This CL is originally based on CL 484838 from rajbarik@uber.com.
Add a new PGO-based devirtualize pass. This pass conditionally
devirtualizes interface calls for the hottest callee. That is, it
performs a transformation like:
type Iface interface {
Foo()
}
type Concrete struct{}
func (Concrete) Foo() {}
func foo(i Iface) {
i.Foo()
}
to:
func foo(i Iface) {
if c, ok := i.(Concrete); ok {
c.Foo()
} else {
i.Foo()
}
}
The primary benefit of this transformation is enabling inlining of the
direct calls.
Today this change has no impact on the escape behavior, as the fallback
interface always forces an escape. But improving escape analysis to take
advantage of this is an area of potential work.
This CL is the bare minimum of a devirtualization implementation. There
are still numerous limitations:
* Callees not directly referenced in the current package can be missed
(even if they are in the transitive dependences).
* Callees not in the transitive dependencies of the current package are
missed.
* Only interface method calls are supported, not other indirect function
calls.
* Multiple calls to compatible interfaces on the same line cannot be
distinguished and will use the same callee target.
* Callees that only partially implement an interface (they are embedded
in another type that completes the interface) cannot be devirtualized.
* Others, mentioned in TODOs.
Fixes#59959
Change-Id: I8bedb516139695ee4069650b099d05957b7ce5ee
Reviewed-on: https://go-review.googlesource.com/c/go/+/492436
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
We will soon have PGO specialization. It doesn't make sense for the
debug flag to have inline in the name, so rename it to pgodebug.
pgoinline is now a flag that can be used to disable PGO inlining.
Devirtualization will have a similar debug flag.
For #59959.
Change-Id: I9770ff1f0d132dfa3cd417018a887a1bd5555bba
Reviewed-on: https://go-review.googlesource.com/c/go/+/494716
Auto-Submit: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
In race mode (or other instrumentation mode), if the caller is in
a regular package and the callee is in a norace (or noinstrument)
package, don't inline. Otherwise, when the caller is instumented
it will also affect the inlined callee.
An example is sync.(*Mutex).Unlock, which is typically not inlined
but with PGO it can be inlined into a regular function, which is
then get instrumented. But the rest of the sync package, in
particular, the Lock function is not instrumented, causing the
race detector to signal false race.
Change-Id: Ia78bb602c6da63a34ec2909b9a82646bf20873f3
Reviewed-on: https://go-review.googlesource.com/c/go/+/495595
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Types are either static (for compiler-created types) or heap
allocated and always reachable (for reflection-created types, held
in the central map). So there is no need to escape types.
With CL 408826 reflect.Value does not always escape. Some functions
that escapes Value.typ would make the Value escape without this CL.
Had to add a special case for the inliner to keep (*Value).Type
still inlineable.
Change-Id: I7c14d35fd26328347b509a06eb5bd1534d40775f
Reviewed-on: https://go-review.googlesource.com/c/go/+/413474
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The RedirectEdges logic is fragile and not quite complete (doesn't
update in-edges), which adds overhead to maintaining this package.
In my opinion, the post-inlining graph doesn't provide as much value as
the pre-inlining graph. Even the latter I am not convinced should be in
the compiler rather than an external tool, but it is comparatively
easier to maintain.
Drop it for now. Perhaps we'll want it back in the future for tracking
follow-up optimizations, but for now keep things simple.
Change-Id: I3133a2eb97893a14a6770547f96a3f1796798d17
Reviewed-on: https://go-review.googlesource.com/c/go/+/494655
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
Delete the "InlineSCCOnePass" debugging flag and the inliner fallback
code that kicks in if it is used. The change it was intended to guard
has been working on tip for some time, no need for the fallback any
more.
Updates #58905.
Change-Id: I2e1dbc7640902d9402213db5ad338be03deb96c5
Reviewed-on: https://go-review.googlesource.com/c/go/+/492015
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
[This is a roll-forward of CL 479095, which was reverted due to a bad
interaction between inlining and escape analysis, then later fixed
first with an attempt in CL 482355, then again in CL 484859, and then
one more time with CL 492135.]
