|
|
|
|
|
|
cmd/compile: simplify asmhdr and plugin exports handling
This CL removes a bunch of obsolete code, which made the overall
possible data flow of the compiler much harder to understand. In
particular, it:
1. Removes typecheck.Declare by inlining its only two remaining uses,
and simplifying them down to just the couple of relevant assignments
for each remaining caller.
2. Renames ir.Package.{Asms,Exports} to {AsmHdrDecls,PluginExports},
respectively, to better describe what they're used for. In particular,
PluginExports now actually holds only the subset of Exports that used
to be confusingly called "ptabs" in package reflectdata.
3. Renames reflectdata.WriteTabs to reflectdata.WritePluginTable, to
make it clearer what it does.
4. Removes the consistency checks on len(Exports) and len(ptabs),
since now it's plainly obvious that only the unified importer ever
appends to PluginExports.
Change-Id: Iedc9d0a4e7648de4e734f7e3e7df302580fed542
Reviewed-on: https://go-review.googlesource.com/c/go/+/518757
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
2023-08-11 12:04:22 -07:00
|
|
|
if base.Flag.AsmHdr != "" && (name.Op() == ir.OLITERAL || name.Op() == ir.OTYPE) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
assert(!sym.Asm())
|
cmd/compile: simplify asmhdr and plugin exports handling
This CL removes a bunch of obsolete code, which made the overall
possible data flow of the compiler much harder to understand. In
particular, it:
1. Removes typecheck.Declare by inlining its only two remaining uses,
and simplifying them down to just the couple of relevant assignments
for each remaining caller.
2. Renames ir.Package.{Asms,Exports} to {AsmHdrDecls,PluginExports},
respectively, to better describe what they're used for. In particular,
PluginExports now actually holds only the subset of Exports that used
to be confusingly called "ptabs" in package reflectdata.
3. Renames reflectdata.WriteTabs to reflectdata.WritePluginTable, to
make it clearer what it does.
4. Removes the consistency checks on len(Exports) and len(ptabs),
since now it's plainly obvious that only the unified importer ever
appends to PluginExports.
Change-Id: Iedc9d0a4e7648de4e734f7e3e7df302580fed542
Reviewed-on: https://go-review.googlesource.com/c/go/+/518757
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
2023-08-11 12:04:22 -07:00
|
|
|
target.AsmHdrDecls = append(target.AsmHdrDecls, name)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
sym.SetAsm(true)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return nodes
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Inlining
|
|
|
|
|
|
2022-08-31 13:48:06 -07:00
|
|
|
// unifiedHaveInlineBody reports whether we have the function body for
|
|
|
|
|
// fn, so we can inline it.
|
|
|
|
|
func unifiedHaveInlineBody(fn *ir.Func) bool {
|
|
|
|
|
if fn.Inl == nil {
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
_, ok := bodyReaderFor(fn)
|
|
|
|
|
return ok
|
|
|
|
|
}
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var inlgen = 0
|
|
|
|
|
|
2022-08-31 13:48:06 -07:00
|
|
|
// unifiedInlineCall implements inline.NewInline by re-reading the function
|
2022-05-24 10:24:05 -07:00
|
|
|
// body from its Unified IR export data.
|
2023-08-17 22:13:22 -07:00
|
|
|
func unifiedInlineCall(callerfn *ir.Func, call *ir.CallExpr, fn *ir.Func, inlIndex int) *ir.InlinedCallExpr {
|
2022-07-26 21:52:42 -07:00
|
|
|
pri, ok := bodyReaderFor(fn)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if !ok {
|
2022-08-31 13:48:06 -07:00
|
|
|
base.FatalfAt(call.Pos(), "cannot inline call to %v: missing inline body", fn)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-26 16:37:24 -07:00
|
|
|
if !fn.Inl.HaveDcl {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
expandInline(fn, pri)
|
|
|
|
|
}
|
|
|
|
|
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pri.asReader(pkgbits.SectionBody, pkgbits.SyncFuncBody)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-08-16 19:45:12 -07:00
|
|
|
tmpfn := ir.NewFunc(fn.Pos(), fn.Nname.Pos(), callerfn.Sym(), fn.Type())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
r.curfn = tmpfn
|
|
|
|
|
|
2021-07-01 15:23:41 -07:00
|
|
|
r.inlCaller = callerfn
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
r.inlCall = call
|
|
|
|
|
r.inlFunc = fn
|
|
|
|
|
r.inlTreeIndex = inlIndex
|
|
|
|
|
r.inlPosBases = make(map[*src.PosBase]*src.PosBase)
|
2023-09-13 19:26:32 -07:00
|
|
|
r.funarghack = true
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
r.closureVars = make([]*ir.Name, len(r.inlFunc.ClosureVars))
|
|
|
|
|
for i, cv := range r.inlFunc.ClosureVars {
|
2023-08-11 18:21:22 -07:00
|
|
|
// TODO(mdempsky): It should be possible to support this case, but
|
|
|
|
|
// for now we rely on the inliner avoiding it.
|
|
|
|
|
if cv.Outer.Curfn != callerfn {
|
|
|
|
|
base.FatalfAt(call.Pos(), "inlining closure call across frames")
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: simplify variable capturing in unified IR
While initially building out unified IR, I didn't have any indexing
scheme. Everything was written out in order. Consequently, if I wanted
to write A before B, I had to compute A before B.
One particular example of this is handling closure variables: the
reader needs the list of closure variables before it can start reading
the function body, so I had to write them out first, and so I had to
compute them first in a separate, dedicated pass.
However, that constraint went away a while ago. For example, it's now
possible to replace the two-pass closure variable capture with a
single pass. We just write out the function body earlier, but then
wait to write out its index.
