[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// UNREVIEWED
|
|
|
|
|
|
|
|
|
|
// 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 (
|
|
|
|
|
"fmt"
|
|
|
|
|
"go/constant"
|
|
|
|
|
|
|
|
|
|
"cmd/compile/internal/base"
|
|
|
|
|
"cmd/compile/internal/ir"
|
|
|
|
|
"cmd/compile/internal/syntax"
|
|
|
|
|
"cmd/compile/internal/types2"
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
type pkgWriter struct {
|
|
|
|
|
pkgEncoder
|
|
|
|
|
|
|
|
|
|
m posMap
|
|
|
|
|
curpkg *types2.Package
|
|
|
|
|
info *types2.Info
|
|
|
|
|
|
|
|
|
|
posBasesIdx map[*syntax.PosBase]int
|
|
|
|
|
pkgsIdx map[*types2.Package]int
|
|
|
|
|
typsIdx map[types2.Type]int
|
|
|
|
|
globalsIdx map[types2.Object]int
|
|
|
|
|
|
|
|
|
|
funDecls map[*types2.Func]*syntax.FuncDecl
|
|
|
|
|
typDecls map[*types2.TypeName]typeDeclGen
|
|
|
|
|
|
|
|
|
|
linknames map[types2.Object]string
|
|
|
|
|
cgoPragmas [][]string
|
|
|
|
|
|
|
|
|
|
dups dupTypes
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func newPkgWriter(m posMap, pkg *types2.Package, info *types2.Info) *pkgWriter {
|
|
|
|
|
return &pkgWriter{
|
|
|
|
|
pkgEncoder: newPkgEncoder(),
|
|
|
|
|
|
|
|
|
|
m: m,
|
|
|
|
|
curpkg: pkg,
|
|
|
|
|
info: info,
|
|
|
|
|
|
|
|
|
|
pkgsIdx: make(map[*types2.Package]int),
|
|
|
|
|
globalsIdx: make(map[types2.Object]int),
|
|
|
|
|
typsIdx: make(map[types2.Type]int),
|
|
|
|
|
|
|
|
|
|
posBasesIdx: make(map[*syntax.PosBase]int),
|
|
|
|
|
|
|
|
|
|
funDecls: make(map[*types2.Func]*syntax.FuncDecl),
|
|
|
|
|
typDecls: make(map[*types2.TypeName]typeDeclGen),
|
|
|
|
|
|
|
|
|
|
linknames: make(map[types2.Object]string),
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) errorf(p poser, msg string, args ...interface{}) {
|
|
|
|
|
base.ErrorfAt(pw.m.pos(p), msg, args...)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) fatalf(p poser, msg string, args ...interface{}) {
|
|
|
|
|
base.FatalfAt(pw.m.pos(p), msg, args...)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) unexpected(what string, p poser) {
|
|
|
|
|
pw.fatalf(p, "unexpected %s: %v (%T)", what, p, p)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type writer struct {
|
|
|
|
|
p *pkgWriter
|
|
|
|
|
|
|
|
|
|
encoder
|
|
|
|
|
|
|
|
|
|
// For writing out object descriptions, ext points to the extension
|
|
|
|
|
// writer for where we can write the compiler's private extension
|
|
|
|
|
// details for the object.
|
|
|
|
|
//
|
|
|
|
|
// TODO(mdempsky): This is a little hacky, but works easiest with
|
|
|
|
|
// the way things are currently.
|
|
|
|
|
ext *writer
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): We should be able to prune localsIdx whenever a
|
|
|
|
|
// scope closes, and then maybe we can just use the same map for
|
|
|
|
|
// storing the TypeParams too (as their TypeName instead).
|
|
|
|
|
|
|
|
|
|
// variables declared within this function
|
[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
|
|
|
localsIdx map[*types2.Var]int
|
|
|
|
|
|
|
|
|
|
closureVars []posObj
|
|
|
|
|
closureVarsIdx map[*types2.Var]int
|
2021-06-28 22:41:50 -07:00
|
|
|
|
|
|
|
|
dict *writerDict
|
|
|
|
|
derived bool
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// A writerDict tracks types and objects that are used by a declaration.
|
|
|
|
|
type writerDict struct {
|
|
|
|
|
implicits []*types2.TypeName
|
|
|
|
|
|
|
|
|
|
// derived is a slice of type indices for computing derived types
|
|
|
|
|
// (i.e., types that depend on the declaration's type parameters).
|
2021-07-13 09:09:32 -07:00
|
|
|
derived []derivedInfo
|
2021-06-28 22:41:50 -07:00
|
|
|
|
|
|
|
|
// derivedIdx maps a Type to its corresponding index within the
|
|
|
|
|
// derived slice, if present.
|
|
|
|
|
derivedIdx map[types2.Type]int
|
2021-07-13 09:09:32 -07:00
|
|
|
|
|
|
|
|
// funcs lists references to generic functions that were
|
|
|
|
|
// instantiated with derived types (i.e., that require
|
|
|
|
|
// sub-dictionaries when called at run time).
|
|
|
|
|
funcs []objInfo
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type derivedInfo struct {
|
|
|
|
|
idx int
|
|
|
|
|
needed bool
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type typeInfo struct {
|
|
|
|
|
idx int
|
|
|
|
|
derived bool
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type objInfo struct {
|
|
|
|
|
idx int // index for the generic function declaration
|
|
|
|
|
explicits []typeInfo // info for the type arguments
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (info objInfo) anyDerived() bool {
|
|
|
|
|
for _, explicit := range info.explicits {
|
|
|
|
|
if explicit.derived {
|
|
|
|
|
return true
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (info objInfo) equals(other objInfo) bool {
|
|
|
|
|
if info.idx != other.idx {
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
assert(len(info.explicits) == len(other.explicits))
|
|
|
|
|
for i, targ := range info.explicits {
|
|
|
|
|
if targ != other.explicits[i] {
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return 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 (pw *pkgWriter) newWriter(k reloc, marker syncMarker) *writer {
|
|
|
|
|
return &writer{
|
|
|
|
|
encoder: pw.newEncoder(k, marker),
|
|
|
|
|
p: pw,
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Positions
|
|
|
|
|
|
|
|
|
|
func (w *writer) pos(p poser) {
|
|
|
|
|
w.sync(syncPos)
|
|
|
|
|
pos := p.Pos()
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Track down the remaining cases here and fix them.
|
|
|
|
|
if !w.bool(pos.IsKnown()) {
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Delta encoding. Also, if there's a b-side, update
|
|
|
|
|
// its position base too (but not vice versa!).
|
|
|
|
|
w.posBase(pos.Base())
|
|
|
|
|
w.uint(pos.Line())
|
|
|
|
|
w.uint(pos.Col())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) posBase(b *syntax.PosBase) {
|
|
|
|
|
w.reloc(relocPosBase, w.p.posBaseIdx(b))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) posBaseIdx(b *syntax.PosBase) int {
|
|
|
|
|
if idx, ok := pw.posBasesIdx[b]; ok {
|
|
|
|
|
return idx
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w := pw.newWriter(relocPosBase, syncPosBase)
|
|
|
|
|
w.p.posBasesIdx[b] = w.idx
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): What exactly does "fileh" do anyway? Is writing
|
|
|
|
|
// out both of these strings really the right thing to do here?
|
|
|
|
|
fn := b.Filename()
|
|
|
|
|
w.string(fn)
|
|
|
|
|
w.string(fileh(fn))
|
|
|
|
|
|
|
|
|
|
if !w.bool(b.IsFileBase()) {
|
|
|
|
|
w.pos(b)
|
|
|
|
|
w.uint(b.Line())
|
|
|
|
|
w.uint(b.Col())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return w.flush()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Packages
|
|
|
|
|
|
|
|
|
|
func (w *writer) pkg(pkg *types2.Package) {
|
|
|
|
|
w.sync(syncPkg)
|
|
|
|
|
w.reloc(relocPkg, w.p.pkgIdx(pkg))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) pkgIdx(pkg *types2.Package) int {
|
|
|
|
|
if idx, ok := pw.pkgsIdx[pkg]; ok {
|
|
|
|
|
return idx
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w := pw.newWriter(relocPkg, syncPkgDef)
|
|
|
|
|
pw.pkgsIdx[pkg] = w.idx
|
|
|
|
|
|
|
|
|
|
if pkg == nil {
|
|
|
|
|
w.string("builtin")
|
|
|
|
|
} else {
|
|
|
|
|
var path string
|
|
|
|
|
if pkg != w.p.curpkg {
|
|
|
|
|
path = pkg.Path()
|
|
|
|
|
}
|
|
|
|
|
w.string(path)
|
|
|
|
|
w.string(pkg.Name())
|
|
|
|
|
w.len(pkg.Height())
|
|
|
|
|
|
|
|
|
|
w.len(len(pkg.Imports()))
|
|
|
|
|
for _, imp := range pkg.Imports() {
|
|
|
|
|
w.pkg(imp)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return w.flush()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Types
|
|
|
|
|
|
|
|
|
|
func (w *writer) typ(typ types2.Type) {
|
2021-07-13 09:09:32 -07:00
|
|
|
w.typInfo(w.p.typIdx(typ, w.dict))
|
|
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
|
2021-07-13 09:09:32 -07:00
|
|
|
func (w *writer) typInfo(info typeInfo) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.sync(syncType)
|
2021-07-13 09:09:32 -07:00
|
|
|
if w.bool(info.derived) {
|
|
|
|
|
w.len(info.idx)
|
2021-06-28 22:41:50 -07:00
|
|
|
w.derived = true
|
|
|
|
|
} else {
|
2021-07-13 09:09:32 -07:00
|
|
|
w.reloc(relocType, info.idx)
|
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
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
// typIdx returns the index where the export data description of type
|
|
|
|
|
// can be read back in. If no such index exists yet, it's created.
|
|
|
|
|
//
|
|
|
|
|
// typIdx also reports whether typ is a derived type; that is, whether
|
|
|
|
|
// its identity depends on type parameters.