Currently, when the inliner is determining if a function is
inlineable, it descends into the bodies of closures constructed by
that function. This has several unfortunate consequences:
- If the closure contains a disallowed operation (e.g., a defer), then
the outer function can't be inlined. It makes sense that the
*closure* can't be inlined in this case, but it doesn't make sense
to punish the function that constructs the closure.
- The hairiness of the closure counts against the inlining budget of
the outer function. Since we currently copy the closure body when
inlining the outer function, this makes sense from the perspective
of export data size and binary size, but ultimately doesn't make
much sense from the perspective of what should be inlineable.
- Since the inliner walks into every closure created by an outer
function in addition to starting a walk at every closure, this adds
an n^2 factor to inlinability analysis.
This CL simply drops this behavior.
In std, this makes 57 more functions inlinable, and disallows inlining
for 10 (due to the basic instability of our bottom-up inlining
approach), for an net increase of 47 inlinable functions (+0.6%).
This will help significantly with the performance of the functions to
be added for #56102, which have a somewhat complicated nesting of
closures with a performance-critical fast path.
The downside of this seems to be a potential increase in export data
and text size, but the practical impact of this seems to be
negligible:
│ before │ after │
│ bytes │ bytes vs base │
Go/binary 15.12Mi ± 0% 15.14Mi ± 0% +0.16% (n=1)
Go/text 5.220Mi ± 0% 5.237Mi ± 0% +0.32% (n=1)
Compile/binary 22.92Mi ± 0% 22.94Mi ± 0% +0.07% (n=1)
Compile/text 8.428Mi ± 0% 8.435Mi ± 0% +0.08% (n=1)
Change-Id: I5f75fcceb177f05853996b75184a486528eafe96
Reviewed-on: https://go-review.googlesource.com/c/go/+/492017
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
[This is a roll-forward of CL 484859, this time including a fix for
issue #59709. The call to do dead function marking was taking place in
the wrong spot, causing it to run more than once if generics were
instantiated.]
This patch generalizes the code in the inliner that marks unreferenced
hidden closure functions as dead. Rather than doing the marking on the
fly (previous approach), this new approach does a single pass at the
end of inlining, which catches more dead functions.
Change-Id: I0e079ad755c21295477201acbd7e1a732a98fffd
Reviewed-on: https://go-review.googlesource.com/c/go/+/492016
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Symbol names in the final executable apply escaping to the final
component of a package path (main in example.com becomes
example%2ecom.main).
ir.PkgFuncName does not perform this escaping, meaning we'd fail to
match functions that are escaped in the profile.
Add ir.LinkFuncName which does perform escaping and use it for PGO.
Fixes#59887.
Change-Id: I10634d63d99d0a6fd2f72b929ab35ea227e1336f
Reviewed-on: https://go-review.googlesource.com/c/go/+/490555
Reviewed-by: Cherry Mui <cherryyz@google.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This reverts commit http://go.dev/cl//484859
Reason for revert: causes linker errors in a number of google-internal tests.
Change-Id: I322252f784a46d2b1d447ebcdca86ce14bc0cc91
Reviewed-on: https://go-review.googlesource.com/c/go/+/485755
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
[This is a roll-forward of CL 479095, which was reverted due to a bad
interaction between inlining and escape analysis, then later fixed
fist with an attempt in CL 482355, then again in 484859 .]
Currently, when the inliner is determining if a function is
inlineable, it descends into the bodies of closures constructed by
that function. This has several unfortunate consequences:
- If the closure contains a disallowed operation (e.g., a defer), then
the outer function can't be inlined. It makes sense that the
*closure* can't be inlined in this case, but it doesn't make sense
to punish the function that constructs the closure.
- The hairiness of the closure counts against the inlining budget of
the outer function. Since we currently copy the closure body when
inlining the outer function, this makes sense from the perspective
of export data size and binary size, but ultimately doesn't make
much sense from the perspective of what should be inlineable.
- Since the inliner walks into every closure created by an outer
function in addition to starting a walk at every closure, this adds
an n^2 factor to inlinability analysis.
This CL simply drops this behavior.