I anticipate this approach will make it easier to implement
dictionaries: rather than needing a separate pass to correctly
recognize and handle all of the generics cases, we can just hook into
the existing logic.
Change-Id: Iab1e07f9202cd5d2b6864eef10116960456214df
Reviewed-on: https://go-review.googlesource.com/c/go/+/330851
Trust: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
2021-06-25 01:54:50 -07:00
|
|
|
r.closureVars[i] = cv.Outer
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2022-07-31 18:48:16 -07:00
|
|
|
if len(r.closureVars) != 0 && r.hasTypeParams() {
|
|
|
|
|
r.dictParam = r.closureVars[len(r.closureVars)-1] // dictParam is last; see reader.funcLit
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
r.declareParams()
|
|
|
|
|
|
|
|
|
|
var inlvars, retvars []*ir.Name
|
|
|
|
|
{
|
|
|
|
|
sig := r.curfn.Type()
|
|
|
|
|
endParams := sig.NumRecvs() + sig.NumParams()
|
|
|
|
|
endResults := endParams + sig.NumResults()
|
|
|
|
|
|
|
|
|
|
inlvars = r.curfn.Dcl[:endParams]
|
|
|
|
|
retvars = r.curfn.Dcl[endParams:endResults]
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
r.delayResults = fn.Inl.CanDelayResults
|
|
|
|
|
|
|
|
|
|
r.retlabel = typecheck.AutoLabel(".i")
|
|
|
|
|
inlgen++
|
|
|
|
|
|
|
|
|
|
init := ir.TakeInit(call)
|
|
|
|
|
|
|
|
|
|
// For normal function calls, the function callee expression
|
2021-07-03 04:53:25 -07:00
|
|
|
// may contain side effects. Make sure to preserve these,
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// if necessary (#42703).
|
|
|
|
|
if call.Op() == ir.OCALLFUNC {
|
2023-10-04 15:22:49 -07:00
|
|
|
inline.CalleeEffects(&init, call.Fun)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
var args ir.Nodes
|
|
|
|
|
if call.Op() == ir.OCALLMETH {
|
2021-06-25 22:24:29 +07:00
|
|
|
base.FatalfAt(call.Pos(), "OCALLMETH missed by typecheck")
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
args.Append(call.Args...)
|
|
|
|
|
|
|
|
|
|
// Create assignment to declare and initialize inlvars.
|
2023-09-13 19:26:32 -07:00
|
|
|
as2 := ir.NewAssignListStmt(call.Pos(), ir.OAS2, ir.ToNodes(inlvars), args)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
as2.Def = true
|
|
|
|
|
var as2init ir.Nodes
|
2023-09-13 19:26:32 -07:00
|
|
|
for _, name := range inlvars {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
if ir.IsBlank(name) {
|
|
|
|
|
continue
|
|
|
|
|
}
|
|
|
|
|
// TODO(mdempsky): Use inlined position of name.Pos() instead?
|
|
|
|
|
as2init.Append(ir.NewDecl(call.Pos(), ir.ODCL, name))
|
|
|
|
|
name.Defn = as2
|
|
|
|
|
}
|
|
|
|
|
as2.SetInit(as2init)
|
|
|
|
|
init.Append(typecheck.Stmt(as2))
|
|
|
|
|
|
|
|
|
|
if !r.delayResults {
|
|
|
|
|
// If not delaying retvars, declare and zero initialize the
|
|
|
|
|
// result variables now.
|
2023-09-13 19:26:32 -07:00
|
|
|
for _, name := range retvars {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// TODO(mdempsky): Use inlined position of name.Pos() instead?
|
|
|
|
|
init.Append(ir.NewDecl(call.Pos(), ir.ODCL, name))
|
|
|
|
|
ras := ir.NewAssignStmt(call.Pos(), name, nil)
|
|
|
|
|
init.Append(typecheck.Stmt(ras))
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Add an inline mark just before the inlined body.
|
|
|
|
|
// This mark is inline in the code so that it's a reasonable spot
|
|
|
|
|
// to put a breakpoint. Not sure if that's really necessary or not
|
|
|
|
|
// (in which case it could go at the end of the function instead).
|
|
|
|
|
// Note issue 28603.
|
|
|
|
|
init.Append(ir.NewInlineMarkStmt(call.Pos().WithIsStmt(), int64(r.inlTreeIndex)))
|
|
|
|
|
|
2021-07-01 15:23:41 -07:00
|
|
|
ir.WithFunc(r.curfn, func() {
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
if !r.syntheticBody(call.Pos()) {
|
|
|
|
|
assert(r.Bool()) // have body
|
|
|
|
|
|
2022-07-31 18:48:16 -07:00
|
|
|
r.curfn.Body = r.stmts()
|
|
|
|
|
r.curfn.Endlineno = r.pos()
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2022-03-06 20:11:21 -08:00
|
|
|
// TODO(mdempsky): This shouldn't be necessary. Inlining might
|
|
|
|
|
// read in new function/method declarations, which could
|
|
|
|
|
// potentially be recursively inlined themselves; but we shouldn't
|
|
|
|
|
// need to read in the non-inlined bodies for the declarations
|
|
|
|
|
// themselves. But currently it's an easy fix to #50552.