|
2021-07-13 09:09:32 -07:00
|
|
|
func (pw *pkgWriter) typIdx(typ types2.Type, dict *writerDict) typeInfo {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 quirksMode() {
|
2021-06-28 22:41:50 -07:00
|
|
|
typ = pw.dups.orig(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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if idx, ok := pw.typsIdx[typ]; ok {
|
2021-07-13 09:09:32 -07:00
|
|
|
return typeInfo{idx: idx, derived: false}
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
|
|
|
|
if dict != nil {
|
|
|
|
|
if idx, ok := dict.derivedIdx[typ]; ok {
|
2021-07-13 09:09:32 -07:00
|
|
|
return typeInfo{idx: idx, derived: true}
|
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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w := pw.newWriter(relocType, syncTypeIdx)
|
2021-06-28 22:41:50 -07:00
|
|
|
w.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
|
|
|
|
|
|
|
|
switch typ := typ.(type) {
|
|
|
|
|
default:
|
|
|
|
|
base.Fatalf("unexpected type: %v (%T)", typ, typ)
|
|
|
|
|
|
|
|
|
|
case *types2.Basic:
|
|
|
|
|
if kind := typ.Kind(); types2.Typ[kind] == typ {
|
|
|
|
|
w.code(typeBasic)
|
|
|
|
|
w.len(int(kind))
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Handle "byte" and "rune" as references to their TypeName.
|
|
|
|
|
obj := types2.Universe.Lookup(typ.Name())
|
|
|
|
|
assert(obj.Type() == typ)
|
|
|
|
|
|
|
|
|
|
w.code(typeNamed)
|
|
|
|
|
w.obj(obj, nil)
|
|
|
|
|
|
|
|
|
|
case *types2.Named:
|
|
|
|
|
// Type aliases can refer to uninstantiated generic types, so we
|
|
|
|
|
// might see len(TParams) != 0 && len(TArgs) == 0 here.
|
|
|
|
|
// TODO(mdempsky): Revisit after #46477 is resolved.
|
|
|
|
|
assert(len(typ.TParams()) == len(typ.TArgs()) || len(typ.TArgs()) == 0)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Why do we need to loop here?
|
|
|
|
|
orig := typ
|
|
|
|
|
for orig.TArgs() != nil {
|
|
|
|
|
orig = orig.Orig()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(typeNamed)
|
|
|
|
|
w.obj(orig.Obj(), typ.TArgs())
|
|
|
|
|
|
|
|
|
|
case *types2.TypeParam:
|
2021-06-28 22:41:50 -07:00
|
|
|
index := func() int {
|
|
|
|
|
for idx, name := range w.dict.implicits {
|
|
|
|
|
if name.Type().(*types2.TypeParam) == typ {
|
|
|
|
|
return idx
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return len(w.dict.implicits) + typ.Index()
|
|
|
|
|
}()
|
|
|
|
|
|
|
|
|
|
w.derived = 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
|
|
|
w.code(typeTypeParam)
|
2021-06-28 22:41:50 -07:00
|
|
|
w.len(index)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 *types2.Array:
|
|
|
|
|
w.code(typeArray)
|
|
|
|
|
w.uint64(uint64(typ.Len()))
|
|
|
|
|
w.typ(typ.Elem())
|
|
|
|
|
|
|
|
|
|
case *types2.Chan:
|
|
|
|
|
w.code(typeChan)
|
|
|
|
|
w.len(int(typ.Dir()))
|
|
|
|
|
w.typ(typ.Elem())
|
|
|
|
|
|
|
|
|
|
case *types2.Map:
|
|
|
|
|
w.code(typeMap)
|
|
|
|
|
w.typ(typ.Key())
|
|
|
|
|
w.typ(typ.Elem())
|
|
|
|
|
|
|
|
|
|
case *types2.Pointer:
|
|
|
|
|
w.code(typePointer)
|
|
|
|
|
w.typ(typ.Elem())
|
|
|
|
|
|
|
|
|
|
case *types2.Signature:
|
|
|
|
|
assert(typ.TParams() == nil)
|
|
|
|
|
w.code(typeSignature)
|
|
|
|
|
w.signature(typ)
|
|
|
|
|
|
|
|
|
|
case *types2.Slice:
|
|
|
|
|
w.code(typeSlice)
|
|
|
|
|
w.typ(typ.Elem())
|
|
|
|
|
|
|
|
|
|
case *types2.Struct:
|
|
|
|
|
w.code(typeStruct)
|
|
|
|
|
w.structType(typ)
|
|
|
|
|
|
|
|
|
|
case *types2.Interface:
|
|
|
|
|
w.code(typeInterface)
|
|
|
|
|
w.interfaceType(typ)
|
|
|
|
|
|
|
|
|
|
case *types2.Union:
|
|
|
|
|
w.code(typeUnion)
|
|
|
|
|
w.unionType(typ)
|
|
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
|
|
|
|
|
if w.derived {
|
|
|
|
|
idx := len(dict.derived)
|
2021-07-13 09:09:32 -07:00
|
|
|
dict.derived = append(dict.derived, derivedInfo{idx: w.flush()})
|
2021-06-28 22:41:50 -07:00
|
|
|
dict.derivedIdx[typ] = idx
|
2021-07-13 09:09:32 -07:00
|
|
|
return typeInfo{idx: idx, derived: true}
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pw.typsIdx[typ] = w.idx
|
2021-07-13 09:09:32 -07:00
|
|
|
return typeInfo{idx: w.flush(), 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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) structType(typ *types2.Struct) {
|
|
|
|
|
w.len(typ.NumFields())
|
|
|
|
|
for i := 0; i < typ.NumFields(); i++ {
|
|
|
|
|
f := typ.Field(i)
|
|
|
|
|
w.pos(f)
|
|
|
|
|
w.selector(f)
|
|
|
|
|
w.typ(f.Type())
|
|
|
|
|
w.string(typ.Tag(i))
|
|
|
|
|
w.bool(f.Embedded())
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) unionType(typ *types2.Union) {
|
|
|
|
|
w.len(typ.NumTerms())
|
|
|
|
|
for i := 0; i < typ.NumTerms(); i++ {
|
|
|
|
|
term, tilde := typ.Term(i)
|
|
|
|
|
w.typ(term)
|
|
|
|
|
w.bool(tilde)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) interfaceType(typ *types2.Interface) {
|
|
|
|
|
w.len(typ.NumExplicitMethods())
|
|
|
|
|
w.len(typ.NumEmbeddeds())
|
|
|
|
|
|
|
|
|
|
for i := 0; i < typ.NumExplicitMethods(); i++ {
|
|
|
|
|
m := typ.ExplicitMethod(i)
|
|
|
|
|
sig := m.Type().(*types2.Signature)
|
|
|
|
|
assert(sig.TParams() == nil)
|
|
|
|
|
|
|
|
|
|
w.pos(m)
|
|
|
|
|
w.selector(m)
|
|
|
|
|
w.signature(sig)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
for i := 0; i < typ.NumEmbeddeds(); i++ {
|
|
|
|
|
w.typ(typ.EmbeddedType(i))
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) signature(sig *types2.Signature) {
|
|
|
|
|
w.sync(syncSignature)
|
|
|
|
|
w.params(sig.Params())
|
|
|
|
|
w.params(sig.Results())
|
|
|
|
|
w.bool(sig.Variadic())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) params(typ *types2.Tuple) {
|
|
|
|
|
w.sync(syncParams)
|
|
|
|
|
w.len(typ.Len())
|
|
|
|
|
for i := 0; i < typ.Len(); i++ {
|
|
|
|
|
w.param(typ.At(i))
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) param(param *types2.Var) {
|
|
|
|
|
w.sync(syncParam)
|
|
|
|
|
w.pos(param)
|
|
|
|
|
w.localIdent(param)
|
|
|
|
|
w.typ(param.Type())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Objects
|
|
|
|
|
|
|
|
|
|
func (w *writer) obj(obj types2.Object, explicits []types2.Type) {
|
2021-07-13 09:09:32 -07:00
|
|
|
explicitInfos := make([]typeInfo, len(explicits))
|
|
|
|
|
for i, explicit := range explicits {
|
|
|
|
|
explicitInfos[i] = w.p.typIdx(explicit, w.dict)
|
|
|
|
|
}
|
|
|
|
|
info := objInfo{idx: w.p.objIdx(obj), explicits: explicitInfos}
|
|
|
|
|
|
|
|
|
|
if _, ok := obj.(*types2.Func); ok && info.anyDerived() {
|
|
|
|
|
idx := -1
|
|
|
|
|
for i, prev := range w.dict.funcs {
|
|
|
|
|
if prev.equals(info) {
|
|
|
|
|
idx = i
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if idx < 0 {
|
|
|
|
|
idx = len(w.dict.funcs)
|
|
|
|
|
w.dict.funcs = append(w.dict.funcs, info)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Push up into expr; this shouldn't appear
|
|
|
|
|
// outside of expression context.
|
|
|
|
|
w.sync(syncObject)
|
|
|
|
|
w.bool(true)
|
|
|
|
|
w.len(idx)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Push up into typIdx; this shouldn't be needed
|
|
|
|
|
// except while writing out types.
|
2021-06-28 22:41:50 -07:00
|
|
|
if isDefinedType(obj) && obj.Pkg() == w.p.curpkg {
|
|
|
|
|
decl, ok := w.p.typDecls[obj.(*types2.TypeName)]
|
|
|
|
|
assert(ok)
|
|
|
|
|
if len(decl.implicits) != 0 {
|
|
|
|
|
w.derived = 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
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
|
|
|
|
|
w.sync(syncObject)
|
2021-07-13 09:09:32 -07:00
|
|
|
w.bool(false)
|
|
|
|
|
w.reloc(relocObj, info.idx)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-07-13 09:09:32 -07:00
|
|
|
w.len(len(info.explicits))
|
|
|
|
|
for _, info := range info.explicits {
|
|
|
|
|
w.typInfo(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
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
func (pw *pkgWriter) objIdx(obj types2.Object) 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
|
|
|
if idx, ok := pw.globalsIdx[obj]; ok {
|
|
|
|
|
return idx
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
dict := &writerDict{
|
|
|
|
|
derivedIdx: make(map[types2.Type]int),
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if isDefinedType(obj) && obj.Pkg() == pw.curpkg {
|
|
|
|
|
decl, ok := pw.typDecls[obj.(*types2.TypeName)]
|
|
|
|
|
assert(ok)
|
|
|
|
|
dict.implicits = decl.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
|
|
|
w := pw.newWriter(relocObj, syncObject1)
|
|
|
|
|
w.ext = pw.newWriter(relocObjExt, syncObject1)
|
2021-06-28 22:41:50 -07:00
|
|
|
wdict := pw.newWriter(relocObjDict, syncObject1)
|
|
|
|
|
|
|
|
|
|
pw.globalsIdx[obj] = w.idx // break cycles
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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(w.ext.idx == w.idx)
|
2021-06-28 22:41:50 -07:00
|
|
|
assert(wdict.idx == w.idx)
|
|
|
|
|
|
|
|
|
|
w.dict = dict
|
|
|
|
|
w.ext.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
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
// Ident goes first so importer can avoid unnecessary work if
|
|
|
|
|
// they've already resolved this object.