In std, this makes 57 more functions inlinable, and disallows inlining
for 10 (due to the basic instability of our bottom-up inlining
approach), for an net increase of 47 inlinable functions (+0.6%).
This will help significantly with the performance of the functions to
be added for #56102, which have a somewhat complicated nesting of
closures with a performance-critical fast path.
The downside of this seems to be a potential increase in export data
and text size, but the practical impact of this seems to be
negligible:
│ before │ after │
│ bytes │ bytes vs base │
Go/binary 15.12Mi ± 0% 15.14Mi ± 0% +0.16% (n=1)
Go/text 5.220Mi ± 0% 5.237Mi ± 0% +0.32% (n=1)
Compile/binary 22.92Mi ± 0% 22.94Mi ± 0% +0.07% (n=1)
Compile/text 8.428Mi ± 0% 8.435Mi ± 0% +0.08% (n=1)
Updates #56102.
Change-Id: I6e938d596992ffb473cf51e7e598f372ce08deb0
Reviewed-on: https://go-review.googlesource.com/c/go/+/484860
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
This patch generalizes the code in the inliner that marks unreferenced
hidden closure functions as dead. Rather than doing the marking on the
fly (previous approach), this new approach does a single pass at the
end of inlining, which catches more dead functions.
Fixes#59638.
Updates #59404.
Updates #59547.
Change-Id: I54fd63e9e37c9123b08a3e7def7d1989919bba91
Reviewed-on: https://go-review.googlesource.com/c/go/+/484859
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
CL 399694 added constant-fold switch early in compilation. So function:
func f() string {
switch intSize {
case 32:
return "32"
case 64:
return "64"
default:
panic("unreachable")
}
}
will be constant-fold to:
func f() string {
switch intSize {
case 64:
return "64"
}
}
When this function get inlined, there is a check whether we can delay
declaring the result parameter until the "return" statement. For the
original function, we can't delay the result, because there's more than
one return statement. However, the constant-fold one can, because
there's on one return statement in the body now. The result parameter
~R0 ends up declaring inside the switch statement scope.
Now, when walking the switch statement, it's re-written into if-else
statement. Without typecheck.EvalConst, the if condition "if 64 == 64"
is passed as-is to the ssa generation pass. Because "64 == 64" is not a
constant, the ssagen creates normal blocks for branching the results.
This confuses the liveness analysis, because ~R0 is only live inside the
if block. With typecheck.EvalConst, "64 == 64" is evaluated to "true",
so ssagen can branch the result without emitting conditional blocks.
Instead, the constant-fold can be re-written as:
switch {
case true:
// Body
}
So it does not depend on the delay results check during inlining. Adding
a test, which will fail when typecheck.EvalConst is removed, so we can
do the cleanup without breaking things.
Change-Id: I638730bb147140de84260653741431b807ff2f15
Reviewed-on: https://go-review.googlesource.com/c/go/+/484316
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
This reverts commit http://go.dev/cl/c/482356.
Reason for revert: Reverting this change again, since it is causing additional failures in google-internal testing.
Change-Id: I9234946f62e5bb18c2f873a65e8b298d04af0809
Reviewed-on: https://go-review.googlesource.com/c/go/+/484735
Reviewed-by: Florian Zenker <floriank@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Auto-Submit: Than McIntosh <thanm@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Standard inlining has a reduced maximum cost of callees (20 instead of
80) when inlining into a "big" function, to limit how much bigger we
will make an already big function.
When adding PGO hot call budget increases, we inadvertently bypassed
this "big" function restriction, allowing hot calls of up to
inlineHotMaxBudget, even into big functions.
Add the restriction back, even for hot calls. If a function is already
very large, we probably shouldn't inline even more.
A very important note here is that function "big"-ness is computed prior
to any inlining. One potential problem with PGO is that many hot calls
inline into an initially-small function and ultimately make it very
large. This CL does nothing to address that case, which would require
recomputing size after inlining.
This CL has no impact on sweet PGO benchmarks. I specifically dug into
tile38, which contained 0 hot big functions. Other benchmarks are
probably similar.