|
2022-10-17 16:57:07 -07:00
|
|
|
readBodies(typecheck.Target, true)
|
2022-03-06 20:11:21 -08:00
|
|
|
|
2021-07-01 15:23:41 -07:00
|
|
|
// Replace any "return" statements within the function body.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
var edit func(ir.Node) ir.Node
|
|
|
|
|
edit = func(n ir.Node) ir.Node {
|
|
|
|
|
if ret, ok := n.(*ir.ReturnStmt); ok {
|
2023-09-13 19:26:32 -07:00
|
|
|
n = typecheck.Stmt(r.inlReturn(ret, retvars))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
ir.EditChildren(n, edit)
|
|
|
|
|
return n
|
|
|
|
|
}
|
|
|
|
|
edit(r.curfn)
|
2021-07-01 15:23:41 -07:00
|
|
|
})
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
body := ir.Nodes(r.curfn.Body)
|
|
|
|
|
|
2023-08-09 04:05:35 -07:00
|
|
|
// Reparent any declarations into the caller function.
|
|
|
|
|
for _, name := range r.curfn.Dcl {
|
|
|
|
|
name.Curfn = callerfn
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
if name.Class != ir.PAUTO {
|
|
|
|
|
name.SetPos(r.inlPos(name.Pos()))
|
|
|
|
|
name.SetInlFormal(true)
|
|
|
|
|
name.Class = ir.PAUTO
|
|
|
|
|
} else {
|
2023-08-09 04:05:35 -07:00
|
|
|
name.SetInlLocal(true)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
}
|
2023-09-13 19:26:32 -07:00
|
|
|
callerfn.Dcl = append(callerfn.Dcl, r.curfn.Dcl...)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
body.Append(ir.NewLabelStmt(call.Pos(), r.retlabel))
|
|
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
res := ir.NewInlinedCallExpr(call.Pos(), body, ir.ToNodes(retvars))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
res.SetInit(init)
|
|
|
|
|
res.SetType(call.Type())
|
|
|
|
|
res.SetTypecheck(1)
|
|
|
|
|
|
|
|
|
|
// Inlining shouldn't add any functions to todoBodies.
|
|
|
|
|
assert(len(todoBodies) == 0)
|
|
|
|
|
|
|
|
|
|
return res
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// inlReturn returns a statement that can substitute for the given
|
|
|
|
|
// return statement when inlining.
|
2023-09-13 19:26:32 -07:00
|
|
|
func (r *reader) inlReturn(ret *ir.ReturnStmt, retvars []*ir.Name) *ir.BlockStmt {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pos := r.inlCall.Pos()
|
|
|
|
|
|
|
|
|
|
block := ir.TakeInit(ret)
|
|
|
|
|
|
|
|
|
|
if results := ret.Results; len(results) != 0 {
|
2023-09-13 19:26:32 -07:00
|
|
|
assert(len(retvars) == len(results))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
as2 := ir.NewAssignListStmt(pos, ir.OAS2, ir.ToNodes(retvars), ret.Results)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
if r.delayResults {
|
2023-09-13 19:26:32 -07:00
|
|
|
for _, name := range retvars {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// TODO(mdempsky): Use inlined position of name.Pos() instead?
|
|
|
|
|
block.Append(ir.NewDecl(pos, ir.ODCL, name))
|
|
|
|
|
name.Defn = as2
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
block.Append(as2)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
block.Append(ir.NewBranchStmt(pos, ir.OGOTO, r.retlabel))
|
|
|
|
|
return ir.NewBlockStmt(pos, block)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// expandInline reads in an extra copy of IR to populate
|
2023-08-26 16:37:24 -07:00
|
|
|
// fn.Inl.Dcl.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
func expandInline(fn *ir.Func, pri pkgReaderIndex) {
|
2021-06-23 14:39:44 +07:00
|
|
|
// TODO(mdempsky): Remove this function. It's currently needed by
|
|
|
|
|
// dwarfgen/dwarf.go:preInliningDcls, which requires fn.Inl.Dcl to
|
|
|
|
|
// create abstract function DIEs. But we should be able to provide it
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// with the same information some other way.
|
|
|
|
|
|
|
|
|
|
fndcls := len(fn.Dcl)
|
2023-08-09 08:39:47 -07:00
|
|
|
topdcls := len(typecheck.Target.Funcs)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-08-16 19:45:12 -07:00
|
|
|
tmpfn := ir.NewFunc(fn.Pos(), fn.Nname.Pos(), fn.Sym(), fn.Type())
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
tmpfn.ClosureVars = fn.ClosureVars
|
|
|
|
|
|
|
|
|
|
{
|
2025-05-12 12:21:03 -04:00
|
|
|
r := pri.asReader(pkgbits.SectionBody, pkgbits.SyncFuncBody)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
// Don't change parameter's Sym/Nname fields.
|
|
|
|
|
r.funarghack = true
|
|
|
|
|
|
|
|
|
|
r.funcBody(tmpfn)
|
2021-07-01 15:23:41 -07:00
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2023-09-13 19:26:32 -07:00
|
|
|
// Move tmpfn's params to fn.Inl.Dcl, and reparent under fn.
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
for _, name := range tmpfn.Dcl {
|
2023-09-13 19:26:32 -07:00
|
|
|
name.Curfn = fn
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
2023-09-13 19:26:32 -07:00
|
|
|
fn.Inl.Dcl = tmpfn.Dcl
|
2023-08-26 16:37:24 -07:00
|
|
|
fn.Inl.HaveDcl = true
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
|
|
|
|
// Double check that we didn't change fn.Dcl by accident.
|
|
|
|
|
assert(fndcls == len(fn.Dcl))
|
|
|
|
|
|
|
|
|
|
// typecheck.Stmts may have added function literals to
|
|
|
|
|
// typecheck.Target.Decls. Remove them again so we don't risk trying
|
|
|
|
|
// to compile them multiple times.
|
2023-08-09 08:39:47 -07:00
|
|
|
typecheck.Target.Funcs = typecheck.Target.Funcs[:topdcls]
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// usedLocals returns a set of local variables that are used within body.