|
|
|
|
|
w.qualifiedIdent(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
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
w.typeParamBounds(objTypeParams(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
|
|
|
|
|
|
|
|
w.doObj(obj)
|
|
|
|
|
|
|
|
|
|
w.flush()
|
|
|
|
|
w.ext.flush()
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
// Done writing out the object description; write out the list of
|
2021-07-13 09:09:32 -07:00
|
|
|
// derived types and instantiated functions found along the way.
|
2021-06-28 22:41:50 -07:00
|
|
|
wdict.len(len(dict.derived))
|
|
|
|
|
for _, typ := range dict.derived {
|
2021-07-13 09:09:32 -07:00
|
|
|
wdict.reloc(relocType, typ.idx)
|
|
|
|
|
wdict.bool(typ.needed)
|
|
|
|
|
}
|
|
|
|
|
wdict.len(len(dict.funcs))
|
|
|
|
|
for _, fn := range dict.funcs {
|
|
|
|
|
wdict.reloc(relocObj, fn.idx)
|
|
|
|
|
wdict.len(len(fn.explicits))
|
|
|
|
|
for _, targ := range fn.explicits {
|
|
|
|
|
wdict.typInfo(targ)
|
|
|
|
|
}
|
2021-06-28 22:41:50 -07:00
|
|
|
}
|
|
|
|
|
wdict.flush()
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 w.idx
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) doObj(obj types2.Object) {
|
|
|
|
|
if obj.Pkg() != w.p.curpkg {
|
|
|
|
|
w.code(objStub)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch obj := obj.(type) {
|
|
|
|
|
default:
|
|
|
|
|
w.p.unexpected("object", obj)
|
|
|
|
|
|
|
|
|
|
case *types2.Const:
|
|
|
|
|
w.code(objConst)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.value(obj.Type(), obj.Val())
|
|
|
|
|
|
|
|
|
|
case *types2.Func:
|
|
|
|
|
decl, ok := w.p.funDecls[obj]
|
|
|
|
|
assert(ok)
|
|
|
|
|
sig := obj.Type().(*types2.Signature)
|
|
|
|
|
|
|
|
|
|
// Rewrite blank methods into blank functions.
|
|
|
|
|
// They aren't included in the receiver type's method set,
|
|
|
|
|
// and we still want to write them out to be compiled
|
|
|
|
|
// for regression tests.
|
|
|
|
|
// TODO(mdempsky): Change regress tests to avoid relying
|
|
|
|
|
// on blank functions/methods, so we can just ignore them
|
|
|
|
|
// altogether.
|
|
|
|
|
if recv := sig.Recv(); recv != nil {
|
|
|
|
|
assert(obj.Name() == "_")
|
|
|
|
|
assert(sig.TParams() == nil)
|
|
|
|
|
|
|
|
|
|
params := make([]*types2.Var, 1+sig.Params().Len())
|
|
|
|
|
params[0] = recv
|
|
|
|
|
for i := 0; i < sig.Params().Len(); i++ {
|
|
|
|
|
params[1+i] = sig.Params().At(i)
|
|
|
|
|
}
|
|
|
|
|
sig = types2.NewSignature(nil, types2.NewTuple(params...), sig.Results(), sig.Variadic())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(objFunc)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.typeParamNames(sig.TParams())
|
|
|
|
|
w.signature(sig)
|
|
|
|
|
w.pos(decl)
|
|
|
|
|
w.ext.funcExt(obj)
|
|
|
|
|
|
|
|
|
|
case *types2.TypeName:
|
|
|
|
|
decl, ok := w.p.typDecls[obj]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
if obj.IsAlias() {
|
|
|
|
|
w.code(objAlias)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.typ(obj.Type())
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
named := obj.Type().(*types2.Named)
|
|
|
|
|
assert(named.TArgs() == nil)
|
|
|
|
|
|
|
|
|
|
w.code(objType)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.typeParamNames(named.TParams())
|
|
|
|
|
w.ext.typeExt(obj)
|
|
|
|
|
w.typExpr(decl.Type)
|
|
|
|
|
|
|
|
|
|
w.len(named.NumMethods())
|
|
|
|
|
for i := 0; i < named.NumMethods(); i++ {
|
|
|
|
|
w.method(named.Method(i))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case *types2.Var:
|
|
|
|
|
w.code(objVar)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.typ(obj.Type())
|
|
|
|
|
w.ext.varExt(obj)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// typExpr writes the type represented by the given expression.
|
|
|
|
|
func (w *writer) typExpr(expr syntax.Expr) {
|
|
|
|
|
tv, ok := w.p.info.Types[expr]
|
|
|
|
|
assert(ok)
|
|
|
|
|
assert(tv.IsType())
|
|
|
|
|
w.typ(tv.Type)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) value(typ types2.Type, val constant.Value) {
|
|
|
|
|
w.sync(syncValue)
|
|
|
|
|
w.typ(typ)
|
|
|
|
|
w.rawValue(val)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) typeParamBounds(tparams []*types2.TypeName) {
|
|
|
|
|
w.sync(syncTypeParamBounds)
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
w.len(len(w.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
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
w.len(len(tparams))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 _, tparam := range tparams {
|
|
|
|
|
w.typ(tparam.Type().(*types2.TypeParam).Bound())
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) typeParamNames(tparams []*types2.TypeName) {
|
|
|
|
|
w.sync(syncTypeParamNames)
|
|
|
|
|
|
|
|
|
|
for _, tparam := range tparams {
|
|
|
|
|
w.pos(tparam)
|
|
|
|
|
w.localIdent(tparam)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) method(meth *types2.Func) {
|
|
|
|
|
decl, ok := w.p.funDecls[meth]
|
|
|
|
|
assert(ok)
|
|
|
|
|
sig := meth.Type().(*types2.Signature)
|
|
|
|
|
|
|
|
|
|
w.sync(syncMethod)
|
|
|
|
|
w.pos(meth)
|
|
|
|
|
w.selector(meth)
|
|
|
|
|
w.typeParamNames(sig.RParams())
|
|
|
|
|
w.param(sig.Recv())
|
|
|
|
|
w.signature(sig)
|
|
|
|
|
|
|
|
|
|
w.pos(decl) // XXX: Hack to workaround linker limitations.
|
|
|
|
|
w.ext.funcExt(meth)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// qualifiedIdent writes out the name of an object declared at package
|
|
|
|
|
// scope. (For now, it's also used to refer to local defined types.)
|
|
|
|
|
func (w *writer) qualifiedIdent(obj types2.Object) {
|
|
|
|
|
w.sync(syncSym)
|
|
|
|
|
|
|
|
|
|
name := obj.Name()
|
2021-06-28 22:41:50 -07:00
|
|
|
if isDefinedType(obj) && obj.Pkg() == w.p.curpkg {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
decl, ok := w.p.typDecls[obj.(*types2.TypeName)]
|
|
|
|
|
assert(ok)
|
2021-06-28 22:41:50 -07:00
|
|
|
if decl.gen != 0 {
|
|
|
|
|
// TODO(mdempsky): Find a better solution than embedding middle
|
|
|
|
|
// dot in the symbol name; this is terrible.
|
|
|
|
|
name = fmt.Sprintf("%s·%v", name, decl.gen)
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.pkg(obj.Pkg())
|
|
|
|
|
w.string(name)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): We should be able to omit pkg from both localIdent
|
|
|
|
|
// and selector, because they should always be known from context.
|
|
|
|
|
// However, past frustrations with this optimization in iexport make
|
|
|
|
|
// me a little nervous to try it again.
|
|
|
|
|
|
|
|
|
|
// localIdent writes the name of a locally declared object (i.e.,
|
|
|
|
|
// objects that can only be accessed by name, within the context of a
|
|
|
|
|
// particular function).
|
|
|
|
|
func (w *writer) localIdent(obj types2.Object) {
|
|
|
|
|
assert(!isGlobal(obj))
|
|
|
|
|
w.sync(syncLocalIdent)
|
|
|
|
|
w.pkg(obj.Pkg())
|
|
|
|
|
w.string(obj.Name())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// selector writes the name of a field or method (i.e., objects that
|
|
|
|
|
// can only be accessed using selector expressions).
|
|
|
|
|
func (w *writer) selector(obj types2.Object) {
|
|
|
|
|
w.sync(syncSelector)
|
|
|
|
|
w.pkg(obj.Pkg())
|
|
|
|
|
w.string(obj.Name())
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Compiler extensions
|
|
|
|
|
|
|
|
|
|
func (w *writer) funcExt(obj *types2.Func) {
|
|
|
|
|
decl, ok := w.p.funDecls[obj]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Extend these pragma validation flags to account
|
|
|
|
|
// for generics. E.g., linkname probably doesn't make sense at
|
|
|
|
|
// least.
|
|
|
|
|
|
|
|
|
|
pragma := asPragmaFlag(decl.Pragma)
|
|
|
|
|
if pragma&ir.Systemstack != 0 && pragma&ir.Nosplit != 0 {
|
|
|
|
|
w.p.errorf(decl, "go:nosplit and go:systemstack cannot be combined")
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if decl.Body != nil {
|
|
|
|
|
if pragma&ir.Noescape != 0 {
|
|
|
|
|
w.p.errorf(decl, "can only use //go:noescape with external func implementations")
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
if base.Flag.Complete || decl.Name.Value == "init" {
|
|
|
|
|
// Linknamed functions are allowed to have no body. Hopefully
|
|
|
|
|
// the linkname target has a body. See issue 23311.