Change-Id: I3b6304eaf7738a219359d4b8bb121d68babfea8b
Reviewed-on: https://go-review.googlesource.com/c/go/+/482157
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
[This is a roll-forward of CL 479095, which was reverted due to a bad
interaction between inlining and escape analysis since fixed in CL 482355.]
Currently, when the inliner is determining if a function is
inlineable, it descends into the bodies of closures constructed by
that function. This has several unfortunate consequences:
- If the closure contains a disallowed operation (e.g., a defer), then
the outer function can't be inlined. It makes sense that the
*closure* can't be inlined in this case, but it doesn't make sense
to punish the function that constructs the closure.
- The hairiness of the closure counts against the inlining budget of
the outer function. Since we currently copy the closure body when
inlining the outer function, this makes sense from the perspective
of export data size and binary size, but ultimately doesn't make
much sense from the perspective of what should be inlineable.
- Since the inliner walks into every closure created by an outer
function in addition to starting a walk at every closure, this adds
an n^2 factor to inlinability analysis.
This CL simply drops this behavior.
In std, this makes 57 more functions inlinable, and disallows inlining
for 10 (due to the basic instability of our bottom-up inlining
approach), for an net increase of 47 inlinable functions (+0.6%).
This will help significantly with the performance of the functions to
be added for #56102, which have a somewhat complicated nesting of
closures with a performance-critical fast path.
The downside of this seems to be a potential increase in export data
and text size, but the practical impact of this seems to be
negligible:
│ before │ after │
│ bytes │ bytes vs base │
Go/binary 15.12Mi ± 0% 15.14Mi ± 0% +0.16% (n=1)
Go/text 5.220Mi ± 0% 5.237Mi ± 0% +0.32% (n=1)
Compile/binary 22.92Mi ± 0% 22.94Mi ± 0% +0.07% (n=1)
Compile/text 8.428Mi ± 0% 8.435Mi ± 0% +0.08% (n=1)
Updates #56102.
Change-Id: I1f4fc96c71609c8feb59fecdb92b69ba7e3b5b41
Reviewed-on: https://go-review.googlesource.com/c/go/+/482356
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
When a closure is inlined, it may contain other hidden closures, which
the inliner will duplicate, rendering the original nested closures as
unreachable. Because they are unreachable, they don't get processed in
escape analysis, meaning that go/defer statements don't get rewritten,
which can then in turn trigger errors in walk. This patch looks for
nested hidden closures and marks them as dead, so that they can be
skipped later on in the compilation flow. NB: if during escape
analysis we rediscover a hidden closure (due to an explicit reference)
that was previously marked dead, revive it at that point.
Fixes#59404.
Change-Id: I76db1e9cf1ee38bd1147aeae823f916dbbbf081b
Reviewed-on: https://go-review.googlesource.com/c/go/+/482355
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Currently, when the inliner is determining if a function is
inlineable, it descends into the bodies of closures constructed by
that function. This has several unfortunate consequences:
- If the closure contains a disallowed operation (e.g., a defer), then
the outer function can't be inlined. It makes sense that the
*closure* can't be inlined in this case, but it doesn't make sense
to punish the function that constructs the closure.
- The hairiness of the closure counts against the inlining budget of
the outer function. Since we currently copy the closure body when
inlining the outer function, this makes sense from the perspective
of export data size and binary size, but ultimately doesn't make
much sense from the perspective of what should be inlineable.
- Since the inliner walks into every closure created by an outer
function in addition to starting a walk at every closure, this adds
an n^2 factor to inlinability analysis.
This CL simply drops this behavior.
In std, this makes 57 more functions inlinable, and disallows inlining
for 10 (due to the basic instability of our bottom-up inlining
approach), for an net increase of 47 inlinable functions (+0.6%).
This will help significantly with the performance of the functions to
be added for #56102, which have a somewhat complicated nesting of
closures with a performance-critical fast path.