|
|
|
|
|
func usedLocals(body []ir.Node) ir.NameSet {
|
|
|
|
|
var used ir.NameSet
|
|
|
|
|
ir.VisitList(body, func(n ir.Node) {
|
|
|
|
|
if n, ok := n.(*ir.Name); ok && n.Op() == ir.ONAME && n.Class == ir.PAUTO {
|
|
|
|
|
used.Add(n)
|
|
|
|
|
}
|
|
|
|
|
})
|
|
|
|
|
return used
|
|
|
|
|
}
|
2021-06-23 21:33:24 -07:00
|
|
|
|
|
|
|
|
// @@@ Method wrappers
|
2024-01-24 00:36:26 -08:00
|
|
|
//
|
|
|
|
|
// Here we handle constructing "method wrappers," alternative entry
|
|
|
|
|
// points that adapt methods to different calling conventions. Given a
|
|
|
|
|
// user-declared method "func (T) M(i int) bool { ... }", there are a
|
|
|
|
|
// few wrappers we may need to construct:
|
|
|
|
|
//
|
|
|
|
|
// - Implicit dereferencing. Methods declared with a value receiver T
|
|
|
|
|
// are also included in the method set of the pointer type *T, so
|
|
|
|
|
// we need to construct a wrapper like "func (recv *T) M(i int)
|
|
|
|
|
// bool { return (*recv).M(i) }".
|
|
|
|
|
//
|
|
|
|
|
// - Promoted methods. If struct type U contains an embedded field of
|
|
|
|
|
// type T or *T, we need to construct a wrapper like "func (recv U)
|
|
|
|
|
// M(i int) bool { return recv.T.M(i) }".
|
|
|
|
|
//
|
|
|
|
|
// - Method values. If x is an expression of type T, then "x.M" is
|
|
|
|
|
// roughly "tmp := x; func(i int) bool { return tmp.M(i) }".
|
|
|
|
|
//
|
|
|
|
|
// At call sites, we always prefer to call the user-declared method
|
|
|
|
|
// directly, if known, so wrappers are only needed for indirect calls
|
|
|
|
|
// (for example, interface method calls that can't be devirtualized).
|
|
|
|
|
// Consequently, we can save some compile time by skipping
|
|
|
|
|
// construction of wrappers that are never needed.
|
|
|
|
|
//
|
|
|
|
|
// Alternatively, because the linker doesn't care which compilation
|
|
|
|
|
// unit constructed a particular wrapper, we can instead construct
|
|
|
|
|
// them as needed. However, if a wrapper is needed in multiple
|
|
|
|
|
// downstream packages, we may end up needing to compile it multiple
|
|
|
|
|
// times, costing us more compile time and object file size. (We mark
|
|
|
|
|
// the wrappers as DUPOK, so the linker doesn't complain about the
|
|
|
|
|
// duplicate symbols.)
|
|
|
|
|
//
|
|
|
|
|
// The current heuristics we use to balance these trade offs are:
|
|
|
|
|
//
|
|
|
|
|
// - For a (non-parameterized) defined type T, we construct wrappers
|
|
|
|
|
// for *T and any promoted methods on T (and *T) in the same
|
|
|
|
|
// compilation unit as the type declaration.
|
|
|
|
|
//
|
|
|
|
|
// - For a parameterized defined type, we construct wrappers in the
|
|
|
|
|
// compilation units in which the type is instantiated. We
|
|
|
|
|
// similarly handle wrappers for anonymous types with methods and
|
|
|
|
|
// compilation units where their type literals appear in source.
|
|
|
|
|
//
|
|
|
|
|
// - Method value expressions are relatively uncommon, so we
|
|
|
|
|
// construct their wrappers in the compilation units that they
|
|
|
|
|
// appear in.
|
|
|
|
|
//
|
|
|
|
|
// Finally, as an opportunistic compile-time optimization, if we know
|
|
|
|
|
// a wrapper was constructed in any imported package's compilation
|
|
|
|
|
// unit, then we skip constructing a duplicate one. However, currently
|
|
|
|
|
// this is only done on a best-effort basis.
|
2021-06-23 21:33:24 -07:00
|
|
|
|
|
|
|
|
// needWrapperTypes lists types for which we may need to generate
|
|
|
|
|
// method wrappers.
|
|
|
|
|
var needWrapperTypes []*types.Type
|
|
|
|
|
|
2021-08-02 23:56:13 -07:00
|
|
|
// haveWrapperTypes lists types for which we know we already have
|
|
|
|
|
// method wrappers, because we found the type in an imported package.
|
|
|
|
|
var haveWrapperTypes []*types.Type
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2021-08-30 01:01:20 -07:00
|
|
|
// needMethodValueWrappers lists methods for which we may need to
|
|
|
|
|
// generate method value wrappers.
|
|
|
|
|
var needMethodValueWrappers []methodValueWrapper
|
|
|
|
|
|
|
|
|
|
// haveMethodValueWrappers lists methods for which we know we already
|
|
|
|
|
// have method value wrappers, because we found it in an imported
|
|
|
|
|
// package.
|
|
|
|
|
var haveMethodValueWrappers []methodValueWrapper
|
|
|
|
|
|
|
|
|
|
type methodValueWrapper struct {
|
|
|
|
|
rcvr *types.Type
|
|
|
|
|
method *types.Field
|
|
|
|
|
}
|
|
|
|
|
|
2024-01-24 00:36:26 -08:00
|
|
|
// needWrapper records that wrapper methods may be needed at link
|
|
|
|
|
// time.