|
|
|
|
|
if _, ok := w.p.linknames[obj]; !ok {
|
|
|
|
|
w.p.errorf(decl, "missing function body")
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
sig, block := obj.Type().(*types2.Signature), decl.Body
|
2021-06-28 22:41:50 -07:00
|
|
|
body, closureVars := w.p.bodyIdx(w.p.curpkg, sig, block, w.dict)
|
[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
|
|
|
assert(len(closureVars) == 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
|
|
|
w.sync(syncFuncExt)
|
|
|
|
|
w.pragmaFlag(pragma)
|
|
|
|
|
w.linkname(obj)
|
|
|
|
|
w.bool(false) // stub extension
|
[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
|
|
|
w.reloc(relocBody, 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
|
|
|
w.sync(syncEOF)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) typeExt(obj *types2.TypeName) {
|
|
|
|
|
decl, ok := w.p.typDecls[obj]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
w.sync(syncTypeExt)
|
|
|
|
|
|
|
|
|
|
w.pragmaFlag(asPragmaFlag(decl.Pragma))
|
|
|
|
|
|
|
|
|
|
// No LSym.SymIdx info yet.
|
|
|
|
|
w.int64(-1)
|
|
|
|
|
w.int64(-1)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) varExt(obj *types2.Var) {
|
|
|
|
|
w.sync(syncVarExt)
|
|
|
|
|
w.linkname(obj)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) linkname(obj types2.Object) {
|
|
|
|
|
w.sync(syncLinkname)
|
|
|
|
|
w.int64(-1)
|
|
|
|
|
w.string(w.p.linknames[obj])
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) pragmaFlag(p ir.PragmaFlag) {
|
|
|
|
|
w.sync(syncPragma)
|
|
|
|
|
w.int(int(p))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Function bodies
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
func (pw *pkgWriter) bodyIdx(pkg *types2.Package, sig *types2.Signature, block *syntax.BlockStmt, dict *writerDict) (idx int, closureVars []posObj) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w := pw.newWriter(relocBody, syncFuncBody)
|
2021-06-28 22:41:50 -07:00
|
|
|
w.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
|
|
|
|
|
|
|
|
w.funcargs(sig)
|
|
|
|
|
if w.bool(block != nil) {
|
|
|
|
|
w.stmts(block.List)
|
|
|
|
|
w.pos(block.Rbrace)
|
|
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
return w.flush(), w.closureVars
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 (w *writer) funcargs(sig *types2.Signature) {
|
|
|
|
|
do := func(params *types2.Tuple, result bool) {
|
|
|
|
|
for i := 0; i < params.Len(); i++ {
|
|
|
|
|
w.funcarg(params.At(i), result)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if recv := sig.Recv(); recv != nil {
|
|
|
|
|
w.funcarg(recv, false)
|
|
|
|
|
}
|
|
|
|
|
do(sig.Params(), false)
|
|
|
|
|
do(sig.Results(), true)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) funcarg(param *types2.Var, result bool) {
|
|
|
|
|
if param.Name() != "" || result {
|
|
|
|
|
w.addLocal(param)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
func (w *writer) addLocal(obj *types2.Var) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.sync(syncAddLocal)
|
|
|
|
|
idx := len(w.localsIdx)
|
[dev.typeparams] cmd/compile: record writer's stack at export data sync points
This CL extends the unified export data format's existing sync
mechanism to save writer stacks, controlled by the -d=syncframes debug
flag. This allows readers to provide more details when reporting
desync errors, which should simplify development of the data format
and the various reader/writer implementations.
For example, CL 328051 updated reader and writer, but missed making a
similar change to the linker (fix in CL 328054). Re-reviewing the CL
in isolation after the failure, it was not immediately obvious what
was going wrong. But the pair of stack traces below identifies exactly
what happened: it should have updated linker.relocFuncExt to write out
the new sync marker too.
```
data sync error: package "internal/abi", section 6, index 4, offset 536
found UseReloc, written at:
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/encoder.go:221: (*encoder).reloc +0x44
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/linker.go:214: (*linker).relocFuncExt +0x580
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/linker.go:233: (*linker).relocTypeExt +0x234
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/linker.go:161: (*linker).relocObj +0x2198
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/linker.go:64: (*linker).relocIdx +0x196
expected ImplicitTypes, reading at:
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/reader.go:796: (*reader).implicitTypes +0x36
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/reader.go:810: (*reader).addBody +0x81
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/reader.go:727: (*reader).funcExt +0x542
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/reader.go:651: (*reader).method +0x324
/home/mdempsky/wd/go/src/cmd/compile/internal/noder/reader.go:557: (*pkgReader).objIdx +0x2704
```
Change-Id: I911193edd2a965f81b7459f15fb613a773584685
Reviewed-on: https://go-review.googlesource.com/c/go/+/328909
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-06-17 01:12:23 -07:00
|
|
|
if enableSync {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.int(idx)
|
|
|
|
|
}
|
[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
|
|
|
if w.localsIdx == nil {
|
|
|
|
|
w.localsIdx = make(map[*types2.Var]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
|
|
|
w.localsIdx[obj] = idx
|
|
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
func (w *writer) useLocal(pos syntax.Pos, obj *types2.Var) {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.sync(syncUseObjLocal)
|
[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
|
|
|
|
|
|
|
|
if idx, ok := w.localsIdx[obj]; w.bool(ok) {
|
|
|
|
|
w.len(idx)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
idx, ok := w.closureVarsIdx[obj]
|
|
|
|
|
if !ok {
|
|
|
|
|
if w.closureVarsIdx == nil {
|
|
|
|
|
w.closureVarsIdx = make(map[*types2.Var]int)
|
|
|
|
|
}
|
|
|
|
|
idx = len(w.closureVars)
|
|
|
|
|
w.closureVars = append(w.closureVars, posObj{pos, obj})
|
|
|
|
|
w.closureVarsIdx[obj] = idx
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.len(idx)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) openScope(pos syntax.Pos) {
|
|
|
|
|
w.sync(syncOpenScope)
|
|
|
|
|
w.pos(pos)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) closeScope(pos syntax.Pos) {
|
|
|
|
|
w.sync(syncCloseScope)
|
|
|
|
|
w.pos(pos)
|
|
|
|
|
w.closeAnotherScope()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) closeAnotherScope() {
|
|
|
|
|
w.sync(syncCloseAnotherScope)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Statements
|
|
|
|
|
|
|
|
|
|
func (w *writer) stmt(stmt syntax.Stmt) {
|
|
|
|
|
var stmts []syntax.Stmt
|
|
|
|
|
if stmt != nil {
|
|
|
|
|
stmts = []syntax.Stmt{stmt}
|
|
|
|
|
}
|
|
|
|
|
w.stmts(stmts)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) stmts(stmts []syntax.Stmt) {
|
|
|
|
|
w.sync(syncStmts)
|
|
|
|
|
for _, stmt := range stmts {
|
|
|
|
|
w.stmt1(stmt)
|
|
|
|
|
}
|
|
|
|
|
w.code(stmtEnd)
|
|
|
|
|
w.sync(syncStmtsEnd)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) stmt1(stmt syntax.Stmt) {
|
|
|
|
|
switch stmt := stmt.(type) {
|
|
|
|
|
default:
|
|
|
|
|
w.p.unexpected("statement", stmt)
|
|
|
|
|
|
|
|
|
|
case nil, *syntax.EmptyStmt:
|
|
|
|
|
return
|
|
|
|
|
|
|
|
|
|
case *syntax.AssignStmt:
|
|
|
|
|
switch {
|
|
|
|
|
case stmt.Rhs == nil:
|
|
|
|
|
w.code(stmtIncDec)
|
|
|
|
|
w.op(binOps[stmt.Op])
|
|
|
|
|
w.expr(stmt.Lhs)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
|
|
|
|
|
case stmt.Op != 0 && stmt.Op != syntax.Def:
|
|
|
|
|
w.code(stmtAssignOp)
|
|
|
|
|
w.op(binOps[stmt.Op])
|
|
|
|
|
w.expr(stmt.Lhs)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.expr(stmt.Rhs)
|
|
|
|
|
|
|
|
|
|
default:
|
|
|
|
|
w.code(stmtAssign)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.exprList(stmt.Rhs)
|
[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
|
|
|
w.assignList(stmt.Lhs)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 *syntax.BlockStmt:
|
|
|
|
|
w.code(stmtBlock)
|
|
|
|
|
w.blockStmt(stmt)
|
|
|
|
|
|
|
|
|
|
case *syntax.BranchStmt:
|
|
|
|
|
w.code(stmtBranch)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.op(branchOps[stmt.Tok])
|
|
|
|
|
w.optLabel(stmt.Label)
|
|
|
|
|
|
|
|
|
|
case *syntax.CallStmt:
|
|
|
|
|
w.code(stmtCall)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.op(callOps[stmt.Tok])
|
|
|
|
|
w.expr(stmt.Call)
|
|
|
|
|
|
|
|
|
|
case *syntax.DeclStmt:
|
|
|
|
|
for _, decl := range stmt.DeclList {
|
|
|
|
|
w.declStmt(decl)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case *syntax.ExprStmt:
|
|
|
|
|
w.code(stmtExpr)
|
|
|
|
|
w.expr(stmt.X)
|
|
|
|
|
|
|
|
|
|
case *syntax.ForStmt:
|
|
|
|
|
w.code(stmtFor)
|
|
|
|
|
w.forStmt(stmt)
|
|
|
|
|
|
|
|
|
|
case *syntax.IfStmt:
|
|
|
|
|
w.code(stmtIf)
|
|
|
|
|
w.ifStmt(stmt)
|
|
|
|
|
|
|
|
|
|
case *syntax.LabeledStmt:
|
|
|
|
|
w.code(stmtLabel)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.label(stmt.Label)
|
|
|
|
|
w.stmt1(stmt.Stmt)
|
|
|
|
|
|
|
|
|
|
case *syntax.ReturnStmt:
|
|
|
|
|
w.code(stmtReturn)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.exprList(stmt.Results)
|
|
|
|
|
|
|
|
|
|
case *syntax.SelectStmt:
|
|
|
|
|
w.code(stmtSelect)
|
|
|
|
|
w.selectStmt(stmt)
|
|
|
|
|
|
|
|
|
|
case *syntax.SendStmt:
|
|
|
|
|
w.code(stmtSend)
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.expr(stmt.Chan)
|
|
|
|
|
w.expr(stmt.Value)
|
|
|
|
|
|
|
|
|
|
case *syntax.SwitchStmt:
|
|
|
|
|
w.code(stmtSwitch)
|
|
|
|
|
w.switchStmt(stmt)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) assignList(expr syntax.Expr) {
|
|
|
|
|
exprs := unpackListExpr(expr)
|
|
|
|
|
w.len(len(exprs))
|
|
|
|
|
|
|
|
|
|
for _, expr := range exprs {
|
|
|
|
|
if name, ok := expr.(*syntax.Name); ok && name.Value != "_" {
|
|
|
|
|
if obj, ok := w.p.info.Defs[name]; ok {
|
[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
|
|
|
obj := obj.(*types2.Var)
|
|
|
|
|
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.bool(true)
|
|
|
|
|
w.pos(obj)
|
|
|
|
|
w.localIdent(obj)
|
|
|
|
|
w.typ(obj.Type())
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Minimize locals index size by deferring
|
|
|
|
|
// this until the variables actually come into scope.