The downside of this seems to be a potential increase in export data
and text size, but the practical impact of this seems to be
negligible:
│ before │ after │
│ bytes │ bytes vs base │
Go/binary 15.12Mi ± 0% 15.14Mi ± 0% +0.16% (n=1)
Go/text 5.220Mi ± 0% 5.237Mi ± 0% +0.32% (n=1)
Compile/binary 22.92Mi ± 0% 22.94Mi ± 0% +0.07% (n=1)
Compile/text 8.428Mi ± 0% 8.435Mi ± 0% +0.08% (n=1)
Change-Id: Ie9e38104fed5689a94c368288653fd7cb4b7a35e
Reviewed-on: https://go-review.googlesource.com/c/go/+/479095
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>
This patch changes the relative order of "CanInline" and "InlineCalls"
operations within the inliner for clumps of functions corresponding to
strongly connected components in the call graph. This helps increase
the amount of inlining within SCCs, particularly in Go's runtime
package, which has a couple of very large SCCs.
For a given SCC of the form { fn1, fn2, ... fnk }, the inliner would
(prior to this point) walk through the list of functions and for each
function first compute inlinability ("CanInline") and then perform
inlining ("InlineCalls"). This meant that if there was an inlinable
call from fn3 to fn4 (for example), this call would never be inlined,
since at the point fn3 was visited, we would not have computed
inlinability for fn4.
We now do inlinability analysis for all functions in an SCC first,
then do actual inlining for everything. This results in 47 additional
inlines in the Go runtime package (a fairly modest increase
percentage-wise of 0.6%).
Updates #58905.
Change-Id: I48dbb1ca16f0b12f256d9eeba8cf7f3e6dd853cd
Reviewed-on: https://go-review.googlesource.com/c/go/+/474955
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
The compiler used to generate ONAME node with nil Func for them, so the
inliner can still analyze, but could not generate inline call for them
anyway.
CL 436961 attempts to create ONAME node with non-nil Func, causing the
inliner complains about missing body reader.
This CL makes inliner recognize type eq/hash functions, and mark them as
non-inlineable. Longer term, if we do want to inline these functions, we
need to integrate the code generation into Unified IR frontend.
Updates #58572
Change-Id: Icdd4dda03711929faa3d48fe2d9886568471f0bc
Reviewed-on: https://go-review.googlesource.com/c/go/+/469017
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
ir.VisitFuncsBottomUp returns recursive==true for functions which
call themselves. It also returns any closures inside that function.
We don't want to report the closures as recursive, as they really
aren't. Only the containing function is recursive.
Fixes#54159
Change-Id: I3b4d6710a389ec1d6b250ba8a7065f2e985bdbe1
Reviewed-on: https://go-review.googlesource.com/c/go/+/463233
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
To simplify backend analysis, we normalize variadic and method calls:
variadic calls are rewritten with an explicit slice argument, and
method calls are turned into function calls that pass the receiver
argument as the first parameter.
But because we've been supporting multiple frontends, this
normalization was scattered in various later passes. Now that we're
back to just one frontend, we can move the normalization forward into
typecheck (where most other IR normalization already happens).
Change-Id: Idd05ae231fc180ae3dd1664452414f6b6d578962
Reviewed-on: https://go-review.googlesource.com/c/go/+/463737
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Unified IR generates uniform IR for "a, b = f()" to be able to insert
implicit conversion expressions, but the result is somewhat more
verbose and trips up the inliner's naive cost metrics.
The hairyVisitor.doNode method was already adjusted to account for
this, but isBigFunc needs the same adjustment.
Fixes#57563.
Change-Id: Ia8d86a6e314ec60190c78f40ace4fb30dadc4413
Reviewed-on: https://go-review.googlesource.com/c/go/+/460395
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Benny Siegert <bsiegert@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This CL removes most of the code specific to the nounified
frontend. To simplify review, it's a strict remove-only CL.
Updates #57410.
Change-Id: Ic3196570aa4286618c1d5e7fd0d0e6601a18195b
Reviewed-on: https://go-review.googlesource.com/c/go/+/458620
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Now that GOEXPERIMENT=nounified is removed, we can assume InlineCall
and HaveInlineBody will always be overridden with the unified
frontend's implementations. Similarly, we can assume expandDecl will
never be called.
This CL changes the code paths into Fatalfs, so subsequent CLs can
remove all the unreachable code.
Updates #57410.