|
2021-08-30 01:01:20 -07:00
|
|
|
func (r *reader) needWrapper(typ *types.Type) {
|
2021-06-23 21:33:24 -07:00
|
|
|
if typ.IsPtr() {
|
2021-08-30 01:01:20 -07:00
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2024-05-17 12:07:15 -04:00
|
|
|
// Special case: runtime must define error even if imported packages mention it (#29304).
|
|
|
|
|
forceNeed := typ == types.ErrorType && base.Ctxt.Pkgpath == "runtime"
|
|
|
|
|
|
2021-08-30 01:01:20 -07:00
|
|
|
// If a type was found in an imported package, then we can assume
|
|
|
|
|
// that package (or one of its transitive dependencies) already
|
|
|
|
|
// generated method wrappers for it.
|
2024-05-17 12:07:15 -04:00
|
|
|
if r.importedDef() && !forceNeed {
|
2021-08-30 01:01:20 -07:00
|
|
|
haveWrapperTypes = append(haveWrapperTypes, typ)
|
|
|
|
|
} else {
|
|
|
|
|
needWrapperTypes = append(needWrapperTypes, typ)
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
2021-08-30 01:01:20 -07:00
|
|
|
}
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2022-05-24 10:24:05 -07:00
|
|
|
// importedDef reports whether r is reading from an imported and
|
|
|
|
|
// non-generic element.
|
|
|
|
|
//
|
|
|
|
|
// If a type was found in an imported package, then we can assume that
|
|
|
|
|
// package (or one of its transitive dependencies) already generated
|
|
|
|
|
// method wrappers for it.
|
|
|
|
|
//
|
|
|
|
|
// Exception: If we're instantiating an imported generic type or
|
|
|
|
|
// function, we might be instantiating it with type arguments not
|
|
|
|
|
// previously seen before.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): Distinguish when a generic function or type was
|
|
|
|
|
// instantiated in an imported package so that we can add types to
|
|
|
|
|
// haveWrapperTypes instead.
|
2021-08-30 01:01:20 -07:00
|
|
|
func (r *reader) importedDef() bool {
|
|
|
|
|
return r.p != localPkgReader && !r.hasTypeParams()
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
|
2024-01-24 00:36:26 -08:00
|
|
|
// MakeWrappers constructs all wrapper methods needed for the target
|
|
|
|
|
// compilation unit.
|
2021-09-01 11:53:36 -07:00
|
|
|
func MakeWrappers(target *ir.Package) {
|
2021-06-23 21:33:24 -07:00
|
|
|
// always generate a wrapper for error.Error (#29304)
|
2021-09-01 11:53:36 -07:00
|
|
|
needWrapperTypes = append(needWrapperTypes, types.ErrorType)
|
2021-06-23 21:33:24 -07:00
|
|
|
|
|
|
|
|
seen := make(map[string]*types.Type)
|
|
|
|
|
|
2021-08-02 23:56:13 -07:00
|
|
|
for _, typ := range haveWrapperTypes {
|
2021-09-01 11:53:36 -07:00
|
|
|
wrapType(typ, target, seen, false)
|
2021-08-02 23:56:13 -07:00
|
|
|
}
|
|
|
|
|
haveWrapperTypes = nil
|
|
|
|
|
|
|
|
|
|
for _, typ := range needWrapperTypes {
|
2021-09-01 11:53:36 -07:00
|
|
|
wrapType(typ, target, seen, true)
|
2021-08-02 23:56:13 -07:00
|
|
|
}
|
2021-06-23 21:33:24 -07:00
|
|
|
needWrapperTypes = nil
|
2021-08-30 01:01:20 -07:00
|
|
|
|
|
|
|
|
for _, wrapper := range haveMethodValueWrappers {
|
2021-09-01 11:53:36 -07:00
|
|
|
wrapMethodValue(wrapper.rcvr, wrapper.method, target, false)
|
2021-08-30 01:01:20 -07:00
|
|
|
}
|
|
|
|
|
haveMethodValueWrappers = nil
|
|
|
|
|
|
|
|
|
|
for _, wrapper := range needMethodValueWrappers {
|
2021-09-01 11:53:36 -07:00
|
|
|
wrapMethodValue(wrapper.rcvr, wrapper.method, target, true)
|
2021-08-30 01:01:20 -07:00
|
|
|
}
|
|
|
|
|
needMethodValueWrappers = nil
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func wrapType(typ *types.Type, target *ir.Package, seen map[string]*types.Type, needed bool) {
|
2021-08-30 01:01:20 -07:00
|
|
|
key := typ.LinkString()
|
|
|
|
|
if prev := seen[key]; prev != nil {
|
|
|
|
|
if !types.Identical(typ, prev) {
|
|
|
|
|
base.Fatalf("collision: types %v and %v have link string %q", typ, prev, key)
|
|
|
|
|
}
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
seen[key] = typ
|
|
|
|
|
|
|
|
|
|
if !needed {
|
|
|
|
|
// Only called to add to 'seen'.