|
|
|
|
|
w.addLocal(obj)
|
|
|
|
|
continue
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.bool(false)
|
|
|
|
|
w.expr(expr)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) declStmt(decl syntax.Decl) {
|
|
|
|
|
switch decl := decl.(type) {
|
|
|
|
|
default:
|
|
|
|
|
w.p.unexpected("declaration", decl)
|
|
|
|
|
|
|
|
|
|
case *syntax.ConstDecl:
|
|
|
|
|
|
|
|
|
|
case *syntax.TypeDecl:
|
|
|
|
|
// Quirk: The legacy inliner doesn't support inlining functions
|
|
|
|
|
// with type declarations. Unified IR doesn't have any need to
|
|
|
|
|
// write out type declarations explicitly (they're always looked
|
|
|
|
|
// up via global index tables instead), so we just write out a
|
|
|
|
|
// marker so the reader knows to synthesize a fake declaration to
|
|
|
|
|
// prevent inlining.
|
|
|
|
|
if quirksMode() {
|
|
|
|
|
w.code(stmtTypeDeclHack)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case *syntax.VarDecl:
|
|
|
|
|
values := unpackListExpr(decl.Values)
|
|
|
|
|
|
|
|
|
|
// Quirk: When N variables are declared with N initialization
|
|
|
|
|
// values, we need to decompose that into N interleaved
|
|
|
|
|
// declarations+initializations, because it leads to different
|
|
|
|
|
// (albeit semantically equivalent) code generation.
|
|
|
|
|
if quirksMode() && len(decl.NameList) == len(values) {
|
|
|
|
|
for i, name := range decl.NameList {
|
|
|
|
|
w.code(stmtAssign)
|
|
|
|
|
w.pos(decl)
|
|
|
|
|
w.exprList(values[i])
|
[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
|
|
|
w.assignList(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
|
|
|
}
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(stmtAssign)
|
|
|
|
|
w.pos(decl)
|
|
|
|
|
w.exprList(decl.Values)
|
[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
|
|
|
w.assignList(namesAsExpr(decl.NameList))
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 (w *writer) blockStmt(stmt *syntax.BlockStmt) {
|
|
|
|
|
w.sync(syncBlockStmt)
|
|
|
|
|
w.openScope(stmt.Pos())
|
|
|
|
|
w.stmts(stmt.List)
|
|
|
|
|
w.closeScope(stmt.Rbrace)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) forStmt(stmt *syntax.ForStmt) {
|
|
|
|
|
w.sync(syncForStmt)
|
|
|
|
|
w.openScope(stmt.Pos())
|
|
|
|
|
|
|
|
|
|
if rang, ok := stmt.Init.(*syntax.RangeClause); w.bool(ok) {
|
|
|
|
|
w.pos(rang)
|
|
|
|
|
w.expr(rang.X)
|
[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
|
|
|
w.assignList(rang.Lhs)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 {
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.stmt(stmt.Init)
|
|
|
|
|
w.expr(stmt.Cond)
|
|
|
|
|
w.stmt(stmt.Post)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.blockStmt(stmt.Body)
|
|
|
|
|
w.closeAnotherScope()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) ifStmt(stmt *syntax.IfStmt) {
|
|
|
|
|
w.sync(syncIfStmt)
|
|
|
|
|
w.openScope(stmt.Pos())
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.stmt(stmt.Init)
|
|
|
|
|
w.expr(stmt.Cond)
|
|
|
|
|
w.blockStmt(stmt.Then)
|
|
|
|
|
w.stmt(stmt.Else)
|
|
|
|
|
w.closeAnotherScope()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) selectStmt(stmt *syntax.SelectStmt) {
|
|
|
|
|
w.sync(syncSelectStmt)
|
|
|
|
|
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.len(len(stmt.Body))
|
|
|
|
|
for i, clause := range stmt.Body {
|
|
|
|
|
if i > 0 {
|
|
|
|
|
w.closeScope(clause.Pos())
|
|
|
|
|
}
|
|
|
|
|
w.openScope(clause.Pos())
|
|
|
|
|
|
|
|
|
|
w.pos(clause)
|
|
|
|
|
w.stmt(clause.Comm)
|
|
|
|
|
w.stmts(clause.Body)
|
|
|
|
|
}
|
|
|
|
|
if len(stmt.Body) > 0 {
|
|
|
|
|
w.closeScope(stmt.Rbrace)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) switchStmt(stmt *syntax.SwitchStmt) {
|
|
|
|
|
w.sync(syncSwitchStmt)
|
|
|
|
|
|
|
|
|
|
w.openScope(stmt.Pos())
|
|
|
|
|
w.pos(stmt)
|
|
|
|
|
w.stmt(stmt.Init)
|
2021-07-09 17:47:15 -07:00
|
|
|
|
|
|
|
|
if guard, ok := stmt.Tag.(*syntax.TypeSwitchGuard); w.bool(ok) {
|
|
|
|
|
w.pos(guard)
|
|
|
|
|
if tag := guard.Lhs; w.bool(tag != nil) {
|
|
|
|
|
w.pos(tag)
|
|
|
|
|
w.string(tag.Value)
|
|
|
|
|
}
|
|
|
|
|
w.expr(guard.X)
|
|
|
|
|
} else {
|
|
|
|
|
w.expr(stmt.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
|
|
|
|
|
|
|
|
w.len(len(stmt.Body))
|
|
|
|
|
for i, clause := range stmt.Body {
|
|
|
|
|
if i > 0 {
|
|
|
|
|
w.closeScope(clause.Pos())
|
|
|
|
|
}
|
|
|
|
|
w.openScope(clause.Pos())
|
|
|
|
|
|
|
|
|
|
w.pos(clause)
|
|
|
|
|
w.exprList(clause.Cases)
|
|
|
|
|
|
|
|
|
|
if obj, ok := w.p.info.Implicits[clause]; ok {
|
|
|
|
|
// TODO(mdempsky): These pos details are quirkish, but also
|
|
|
|
|
// necessary so the variable's position is correct for DWARF
|
|
|
|
|
// scope assignment later. It would probably be better for us to
|
|
|
|
|
// instead just set the variable's DWARF scoping info earlier so
|
|
|
|
|
// we can give it the correct position information.
|
|
|
|
|
pos := clause.Pos()
|
|
|
|
|
if typs := unpackListExpr(clause.Cases); len(typs) != 0 {
|
|
|
|
|
pos = typeExprEndPos(typs[len(typs)-1])
|
|
|
|
|
}
|
|
|
|
|
w.pos(pos)
|
|
|
|
|
|
|
|
|
|
obj := obj.(*types2.Var)
|
|
|
|
|
w.typ(obj.Type())
|
|
|
|
|
w.addLocal(obj)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.stmts(clause.Body)
|
|
|
|
|
}
|
|
|
|
|
if len(stmt.Body) > 0 {
|
|
|
|
|
w.closeScope(stmt.Rbrace)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.closeScope(stmt.Rbrace)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) label(label *syntax.Name) {
|
|
|
|
|
w.sync(syncLabel)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): Replace label strings with dense indices.
|
|
|
|
|
w.string(label.Value)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) optLabel(label *syntax.Name) {
|
|
|
|
|
w.sync(syncOptLabel)
|
|
|
|
|
if w.bool(label != nil) {
|
|
|
|
|
w.label(label)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Expressions
|
|
|
|
|
|
|
|
|
|
func (w *writer) expr(expr syntax.Expr) {
|
|
|
|
|
expr = unparen(expr) // skip parens; unneeded after typecheck
|
|
|
|
|
|
|
|
|
|
obj, targs := lookupObj(w.p.info, expr)
|
|
|
|
|
|
|
|
|
|
if tv, ok := w.p.info.Types[expr]; ok {
|
2021-07-13 09:09:32 -07:00
|
|
|
// TODO(mdempsky): Be more judicious about which types are marked as "needed".
|
|
|
|
|
w.needType(tv.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
|
|
|
if tv.IsType() {
|
|
|
|
|
w.code(exprType)
|
|
|
|
|
w.typ(tv.Type)
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if tv.Value != nil {
|
|
|
|
|
pos := expr.Pos()
|
|
|
|
|
if quirksMode() {
|
|
|
|
|
if obj != nil {
|
|
|
|
|
// Quirk: IR (and thus iexport) doesn't track position
|
|
|
|
|
// information for uses of declared objects.
|
|
|
|
|
pos = syntax.Pos{}
|
|
|
|
|
} else if tv.Value.Kind() == constant.String {
|
|
|
|
|
// Quirk: noder.sum picks a particular position for certain
|
|
|
|
|
// string concatenations.
|
|
|
|
|
pos = sumPos(expr)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(exprConst)
|
|
|
|
|
w.pos(pos)
|
|
|
|
|
w.value(tv.Type, tv.Value)
|
|
|
|
|
|
|
|
|
|
// TODO(mdempsky): These details are only important for backend
|
|
|
|
|
// diagnostics. Explore writing them out separately.