Change-Id: I2a0c3edb32916c30dd63c4dce4f1bd6f18e07468
Reviewed-on: https://go-review.googlesource.com/c/go/+/458618
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
This flag forced the compiler to eagerly type check all available
inline function bodies, which presumably was useful in the early days
of implementing inlining support. However, it shouldn't have any
significance with the unified frontend, since the same code paths are
used for constructing normal function bodies as for inlining.
Updates #57410.
Change-Id: I6842cf86bcd0fbf22ac336f2fc0b7b8fe14bccca
Reviewed-on: https://go-review.googlesource.com/c/go/+/458617
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This CL removes the GOEXPERIMENT=nounified knob, and any conditional
statements that depend on that knob. Further CLs to remove unreachable
code follow this one.
Updates #57410.
Change-Id: I39c147e1a83601c73f8316a001705778fee64a91
Reviewed-on: https://go-review.googlesource.com/c/go/+/458615
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
When -m=N (where N > 1) is in effect, include a note in the trace
output if a given function is considered "big" during inlining
analysis, since this causes the inliner to be less aggressive. If a
small change to a large function happens to nudge it over the large
function threshold, it can be confusing for developers, thus it's
probably worth including this info in the remark output.
Change-Id: Id31a1b76371ab1ef9265ba28a377f97b0247d0a7
Reviewed-on: https://go-review.googlesource.com/c/go/+/460317
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Keith Randall <khr@google.com>
This is the second round to look for spelling mistakes. This time the
manual sifting of the result list was made easier by filtering out
capitalized and camelcase words.
grep -r --include '*.go' -E '^// .*$' . | aspell list | grep -E -x '[A-Za-z]{1}[a-z]*' | sort | uniq
This PR will be imported into Gerrit with the title and first
comment (this text) used to generate the subject and body of
the Gerrit change.
Change-Id: Ie8a2092aaa7e1f051aa90f03dbaf2b9aaf5664a9
GitHub-Last-Rev: fc2bd6e0c5
GitHub-Pull-Request: golang/go#57737
Reviewed-on: https://go-review.googlesource.com/c/go/+/461595
Auto-Submit: Ian Lance Taylor <iant@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Run-TryBot: Ian Lance Taylor <iant@google.com>
Reviewed-by: Robert Griesemer <gri@google.com>
Currently, we use CDF to compute a weight threshold and then use
the weight threshold to determine whether a call site is hot. As
when we compute the CDF we already have a list of hot call sites
that make up the given percentage of the CDF, just use that list.
Also, when computing the CDF threshold, include the very last node
that makes it to go over the threshold. (I.e. if the CDF threshold
is 50% and one hot node takes 60% of weight, we should include that
node instead of excluding it. In practice it rarely matters,
probably only for testing and micro-benchmarks.)
Change-Id: I535ae9cd6b679609e247c3d0d9ee572c1a1187cc
Reviewed-on: https://go-review.googlesource.com/c/go/+/450737
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Adjust PGO inlining default parameters to 99% CDF threshold and
2000 budget. Benchmark results (mostly from Sweet) show that this
set of parameters performs reasonably well, with a few percent
speedup at the cost of a few percent binary size increase.
Also rename the debug flags to start with "pgo", to make it clear
that they are related to PGO.
Change-Id: I0749249b1298d1dc55a28993c37b3185f9d7639d
Reviewed-on: https://go-review.googlesource.com/c/go/+/449477
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
If an imported, non-generic function F transitively calls a generic
function G[T], we may need to call CanInline on G[T].
While here, we can also take advantage of the fact that we know G[T]
was already seen and compiled in an imported package, so we don't need
to call InlineCalls or add it to typecheck.Target.Decls. This saves us
from wasting compile time re-creating DUPOK symbols that we know
already exist in the imported package's link objects.
Fixes#56280.
Change-Id: I3336786bee01616ee9f2b18908738e4ca41c8102
Reviewed-on: https://go-review.googlesource.com/c/go/+/443535
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
- Include the callee names in hot call inlining message.
- Print the graph when pgoinline >= 2.
Change-Id: Iceb89b5f18cefc69ab9256aca9a910743d22ec0f
Reviewed-on: https://go-review.googlesource.com/c/go/+/448496
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Appears to be a typo in CL 447315.