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-23 21:33:24 -07:00
|
|
|
if !typ.IsInterface() {
|
|
|
|
|
typecheck.CalcMethods(typ)
|
|
|
|
|
}
|
2023-08-20 10:05:29 -07:00
|
|
|
for _, meth := range typ.AllMethods() {
|
2021-06-23 21:33:24 -07:00
|
|
|
if meth.Sym.IsBlank() || !meth.IsMethod() {
|
|
|
|
|
base.FatalfAt(meth.Pos, "invalid method: %v", meth)
|
|
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
methodWrapper(0, typ, meth, target)
|
2021-06-23 21:33:24 -07:00
|
|
|
|
|
|
|
|
// For non-interface types, we also want *T wrappers.
|
|
|
|
|
if !typ.IsInterface() {
|
2021-09-01 11:53:36 -07:00
|
|
|
methodWrapper(1, typ, meth, target)
|
2021-06-23 21:33:24 -07:00
|
|
|
|
|
|
|
|
// For not-in-heap types, *T is a scalar, not pointer shaped,
|
|
|
|
|
// so the interface wrappers use **T.
|
|
|
|
|
if typ.NotInHeap() {
|
2021-09-01 11:53:36 -07:00
|
|
|
methodWrapper(2, typ, meth, target)
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func methodWrapper(derefs int, tbase *types.Type, method *types.Field, target *ir.Package) {
|
2021-06-23 21:33:24 -07:00
|
|
|
wrapper := tbase
|
|
|
|
|
for i := 0; i < derefs; i++ {
|
|
|
|
|
wrapper = types.NewPtr(wrapper)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
sym := ir.MethodSym(wrapper, method.Sym)
|
2021-08-30 01:01:20 -07:00
|
|
|
base.Assertf(!sym.Siggen(), "already generated wrapper %v", sym)
|
2021-06-23 21:33:24 -07:00
|
|
|
sym.SetSiggen(true)
|
|
|
|
|
|
|
|
|
|
wrappee := method.Type.Recv().Type
|
|
|
|
|
if types.Identical(wrapper, wrappee) ||
|
|
|
|
|
!types.IsMethodApplicable(wrapper, method) ||
|
|
|
|
|
!reflectdata.NeedEmit(tbase) {
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Use method.Pos instead?
|
|
|
|
|
pos := base.AutogeneratedPos
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
fn := newWrapperFunc(pos, sym, wrapper, method)
|
2021-06-26 12:44:53 -07:00
|
|
|
|
|
|
|
|
var recv ir.Node = fn.Nname.Type().Recv().Nname.(*ir.Name)
|
|
|
|
|
|
|
|
|
|
// For simple *T wrappers around T methods, panicwrap produces a
|
|
|
|
|
// nicer panic message.
|
|
|
|
|
if wrapper.IsPtr() && types.Identical(wrapper.Elem(), wrappee) {
|
|
|
|
|
cond := ir.NewBinaryExpr(pos, ir.OEQ, recv, types.BuiltinPkg.Lookup("nil").Def.(ir.Node))
|
|
|
|
|
then := []ir.Node{ir.NewCallExpr(pos, ir.OCALL, typecheck.LookupRuntime("panicwrap"), nil)}
|
|
|
|
|
fn.Body.Append(ir.NewIfStmt(pos, cond, then, nil))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// typecheck will add one implicit deref, if necessary,
|
|
|
|
|
// but not-in-heap types require more for their **T wrappers.
|
|
|
|
|
for i := 1; i < derefs; i++ {
|
|
|
|
|
recv = Implicit(ir.NewStarExpr(pos, recv))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
addTailCall(pos, fn, recv, method)
|
2021-06-30 19:20:28 -07:00
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
finishWrapperFunc(fn, target)
|
2021-06-26 12:44:53 -07:00
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func wrapMethodValue(recvType *types.Type, method *types.Field, target *ir.Package, needed bool) {
|
2021-06-26 12:44:53 -07:00
|
|
|
sym := ir.MethodSymSuffix(recvType, method.Sym, "-fm")
|
2021-08-30 01:01:20 -07:00
|
|
|
if sym.Uniq() {
|
|
|
|
|
return
|
|
|
|
|
}
|
2021-06-26 12:44:53 -07:00
|
|
|
sym.SetUniq(true)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Use method.Pos instead?
|
|
|
|
|
pos := base.AutogeneratedPos
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
fn := newWrapperFunc(pos, sym, nil, method)
|
2021-08-03 10:37:33 -07:00
|
|
|
sym.Def = fn.Nname
|
2021-06-26 12:44:53 -07:00
|
|
|
|
|
|
|
|
// Declare and initialize variable holding receiver.
|
2021-07-02 16:29:42 -07:00
|
|
|
recv := ir.NewHiddenParam(pos, fn, typecheck.Lookup(".this"), recvType)
|
2021-06-26 12:44:53 -07:00
|
|
|
|
2021-08-30 01:01:20 -07:00
|
|
|
if !needed {
|
2021-08-02 23:56:13 -07:00
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-26 12:44:53 -07:00
|
|
|
addTailCall(pos, fn, recv, method)
|
2021-06-30 19:20:28 -07:00
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
finishWrapperFunc(fn, target)
|
2021-06-26 12:44:53 -07:00
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func newWrapperFunc(pos src.XPos, sym *types.Sym, wrapper *types.Type, method *types.Field) *ir.Func {
|
2023-08-16 17:23:52 -07:00
|
|
|
sig := newWrapperType(wrapper, method)
|
2023-08-16 19:45:12 -07:00
|
|
|
fn := ir.NewFunc(pos, pos, sym, sig)
|
2023-09-13 19:26:32 -07:00
|
|
|
fn.DeclareParams(true)
|
2021-06-26 12:44:53 -07:00
|
|
|
fn.SetDupok(true) // TODO(mdempsky): Leave unset for local, non-generic wrappers?