|
|
|
|
|
w.op(constExprOp(expr))
|
|
|
|
|
w.string(syntax.String(expr))
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if obj != nil {
|
[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
|
|
|
if isGlobal(obj) {
|
|
|
|
|
w.code(exprName)
|
|
|
|
|
w.obj(obj, targs)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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
|
|
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
obj := obj.(*types2.Var)
|
|
|
|
|
assert(len(targs) == 0)
|
|
|
|
|
|
|
|
|
|
w.code(exprLocal)
|
|
|
|
|
w.useLocal(expr.Pos(), 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
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch expr := expr.(type) {
|
|
|
|
|
default:
|
|
|
|
|
w.p.unexpected("expression", expr)
|
|
|
|
|
|
|
|
|
|
case nil: // absent slice index, for condition, or switch tag
|
|
|
|
|
w.code(exprNone)
|
|
|
|
|
|
|
|
|
|
case *syntax.Name:
|
|
|
|
|
assert(expr.Value == "_")
|
|
|
|
|
w.code(exprBlank)
|
|
|
|
|
|
|
|
|
|
case *syntax.CompositeLit:
|
|
|
|
|
w.code(exprCompLit)
|
|
|
|
|
w.compLit(expr)
|
|
|
|
|
|
|
|
|
|
case *syntax.FuncLit:
|
|
|
|
|
w.code(exprFuncLit)
|
|
|
|
|
w.funcLit(expr)
|
|
|
|
|
|
|
|
|
|
case *syntax.SelectorExpr:
|
|
|
|
|
sel, ok := w.p.info.Selections[expr]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
w.code(exprSelector)
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.selector(sel.Obj())
|
|
|
|
|
|
|
|
|
|
case *syntax.IndexExpr:
|
|
|
|
|
tv, ok := w.p.info.Types[expr.Index]
|
|
|
|
|
assert(ok && tv.IsValue())
|
|
|
|
|
|
|
|
|
|
w.code(exprIndex)
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.expr(expr.Index)
|
|
|
|
|
|
|
|
|
|
case *syntax.SliceExpr:
|
|
|
|
|
w.code(exprSlice)
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
for _, n := range &expr.Index {
|
|
|
|
|
w.expr(n)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case *syntax.AssertExpr:
|
|
|
|
|
w.code(exprAssert)
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.expr(expr.Type)
|
|
|
|
|
|
|
|
|
|
case *syntax.Operation:
|
|
|
|
|
if expr.Y == nil {
|
|
|
|
|
w.code(exprUnaryOp)
|
|
|
|
|
w.op(unOps[expr.Op])
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(exprBinaryOp)
|
|
|
|
|
w.op(binOps[expr.Op])
|
|
|
|
|
w.expr(expr.X)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.expr(expr.Y)
|
|
|
|
|
|
|
|
|
|
case *syntax.CallExpr:
|
2021-07-09 17:47:15 -07:00
|
|
|
tv, ok := w.p.info.Types[expr.Fun]
|
|
|
|
|
assert(ok)
|
|
|
|
|
if tv.IsType() {
|
|
|
|
|
assert(len(expr.ArgList) == 1)
|
|
|
|
|
assert(!expr.HasDots)
|
|
|
|
|
|
|
|
|
|
w.code(exprConvert)
|
|
|
|
|
w.typ(tv.Type)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.expr(expr.ArgList[0])
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
2021-07-09 19:47:45 -07:00
|
|
|
writeFunExpr := func() {
|
|
|
|
|
if selector, ok := unparen(expr.Fun).(*syntax.SelectorExpr); ok {
|
|
|
|
|
if sel, ok := w.p.info.Selections[selector]; ok && sel.Kind() == types2.MethodVal {
|
|
|
|
|
w.expr(selector.X)
|
|
|
|
|
w.bool(true) // method call
|
|
|
|
|
w.pos(selector)
|
|
|
|
|
w.selector(sel.Obj())
|
|
|
|
|
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
|
|
|
|
2021-07-09 19:47:45 -07:00
|
|
|
if inf, ok := w.p.info.Inferred[expr]; ok {
|
|
|
|
|
obj, _ := lookupObj(w.p.info, expr.Fun)
|
|
|
|
|
assert(obj != 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
|
|
|
|
2021-07-09 19:47:45 -07:00
|
|
|
// As if w.expr(expr.Fun), but using inf.TArgs instead.
|
|
|
|
|
w.code(exprName)
|
|
|
|
|
w.obj(obj, inf.TArgs)
|
|
|
|
|
} else {
|
|
|
|
|
w.expr(expr.Fun)
|
|
|
|
|
}
|
|
|
|
|
w.bool(false) // not a method call (i.e., normal function 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
|
|
|
}
|
|
|
|
|
|
2021-07-09 19:47:45 -07:00
|
|
|
w.code(exprCall)
|
|
|
|
|
writeFunExpr()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
w.pos(expr)
|
|
|
|
|
w.exprs(expr.ArgList)
|
|
|
|
|
w.bool(expr.HasDots)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) compLit(lit *syntax.CompositeLit) {
|
|
|
|
|
tv, ok := w.p.info.Types[lit]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
w.sync(syncCompLit)
|
|
|
|
|
w.pos(lit)
|
|
|
|
|
w.typ(tv.Type)
|
|
|
|
|
|
|
|
|
|
typ := tv.Type
|
|
|
|
|
if ptr, ok := typ.Underlying().(*types2.Pointer); ok {
|
|
|
|
|
typ = ptr.Elem()
|
|
|
|
|
}
|
|
|
|
|
str, isStruct := typ.Underlying().(*types2.Struct)
|
|
|
|
|
|
|
|
|
|
w.len(len(lit.ElemList))
|
|
|
|
|
for i, elem := range lit.ElemList {
|
|
|
|
|
if isStruct {
|
|
|
|
|
if kv, ok := elem.(*syntax.KeyValueExpr); ok {
|
|
|
|
|
// use position of expr.Key rather than of elem (which has position of ':')
|
|
|
|
|
w.pos(kv.Key)
|
|
|
|
|
w.len(fieldIndex(w.p.info, str, kv.Key.(*syntax.Name)))
|
|
|
|
|
elem = kv.Value
|
|
|
|
|
} else {
|
|
|
|
|
w.pos(elem)
|
|
|
|
|
w.len(i)
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
if kv, ok := elem.(*syntax.KeyValueExpr); w.bool(ok) {
|
|
|
|
|
// use position of expr.Key rather than of elem (which has position of ':')
|
|
|
|
|
w.pos(kv.Key)
|
|
|
|
|
w.expr(kv.Key)
|
|
|
|
|
elem = kv.Value
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
w.pos(elem)
|
|
|
|
|
w.expr(elem)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) funcLit(expr *syntax.FuncLit) {
|
|
|
|
|
tv, ok := w.p.info.Types[expr]
|
|
|
|
|
assert(ok)
|
|
|
|
|
sig := tv.Type.(*types2.Signature)
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
body, closureVars := w.p.bodyIdx(w.p.curpkg, sig, expr.Body, w.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
|
|
|
w.sync(syncFuncLit)
|
|
|
|
|
w.pos(expr)
|
|
|
|
|
w.pos(expr.Type) // for QuirksMode
|
|
|
|
|
w.signature(sig)
|
|
|
|
|
|
|
|
|
|
w.len(len(closureVars))
|
[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
|
|
|
for _, cv := range closureVars {
|
|
|
|
|
w.pos(cv.pos)
|
|
|
|
|
if quirksMode() {
|
|
|
|
|
cv.pos = expr.Body.Rbrace
|
|
|
|
|
}
|
|
|
|
|
w.useLocal(cv.pos, cv.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
|
|
|
}
|
|
|
|
|
|
[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
|
|
|
w.reloc(relocBody, 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
|
|
|
}
|
|
|
|
|
|
|
|
|
|
type posObj struct {
|
|
|
|
|
pos syntax.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
|
|
|
obj *types2.Var
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 (w *writer) exprList(expr syntax.Expr) {
|
|
|
|
|
w.sync(syncExprList)
|
|
|
|
|
w.exprs(unpackListExpr(expr))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) exprs(exprs []syntax.Expr) {
|
|
|
|
|
if len(exprs) == 0 {
|
|
|
|
|
assert(exprs == nil)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.sync(syncExprs)
|
|
|
|
|
w.len(len(exprs))
|
|
|
|
|
for _, expr := range exprs {
|
|
|
|
|
w.expr(expr)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) op(op ir.Op) {
|
|
|
|
|
// TODO(mdempsky): Remove in favor of explicit codes? Would make
|
|
|
|
|
// export data more stable against internal refactorings, but low
|
|
|
|
|
// priority at the moment.
|
|
|
|
|
assert(op != 0)
|
|
|
|
|
w.sync(syncOp)
|
|
|
|
|
w.len(int(op))
|
|
|
|
|
}
|
|
|
|
|
|
2021-07-13 09:09:32 -07:00
|
|
|
func (w *writer) needType(typ types2.Type) {
|
|
|
|
|
// Decompose tuple into component element types.
|
|
|
|
|
if typ, ok := typ.(*types2.Tuple); ok {
|
|
|
|
|
for i := 0; i < typ.Len(); i++ {
|
|
|
|
|
w.needType(typ.At(i).Type())
|
|
|
|
|
}
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if info := w.p.typIdx(typ, w.dict); info.derived {
|
|
|
|
|
w.dict.derived[info.idx].needed = 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
|
|
|
// @@@ Package initialization
|
|
|
|
|
|
|
|
|
|
// Caution: This code is still clumsy, because toolstash -cmp is
|
|
|
|
|
// particularly sensitive to it.
|
|
|
|
|
|
|
|
|
|
type typeDeclGen struct {
|
|
|
|
|
*syntax.TypeDecl
|
|
|
|
|
gen int
|
2021-06-28 22:41:50 -07:00
|
|
|
|
|
|
|
|
// Implicit type parameters in scope at this type declaration.