Change-Id: I9f380a3c7521f5ac5a1d7e271eaa60bd4bbcfb29
Reviewed-on: https://go-review.googlesource.com/c/go/+/448515
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Rather than matching calls to edges in the profile based directly on
line number in the source file, use the line offset from the start of
the function. This makes matching robust to changes in the source file
above the function containing the call.
The start line in the profile comes from Function.start_line, which is
included in Go pprof output since CL 438255.
Currently it is an error if no samples set start_line to help users
detect profiles missing this information. In the future, we should
fallback to using absolute lines, which is better than nothing.
For #55022.
Change-Id: Ie621950cfee1fef8fb200907a2a3f1ded41d04fa
Reviewed-on: https://go-review.googlesource.com/c/go/+/447315
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
With CL 447015, we identify hot callees from edge weights, but
the code only traverses edges for calls from the current package.
If the callee is in a different package, when compiling that
package, the edge was not visited, so the callee was not actually
marked inline candidate. This CL fixes it by traversing all hot
edges.
Change-Id: If668c1a16ebe34e3474376b88ab3a84be76b8562
Reviewed-on: https://go-review.googlesource.com/c/go/+/448015
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Currently in PGO we use a percentage threshold to determine if a
callsite is hot. This CL uses a different method -- treating the
hottest callsites that make up cumulatively top X% of total edge
weights as hot (X=95 for now). This default might work better for
a wider range of profiles. (The absolute threshold can still be
changed by a flag.)
For #55022.
Change-Id: I7e3b6f0c3cf23f9a89dd5994c10075b498bf14ee
Reviewed-on: https://go-review.googlesource.com/c/go/+/447016
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
Currently, with PGO, the inliner uses node weights to decide if a
function is inlineable (with a larger budget). But the actual
inlining is determined by the weight of the call edge. There is a
discrepancy that, if a callee node is hot but the call edge is not,
it would not inlined, and marking the callee inlineable would of
no use.
Instead of using two kinds of weights, we just use the edge
weights to decide inlineability. If a function is the callee of a
hot call edge, its inlineability is determined with a larger
threshold. For a function that exceeds the regular inlining budget,
it is still inlined only when the call edge is hot, as it would
exceed the regular inlining cost for non-hot call sites, even if
it is marked inlineable.
For #55022.
Change-Id: I93fa9919fc6bcbb394e6cfe54ec96a96eede08f7
Reviewed-on: https://go-review.googlesource.com/c/go/+/447015
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The global ListOfHotCallSites set is used to communicate between
CanInline and InlineCalls the set of call sites that InlineCalls may
increase the budget for.
CanInline clears this map on each call, thus assuming that
InlineCalls(x) is called immediately after CanInline(x). This assumption
is false, as CanInline (among other cases) is recursive (CanInline ->
hairyVisitor.doNode -> inlCallee -> CanInline).
When this assumption proves false, we will lose the opportunity to
inline hot calls.
This CL is the least invasive fix for this. ListOfHotCallSites is
actually just a subset of the candHotEdgeMap, with CallSiteInfo.Callee
cleared. candHotEdgeMap doesn't actually need to distinguish based on
Callee, so we can drop callee from candHotEdgeMap as well and just use
that directly [1].
Later CLs should do more work to remove the globals entirely.
For cmd/compile, this inceases the number of PGO inlined functions by
~50% for one set of PGO parameters. I have no evaluated performance
impact.
[1] This is something that we likely want to change in the future.
For #55022.
Change-Id: I57735958d651f6dfa9bd296499841213d20e1706
Reviewed-on: https://go-review.googlesource.com/c/go/+/446755
Auto-Submit: Michael Pratt <mpratt@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This patch fixes a typo/bug introduced in CL 441858 where when pattern
matching a coverage counter access we were looking at an assingment
node instead of the assignment LHS, and fixes a similar problem in
atomic counter update pattern matching introduced in CL 444835. In
both of these cases the bug was not caught because the test intended
to lock down the behavior was written incorrectly (wasn't
instrumenting what the test author thought it was instrumenting,
ouch).