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2021-06-26 12:44:53 -07:00
|
|
|
return fn
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
func finishWrapperFunc(fn *ir.Func, target *ir.Package) {
|
2021-06-30 19:20:28 -07:00
|
|
|
ir.WithFunc(fn, func() {
|
|
|
|
|
typecheck.Stmts(fn.Body)
|
|
|
|
|
})
|
|
|
|
|
|
2021-09-01 11:53:36 -07:00
|
|
|
// We generate wrappers after the global inlining pass,
|
|
|
|
|
// so we're responsible for applying inlining ourselves here.
|
2022-10-28 17:34:43 -04:00
|
|
|
// TODO(prattmic): plumb PGO.
|
2023-11-18 20:16:47 -08:00
|
|
|
interleaved.DevirtualizeAndInlineFunc(fn, nil)
|
2021-09-01 11:53:36 -07:00
|
|
|
|
2022-05-24 00:41:13 +07:00
|
|
|
// The body of wrapper function after inlining may reveal new ir.OMETHVALUE node,
|
|
|
|
|
// we don't know whether wrapper function has been generated for it or not, so
|
|
|
|
|
// generate one immediately here.
|
2023-06-23 01:04:29 +07:00
|
|
|
//
|
|
|
|
|
// Further, after CL 492017, function that construct closures is allowed to be inlined,
|
|
|
|
|
// even though the closure itself can't be inline. So we also need to visit body of any
|
|
|
|
|
// closure that we see when visiting body of the wrapper function.
|
|
|
|
|
ir.VisitFuncAndClosures(fn, func(n ir.Node) {
|
2022-05-24 00:41:13 +07:00
|
|
|
if n, ok := n.(*ir.SelectorExpr); ok && n.Op() == ir.OMETHVALUE {
|
|
|
|
|
wrapMethodValue(n.X.Type(), n.Selection, target, true)
|
|
|
|
|
}
|
|
|
|
|
})
|
|
|
|
|
|
2023-08-16 19:45:12 -07:00
|
|
|
fn.Nname.Defn = fn
|
2023-08-09 08:39:47 -07:00
|
|
|
target.Funcs = append(target.Funcs, fn)
|
2021-06-30 19:20:28 -07:00
|
|
|
}
|
|
|
|
|
|
2021-06-23 21:33:24 -07:00
|
|
|
// newWrapperType returns a copy of the given signature type, but with
|
2021-06-26 12:44:53 -07:00
|
|
|
// the receiver parameter type substituted with recvType.
|
|
|
|
|
// If recvType is nil, newWrapperType returns a signature
|
|
|
|
|
// without a receiver parameter.
|
|
|
|
|
func newWrapperType(recvType *types.Type, method *types.Field) *types.Type {
|
2021-06-23 21:33:24 -07:00
|
|
|
clone := func(params []*types.Field) []*types.Field {
|
|
|
|
|
res := make([]*types.Field, len(params))
|
|
|
|
|
for i, param := range params {
|
2023-09-13 19:26:32 -07:00
|
|
|
res[i] = types.NewField(param.Pos, param.Sym, param.Type)
|
2021-06-23 21:33:24 -07:00
|
|
|
res[i].SetIsDDD(param.IsDDD())
|
|
|
|
|
}
|
|
|
|
|
return res
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-26 12:44:53 -07:00
|
|
|
sig := method.Type
|
|
|
|
|
|
|
|
|
|
var recv *types.Field
|
|
|
|
|
if recvType != nil {
|
2023-09-13 19:26:32 -07:00
|
|
|
recv = types.NewField(sig.Recv().Pos, sig.Recv().Sym, recvType)
|
2021-06-26 12:44:53 -07:00
|
|
|
}
|
2023-08-20 10:05:29 -07:00
|
|
|
params := clone(sig.Params())
|
|
|
|
|
results := clone(sig.Results())
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
return types.NewSignature(recv, params, results)
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
|
2021-06-26 12:44:53 -07:00
|
|
|
func addTailCall(pos src.XPos, fn *ir.Func, recv ir.Node, method *types.Field) {
|
|
|
|
|
sig := fn.Nname.Type()
|
|
|
|
|
args := make([]ir.Node, sig.NumParams())
|
2023-08-20 10:05:29 -07:00
|
|
|
for i, param := range sig.Params() {
|
2021-06-26 12:44:53 -07:00
|
|
|
args[i] = param.Nname.(*ir.Name)
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-18 19:40:33 -07:00
|
|
|
dot := typecheck.XDotMethod(pos, recv, method.Sym, true)
|
2021-07-03 11:55:31 -07:00
|
|
|
call := typecheck.Call(pos, dot, args, method.Type.IsVariadic()).(*ir.CallExpr)
|
2021-06-23 21:33:24 -07:00
|
|
|
|
2024-04-11 10:57:32 +03:00
|
|
|
if recv.Type() != nil && recv.Type().IsPtr() && method.Type.Recv().Type.IsPtr() &&
|
|
|
|
|
method.Embedded != 0 && !types.IsInterfaceMethod(method.Type) &&
|
2025-06-03 15:44:32 -04:00
|
|
|
!unifiedHaveInlineBody(ir.MethodExprName(dot).Func) &&
|
2024-04-11 10:57:32 +03:00
|
|
|
!(base.Ctxt.Arch.Name == "ppc64le" && base.Ctxt.Flag_dynlink) {
|
|
|
|
|
if base.Debug.TailCall != 0 {
|
|
|
|
|
base.WarnfAt(fn.Nname.Type().Recv().Type.Elem().Pos(), "tail call emitted for the method %v wrapper", method.Nname)
|
|
|
|
|
}
|
2025-06-03 15:44:32 -04:00
|
|
|
// Prefer OTAILCALL to reduce code size (except the case when the called method can be inlined).