|
|
|
|
|
implicits []*types2.TypeName
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
}
|
|
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
type fileImports struct {
|
|
|
|
|
importedEmbed, importedUnsafe 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
type declCollector struct {
|
|
|
|
|
pw *pkgWriter
|
|
|
|
|
typegen *int
|
|
|
|
|
file *fileImports
|
|
|
|
|
withinFunc bool
|
2021-06-28 22:41:50 -07:00
|
|
|
implicits []*types2.TypeName
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (c *declCollector) withTParams(obj types2.Object) *declCollector {
|
|
|
|
|
tparams := objTypeParams(obj)
|
|
|
|
|
if len(tparams) == 0 {
|
|
|
|
|
return c
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
copy := *c
|
|
|
|
|
copy.implicits = copy.implicits[:len(copy.implicits):len(copy.implicits)]
|
|
|
|
|
copy.implicits = append(copy.implicits, objTypeParams(obj)...)
|
|
|
|
|
return ©
|
2021-06-23 14:04:11 -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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
func (c *declCollector) Visit(n syntax.Node) syntax.Visitor {
|
|
|
|
|
pw := c.pw
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
switch n := n.(type) {
|
|
|
|
|
case *syntax.File:
|
|
|
|
|
pw.checkPragmas(n.Pragma, ir.GoBuildPragma, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
case *syntax.ImportDecl:
|
|
|
|
|
pw.checkPragmas(n.Pragma, 0, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
switch pkgNameOf(pw.info, n).Imported().Path() {
|
|
|
|
|
case "embed":
|
|
|
|
|
c.file.importedEmbed = true
|
|
|
|
|
case "unsafe":
|
|
|
|
|
c.file.importedUnsafe = 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
case *syntax.ConstDecl:
|
|
|
|
|
pw.checkPragmas(n.Pragma, 0, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
case *syntax.FuncDecl:
|
|
|
|
|
pw.checkPragmas(n.Pragma, funcPragmas, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
obj := pw.info.Defs[n.Name].(*types2.Func)
|
|
|
|
|
pw.funDecls[obj] = 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
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
return c.withTParams(obj)
|
|
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
case *syntax.TypeDecl:
|
|
|
|
|
obj := pw.info.Defs[n.Name].(*types2.TypeName)
|
2021-06-28 22:41:50 -07:00
|
|
|
d := typeDeclGen{TypeDecl: n, implicits: c.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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
if n.Alias {
|
|
|
|
|
pw.checkPragmas(n.Pragma, 0, false)
|
|
|
|
|
} else {
|
|
|
|
|
pw.checkPragmas(n.Pragma, typePragmas, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
// Assign a unique ID to function-scoped defined types.
|
2021-06-28 22:41:50 -07:00
|
|
|
if c.withinFunc {
|
2021-06-23 14:04:11 -07:00
|
|
|
*c.typegen++
|
|
|
|
|
d.gen = *c.typegen
|
|
|
|
|
}
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
pw.typDecls[obj] = d
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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
|
|
|
// TODO(mdempsky): Omit? Not strictly necessary; only matters for
|
|
|
|
|
// type declarations within function literals within parameterized
|
|
|
|
|
// type declarations, but types2 the function literals will be
|
|
|
|
|
// constant folded away.
|
|
|
|
|
return c.withTParams(obj)
|
|
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
case *syntax.VarDecl:
|
|
|
|
|
pw.checkPragmas(n.Pragma, 0, 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
|
|
|
|
2021-06-23 14:04:11 -07:00
|
|
|
if p, ok := n.Pragma.(*pragmas); ok && len(p.Embeds) > 0 {
|
|
|
|
|
if err := checkEmbed(n, c.file.importedEmbed, c.withinFunc); err != nil {
|
|
|
|
|
pw.errorf(p.Embeds[0].Pos, "%s", err)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Workaround for #46208. For variable declarations that
|
|
|
|
|
// declare multiple variables and have an explicit type
|
|
|
|
|
// expression, the type expression is evaluated multiple
|
|
|
|
|
// times. This affects toolstash -cmp, because iexport is
|
|
|
|
|
// sensitive to *types.Type pointer identity.
|
|
|
|
|
if quirksMode() && n.Type != nil {
|
|
|
|
|
tv, ok := pw.info.Types[n.Type]
|
|
|
|
|
assert(ok)
|
|
|
|
|
assert(tv.IsType())
|
|
|
|
|
for _, name := range n.NameList {
|
|
|
|
|
obj := pw.info.Defs[name].(*types2.Var)
|
|
|
|
|
pw.dups.add(obj.Type(), tv.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
|
|
|
}
|
2021-06-23 14:04:11 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
case *syntax.BlockStmt:
|
|
|
|
|
if !c.withinFunc {
|
|
|
|
|
copy := *c
|
|
|
|
|
copy.withinFunc = true
|
|
|
|
|
return ©
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return c
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) collectDecls(noders []*noder) {
|
|
|
|
|
var typegen int
|
|
|
|
|
for _, p := range noders {
|
|
|
|
|
var file fileImports
|
|
|
|
|
|
|
|
|
|
syntax.Walk(p.file, &declCollector{
|
|
|
|
|
pw: pw,
|
|
|
|
|
typegen: &typegen,
|
|
|
|
|
file: &file,
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
})
|
|
|
|
|
|
|
|
|
|
pw.cgoPragmas = append(pw.cgoPragmas, p.pragcgobuf...)
|
|
|
|
|
|
|
|
|
|
for _, l := range p.linknames {
|
2021-06-23 14:04:11 -07:00
|
|
|
if !file.importedUnsafe {
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
pw.errorf(l.pos, "//go:linkname only allowed in Go files that import \"unsafe\"")
|
|
|
|
|
continue
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
switch obj := pw.curpkg.Scope().Lookup(l.local).(type) {
|
|
|
|
|
case *types2.Func, *types2.Var:
|
|
|
|
|
if _, ok := pw.linknames[obj]; !ok {
|
|
|
|
|
pw.linknames[obj] = l.remote
|
|
|
|
|
} else {
|
|
|
|
|
pw.errorf(l.pos, "duplicate //go:linkname for %s", l.local)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
default:
|
|
|
|
|
// TODO(mdempsky): Enable after #42938 is fixed.
|
|
|
|
|
if false {
|
|
|
|
|
pw.errorf(l.pos, "//go:linkname must refer to declared function or variable")
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (pw *pkgWriter) checkPragmas(p syntax.Pragma, allowed ir.PragmaFlag, embedOK bool) {
|
|
|
|
|
if p == nil {
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
pragma := p.(*pragmas)
|
|
|
|
|
|
|
|
|
|
for _, pos := range pragma.Pos {
|
|
|
|
|
if pos.Flag&^allowed != 0 {
|
|
|
|
|
pw.errorf(pos.Pos, "misplaced compiler directive")
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if !embedOK {
|
|
|
|
|
for _, e := range pragma.Embeds {
|
|
|
|
|
pw.errorf(e.Pos, "misplaced go:embed directive")
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) pkgInit(noders []*noder) {
|
|
|
|
|
if quirksMode() {
|
|
|
|
|
posBases := posBasesOf(noders)
|
|
|
|
|
w.len(len(posBases))
|
|
|
|
|
for _, posBase := range posBases {
|
|
|
|
|
w.posBase(posBase)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
objs := importedObjsOf(w.p.curpkg, w.p.info, noders)
|
|
|
|
|
w.len(len(objs))
|
|
|
|
|
for _, obj := range objs {
|
|
|
|
|
w.qualifiedIdent(obj)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.len(len(w.p.cgoPragmas))
|
|
|
|
|
for _, cgoPragma := range w.p.cgoPragmas {
|
|
|
|
|
w.strings(cgoPragma)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.sync(syncDecls)
|
|
|
|
|
for _, p := range noders {
|
|
|
|
|
for _, decl := range p.file.DeclList {
|
|
|
|
|
w.pkgDecl(decl)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
w.code(declEnd)
|
|
|
|
|
|
|
|
|
|
w.sync(syncEOF)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) pkgDecl(decl syntax.Decl) {
|
|
|
|
|
switch decl := decl.(type) {
|
|
|
|
|
default:
|
|
|
|
|
w.p.unexpected("declaration", decl)
|
|
|
|
|
|
|
|
|
|
case *syntax.ImportDecl:
|
|
|
|
|
|
|
|
|
|
case *syntax.ConstDecl:
|
|
|
|
|
w.code(declOther)
|
|
|
|
|
w.pkgObjs(decl.NameList...)
|
|
|
|
|
|
|
|
|
|
case *syntax.FuncDecl:
|
|
|
|
|
obj := w.p.info.Defs[decl.Name].(*types2.Func)
|
|
|
|
|
sig := obj.Type().(*types2.Signature)
|
|
|
|
|
|
|
|
|
|
if sig.RParams() != nil || sig.TParams() != nil {
|
|
|
|
|
break // skip generic functions
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if recv := sig.Recv(); recv != nil && obj.Name() != "_" {
|
|
|
|
|
w.code(declMethod)
|
|
|
|
|
w.typ(recvBase(recv))
|
|
|
|
|
w.selector(obj)
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(declFunc)
|
|
|
|
|
w.pkgObjs(decl.Name)
|
|
|
|
|
|
|
|
|
|
case *syntax.TypeDecl:
|
|
|
|
|
if len(decl.TParamList) != 0 {
|
|
|
|
|
break // skip generic type decls
|
|
|
|
|
}
|
|
|
|
|
|
2021-06-28 22:41:50 -07:00
|
|
|
if decl.Name.Value == "_" {
|
|
|
|
|
break // skip blank type decls
|
|
|
|
|
}
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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
|
|
|
name := w.p.info.Defs[decl.Name].(*types2.TypeName)
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
2021-05-13 20:23:13 -07:00
|
|
|
// Skip type declarations for interfaces that are only usable as
|
|
|
|
|
// type parameter bounds.
|
|
|
|
|
if iface, ok := name.Type().Underlying().(*types2.Interface); ok && iface.IsConstraint() {
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Skip aliases to uninstantiated generic types.
|
|
|
|
|
// TODO(mdempsky): Revisit after #46477 is resolved.