Change-Id: I6e6ac3beacf12ef1a817de5527340b639f0bb044
Reviewed-on: https://go-review.googlesource.com/c/go/+/446258
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Than McIntosh <thanm@google.com>
Since pgo is a new package, it is reasonably straightforward to
encapsulate its state into a non-global object that we pass around,
which will help keep it isolated.
There are no functional changes in this CL, just packaging up the
globals into a new object.
There are two major pieces of cleanup remaining:
1. reflectdata and noder have separate InlineCalls calls for method
wrappers. The Profile is not plumbed there yet, but this is not a
regression as the globals were previously set only right around the
main inlining pass in gc.Main.
2. pgo.ListOfHotCallSites is still global, as it will require more work
to clean up. It is effectively a local variable in InlinePackage,
except that it assumes that InlineCalls is immediately preceded by a
CanInline call for the same function. This is not necessarily true
due to the recursive nature of CanInline. This also means that some
InlineCalls calls may be missing the list of hot callsites right now.
For #55022.
Change-Id: Ic1fe41f73df96861c65f8bfeecff89862b367290
Reviewed-on: https://go-review.googlesource.com/c/go/+/446303
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
Parts of package pgo fetch the line number of a node by parsing the
number out of the string returned from ir.Line().
This is indirect and inefficient, so it should be replaced with a more
direct lookup. It is also potentially buggy: ir.Line uses
ctxt.OutermostPos, i.e., the line number where an inlined node in
inlined. We want ctxt.InnermostPos, because that is the line number used
in pprof profiles that we are matching against (See comments on
OutermostPos and InnermostPos).
I'm not sure whether this was an active, as we use ir.Line before and
during inlining. I think we could see CALL nodes with OutermostPos !=
InnermostPos during midstack inlining, but I am not sure. Regardless,
explicitly using the desired position is clearer.
For #55022.
Change-Id: Ic640761c9e1d01cacbf91f3aaeaf284ad7e38dbd
Reviewed-on: https://go-review.googlesource.com/c/go/+/446302
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Michael Pratt <mpratt@google.com>
Auto-Submit: Michael Pratt <mpratt@google.com>
This patch fixes up a bug in the inliner's special case code for
coverage counter updates, which was not properly working for
-covermode=atomic compilations.
Updates #56044.
Change-Id: I9e309312b123121c3df02862623bdbab1f6c6a4b
Reviewed-on: https://go-review.googlesource.com/c/go/+/441858
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
The linker needs FuncInfo metadata for all inlined functions. This is
typically handled by gc.enqueueFunc calling ir.InitLSym for all function
declarations in typecheck.Target.Decls (ir.UseClosure adds all closures
to Decls).
However, non-trivial closures in Decls are ignored, and are insteaded
enqueued when walk of the calling function discovers them.
This presents a problem for direct calls to closures. Inlining will
replace the entire closure definition with its body, which hides the
closure from walk and thus suppresses symbol creation.
Explicitly create a symbol early in this edge case to ensure we keep
this metadata.
InitLSym needs to move out of ssagen to avoid a circular dependency (it
doesn't have anything to do with ssa anyway). There isn't a great place
for it, so I placed it in ir, which seemed least objectionable.
The added test triggers one of these inlined direct non-trivial closure
calls, though the test needs CL 429637 to fail, which adds a FuncInfo
assertion to the linker. Note that the test must use "run" instead of
"compile" since the assertion is in the linker, and "compiler" doesn't
run the linker.
Fixes#54959.
Change-Id: I0bd1db4f3539a78da260934cd968372b7aa92546
Reviewed-on: https://go-review.googlesource.com/c/go/+/436240
Run-TryBot: Michael Pratt <mpratt@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Add a new "coverage counter" classification for variables to be used
for storing code coverage counter values (somewhat in the same way
that we identify fuzzer counters). Tagging such variables allows us to
aggregate them in the linker, and to treat updates specially.
Updates #51430.
Change-Id: Ib49fb05736ffece98bcc2f7a7c37e991b7f67bbb
Reviewed-on: https://go-review.googlesource.com/c/go/+/401235
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