|
2024-04-11 10:57:32 +03:00
|
|
|
fn.Body.Append(ir.NewTailCallStmt(pos, call))
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fn.SetWrapper(true)
|
|
|
|
|
|
2021-06-23 21:33:24 -07:00
|
|
|
if method.Type.NumResults() == 0 {
|
2021-06-26 12:44:53 -07:00
|
|
|
fn.Body.Append(call)
|
|
|
|
|
return
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ret := ir.NewReturnStmt(pos, nil)
|
|
|
|
|
ret.Results = []ir.Node{call}
|
2021-06-26 12:44:53 -07:00
|
|
|
fn.Body.Append(ret)
|
2021-06-23 21:33:24 -07:00
|
|
|
}
|
2022-05-25 23:27:45 +07:00
|
|
|
|
|
|
|
|
func setBasePos(pos src.XPos) {
|
|
|
|
|
// Set the position for any error messages we might print (e.g. too large types).
|
|
|
|
|
base.Pos = pos
|
|
|
|
|
}
|
2022-07-31 18:48:16 -07:00
|
|
|
|
|
|
|
|
// dictParamName is the name of the synthetic dictionary parameter
|
|
|
|
|
// added to shaped functions.
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
//
|
|
|
|
|
// N.B., this variable name is known to Delve:
|
|
|
|
|
// https://github.com/go-delve/delve/blob/cb91509630529e6055be845688fd21eb89ae8714/pkg/proc/eval.go#L28
|
2023-05-02 17:37:00 +00:00
|
|
|
const dictParamName = typecheck.LocalDictName
|
2022-07-31 18:48:16 -07:00
|
|
|
|
|
|
|
|
// shapeSig returns a copy of fn's signature, except adding a
|
|
|
|
|
// dictionary parameter and promoting the receiver parameter (if any)
|
|
|
|
|
// to a normal parameter.
|
|
|
|
|
//
|
|
|
|
|
// The parameter types.Fields are all copied too, so their Nname
|
|
|
|
|
// fields can be initialized for use by the shape function.
|
|
|
|
|
func shapeSig(fn *ir.Func, dict *readerDict) *types.Type {
|
|
|
|
|
sig := fn.Nname.Type()
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
oldRecv := sig.Recv()
|
|
|
|
|
|
|
|
|
|
var recv *types.Field
|
|
|
|
|
if oldRecv != nil {
|
|
|
|
|
recv = types.NewField(oldRecv.Pos, oldRecv.Sym, oldRecv.Type)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-20 10:05:29 -07:00
|
|
|
params := make([]*types.Field, 1+sig.NumParams())
|
2022-07-31 18:48:16 -07:00
|
|
|
params[0] = types.NewField(fn.Pos(), fn.Sym().Pkg.Lookup(dictParamName), types.NewPtr(dict.varType()))
|
2023-08-20 10:05:29 -07:00
|
|
|
for i, param := range sig.Params() {
|
2022-07-31 18:48:16 -07:00
|
|
|
d := types.NewField(param.Pos, param.Sym, param.Type)
|
|
|
|
|
d.SetIsDDD(param.IsDDD())
|
cmd/compile/internal/noder: shape-based stenciling for unified IR
This CL switches unified IR to use shape-based stenciling with runtime
dictionaries, like the existing non-unified frontend. Specifically,
when instantiating generic functions and types `X[T]`, we now also
instantiated shaped variants `X[shapify(T)]` that can be shared by
`T`'s with common underlying types.
For example, for generic function `F`, `F[int](args...)` will be
rewritten to `F[go.shape.int](&.dict.F[int], args...)`.
For generic type `T` with method `M` and value `t` of type `T[int]`,
`t.M(args...)` will be rewritten to `T[go.shape.int].M(t,
&.dict.T[int], args...)`.
Two notable distinctions from the non-unified frontend:
1. For simplicity, currently shaping is limited to simply converting
type arguments to their underlying type. Subsequent CLs will implement
more aggressive shaping.
2. For generic types, a single dictionary is generated to be shared by
all methods, rather than separate dictionaries for each method. I
originally went with this design because I have an idea of changing
interface calls to pass the itab pointer via the closure
register (which should have zero overhead), and then the interface
wrappers for generic methods could use the *runtime.itab to find the
runtime dictionary that corresponds to the dynamic type. This would
allow emitting fewer method wrappers.
However, this choice does have the consequence that currently even if
a method is unused and its code is pruned by the linker, it may have
produced runtime dictionary entries that need to be kept alive anyway.
I'm open to changing this to generate per-method dictionaries, though
this would require changing the unified IR export data format; so it
would be best to make this decision before Go 1.20.
The other option is making the linker smarter about pruning unneeded
dictionary entries, like how it already prunes itab entries. For
example, the runtime dictionary for `T[int]` could have a `R_DICTTYPE`
meta-relocation against symbol `.dicttype.T[go.shape.int]` that
declares it's a dictionary associated with that type; and then each
method on `T[go.shape.T]` could have `R_DICTUSE` meta-relocations
against `.dicttype.T[go.shape.T]+offset` indicating which fields
within dictionaries of that type need to be preserved.
Change-Id: I369580b1d93d19640a4b5ecada4f6231adcce3fd
Reviewed-on: https://go-review.googlesource.com/c/go/+/421821
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
2022-08-06 16:40:56 -07:00
|
|
|
params[1+i] = d
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|
|
|
|
|
|
2023-08-20 10:05:29 -07:00
|
|
|
results := make([]*types.Field, sig.NumResults())
|
|
|
|
|
for i, result := range sig.Results() {
|
2022-07-31 18:48:16 -07:00
|
|
|
results[i] = types.NewField(result.Pos, result.Sym, result.Type)
|
|
|
|
|
}
|
|
|
|
|
|
2022-12-01 17:42:02 -08:00
|
|
|
return types.NewSignature(recv, params, results)
|
2022-07-31 18:48:16 -07:00
|
|
|
}
|