|
|
|
|
|
if name.IsAlias() {
|
|
|
|
|
named, ok := name.Type().(*types2.Named)
|
|
|
|
|
if ok && len(named.TParams()) != 0 && len(named.TArgs()) == 0 {
|
|
|
|
|
break
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
w.code(declOther)
|
|
|
|
|
w.pkgObjs(decl.Name)
|
|
|
|
|
|
|
|
|
|
case *syntax.VarDecl:
|
|
|
|
|
w.code(declVar)
|
|
|
|
|
w.pos(decl)
|
|
|
|
|
w.pkgObjs(decl.NameList...)
|
|
|
|
|
w.exprList(decl.Values)
|
|
|
|
|
|
|
|
|
|
var embeds []pragmaEmbed
|
|
|
|
|
if p, ok := decl.Pragma.(*pragmas); ok {
|
|
|
|
|
embeds = p.Embeds
|
|
|
|
|
}
|
|
|
|
|
w.len(len(embeds))
|
|
|
|
|
for _, embed := range embeds {
|
|
|
|
|
w.pos(embed.Pos)
|
|
|
|
|
w.strings(embed.Patterns)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func (w *writer) pkgObjs(names ...*syntax.Name) {
|
|
|
|
|
w.sync(syncDeclNames)
|
|
|
|
|
w.len(len(names))
|
|
|
|
|
|
|
|
|
|
for _, name := range names {
|
|
|
|
|
obj, ok := w.p.info.Defs[name]
|
|
|
|
|
assert(ok)
|
|
|
|
|
|
|
|
|
|
w.sync(syncDeclName)
|
|
|
|
|
w.obj(obj, nil)
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// @@@ Helpers
|
|
|
|
|
|
|
|
|
|
// isDefinedType reports whether obj is a defined type.
|
|
|
|
|
func isDefinedType(obj types2.Object) bool {
|
|
|
|
|
if obj, ok := obj.(*types2.TypeName); ok {
|
|
|
|
|
return !obj.IsAlias()
|
|
|
|
|
}
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// isGlobal reports whether obj was declared at package scope.
|
|
|
|
|
//
|
|
|
|
|
// Caveat: blank objects are not declared.
|
|
|
|
|
func isGlobal(obj types2.Object) bool {
|
|
|
|
|
return obj.Parent() == obj.Pkg().Scope()
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// lookupObj returns the object that expr refers to, if any. If expr
|
|
|
|
|
// is an explicit instantiation of a generic object, then the type
|
|
|
|
|
// arguments are returned as well.
|
|
|
|
|
func lookupObj(info *types2.Info, expr syntax.Expr) (obj types2.Object, targs []types2.Type) {
|
|
|
|
|
if index, ok := expr.(*syntax.IndexExpr); ok {
|
|
|
|
|
if inf, ok := info.Inferred[index]; ok {
|
|
|
|
|
targs = inf.TArgs
|
|
|
|
|
} else {
|
|
|
|
|
args := unpackListExpr(index.Index)
|
|
|
|
|
|
|
|
|
|
if len(args) == 1 {
|
|
|
|
|
tv, ok := info.Types[args[0]]
|
|
|
|
|
assert(ok)
|
|
|
|
|
if tv.IsValue() {
|
|
|
|
|
return // normal index expression
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
targs = make([]types2.Type, len(args))
|
|
|
|
|
for i, arg := range args {
|
|
|
|
|
tv, ok := info.Types[arg]
|
|
|
|
|
assert(ok)
|
|
|
|
|
assert(tv.IsType())
|
|
|
|
|
targs[i] = tv.Type
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
expr = index.X
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Strip package qualifier, if present.
|
|
|
|
|
if sel, ok := expr.(*syntax.SelectorExpr); ok {
|
|
|
|
|
if !isPkgQual(info, sel) {
|
|
|
|
|
return // normal selector expression
|
|
|
|
|
}
|
|
|
|
|
expr = sel.Sel
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if name, ok := expr.(*syntax.Name); ok {
|
|
|
|
|
obj, _ = info.Uses[name]
|
|
|
|
|
}
|
|
|
|
|
return
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// isPkgQual reports whether the given selector expression is a
|
|
|
|
|
// package-qualified identifier.
|
|
|
|
|
func isPkgQual(info *types2.Info, sel *syntax.SelectorExpr) bool {
|
|
|
|
|
if name, ok := sel.X.(*syntax.Name); ok {
|
|
|
|
|
_, isPkgName := info.Uses[name].(*types2.PkgName)
|
|
|
|
|
return isPkgName
|
|
|
|
|
}
|
|
|
|
|
return false
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// recvBase returns the base type for the given receiver parameter.
|
|
|
|
|
func recvBase(recv *types2.Var) *types2.Named {
|
|
|
|
|
typ := recv.Type()
|
|
|
|
|
if ptr, ok := typ.(*types2.Pointer); ok {
|
|
|
|
|
typ = ptr.Elem()
|
|
|
|
|
}
|
|
|
|
|
return typ.(*types2.Named)
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// namesAsExpr returns a list of names as a syntax.Expr.
|
|
|
|
|
func namesAsExpr(names []*syntax.Name) syntax.Expr {
|
|
|
|
|
if len(names) == 1 {
|
|
|
|
|
return names[0]
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
exprs := make([]syntax.Expr, len(names))
|
|
|
|
|
for i, name := range names {
|
|
|
|
|
exprs[i] = name
|
|
|
|
|
}
|
|
|
|
|
return &syntax.ListExpr{ElemList: exprs}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// fieldIndex returns the index of the struct field named by key.
|
|
|
|
|
func fieldIndex(info *types2.Info, str *types2.Struct, key *syntax.Name) int {
|
|
|
|
|
field := info.Uses[key].(*types2.Var)
|
|
|
|
|
|
|
|
|
|
for i := 0; i < str.NumFields(); i++ {
|
|
|
|
|
if str.Field(i) == field {
|
|
|
|
|
return i
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
panic(fmt.Sprintf("%s: %v is not a field of %v", key.Pos(), field, str))
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// objTypeParams returns the type parameters on the given object.
|
|
|
|
|
func objTypeParams(obj types2.Object) []*types2.TypeName {
|
|
|
|
|
switch obj := obj.(type) {
|
|
|
|
|
case *types2.Func:
|
2021-06-28 22:41:50 -07:00
|
|
|
sig := obj.Type().(*types2.Signature)
|
|
|
|
|
if sig.Recv() != nil {
|
|
|
|
|
return sig.RParams()
|
|
|
|
|
}
|
|
|
|
|
return sig.TParams()
|
[dev.typeparams] cmd/compile: unified IR construction
This CL adds a new unified IR construction mode to the frontend. It's
purely additive, and all files include "UNREVIEWED" at the top, like
how types2 was initially imported. The next CL adds a -d=unified flag
to actually enable unified IR mode.
See below for more details, but some highlights:
1. It adds ~6kloc (excluding enum listings and stringer output), but I
estimate it will allow removing ~14kloc (see CL 324670, including its
commit message);
2. When enabled by default, it passes more tests than -G=3 does (see
CL 325213 and CL 324673);
3. Without requiring any new code, it supports inlining of more code
than the current inliner (see CL 324574; contrast CL 283112 and CL
266203, which added support for inlining function literals and type
switches, respectively);
4. Aside from dictionaries (which I intend to add still), its support
for generics is more complete (e.g., it fully supports local types,
including local generic types within generic functions and
instantiating generic types with local types; see
test/typeparam/nested.go);
5. It supports lazy loading of types and objects for types2 type
checking;
6. It supports re-exporting of types, objects, and inline bodies
without needing to parse them into IR;
7. The new export data format has extensive support for debugging with
"sync" markers, so mistakes during development are easier to catch;
8. When compiling with -d=inlfuncswithclosures=0, it enables "quirks
mode" where it generates output that passes toolstash -cmp.
--
The new unified IR pipeline combines noding, stenciling, inlining, and
import/export into a single, shared code path. Previously, IR trees
went through multiple phases of copying during compilation:
1. "Noding": the syntax AST is copied into the initial IR form. To
support generics, there's now also "irgen", which implements the same
idea, but takes advantage of types2 type-checking results to more
directly construct IR.
2. "Stenciling": generic IR forms are copied into instantiated IR
forms, substituting type parameters as appropriate.
3. "Inlining": the inliner made backup copies of inlinable functions,
and then copied them again when inlining into a call site, with some
modifications (e.g., updating position information, rewriting variable
references, changing "return" statements into "goto").
4. "Importing/exporting": the exporter wrote out the IR as saved by
the inliner, and then the importer read it back as to be used by the
inliner again. Normal functions are imported/exported "desugared",
while generic functions are imported/exported in source form.
These passes are all conceptually the same thing: make a copy of a
function body, maybe with some minor changes/substitutions. However,
they're all completely separate implementations that frequently run
into the same issues because IR has many nuanced corner cases.
For example, inlining currently doesn't support local defined types,
"range" loops, or labeled "for"/"switch" statements, because these
require special handling around Sym references. We've recently
extended the inliner to support new features like inlining type
switches and function literals, and they've had issues. The exporter
only knows how to export from IR form, so when re-exporting inlinable
functions (e.g., methods on imported types that are exposed via
exported APIs), these functions may need to be imported as IR for the
sole purpose of being immediately exported back out again.
By unifying all of these modes of copying into a single code path that
cleanly separates concerns, we eliminate many of these possible
issues. Some recent examples:
1. Issues #45743 and #46472 were issues where type switches were
mishandled by inlining and stenciling, respectively; but neither of
these affected unified IR, because it constructs type switches using
the exact same code as for normal functions.
2. CL 325409 fixes an issue in stenciling with implicit conversion of
values of type-parameter type to variables of interface type, but this
issue did not affect unified IR.
Change-Id: I5a05991fe16d68bb0f712503e034cb9f2d19e296
Reviewed-on: https://go-review.googlesource.com/c/go/+/324573
Trust: Matthew Dempsky <mdempsky@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@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 *types2.TypeName:
|
|
|
|
|
if !obj.IsAlias() {
|
|
|
|
|
return obj.Type().(*types2.Named).TParams()
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
return nil
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
func asPragmaFlag(p syntax.Pragma) ir.PragmaFlag {
|
|
|
|
|
if p == nil {
|
|
|
|
|
return 0
|
|
|
|
|
}
|
|
|
|
|
return p.(*pragmas).Flag
|
|
|
|
|
}
|