cmd/compile: build compiler with PGO

Reapplies CL 451292, which is reverted at CL 495475. The failure
is fixed at CL 495595.

Build the compiler with PGO. As go build -pgo=auto is enabled by
default, we just need to store a profile in the compiler's
directory.

The profile is collected from building all std and cmd packages on
Linux/AMD64 machine, using profile.sh.

This improves the compiler speed. On Linux/AMD64,
name        old time/op       new time/op       delta
Template          138ms ± 5%        136ms ± 4%  -1.44%  (p=0.005 n=36+39)
Unicode           147ms ± 4%        140ms ± 4%  -4.99%  (p=0.000 n=40+39)
GoTypes           780ms ± 3%        778ms ± 4%    ~     (p=0.172 n=39+39)
Compiler          105ms ± 5%         99ms ± 7%  -5.64%  (p=0.000 n=40+40)
SSA               5.83s ± 6%        5.80s ± 6%    ~     (p=0.556 n=40+40)
Flate            89.0ms ± 5%       87.0ms ± 6%  -2.18%  (p=0.000 n=40+40)
GoParser          172ms ± 4%        167ms ± 4%  -2.72%  (p=0.000 n=39+40)
Reflect           333ms ± 4%        333ms ± 3%    ~     (p=0.426 n=40+39)
Tar               128ms ± 4%        126ms ± 4%  -1.82%  (p=0.000 n=39+39)
XML               173ms ± 4%        170ms ± 4%  -1.39%  (p=0.000 n=39+40)
[Geo mean]        253ms             248ms       -2.13%

The profile is pretty transferable. Using the same profile, we
see a bigger win on Darwin/ARM64,
name        old time/op       new time/op       delta
Template         71.0ms ± 2%       68.3ms ± 2%  -3.90%  (p=0.000 n=20+20)
Unicode          71.8ms ± 2%       66.8ms ± 2%  -6.90%  (p=0.000 n=20+20)
GoTypes           444ms ± 1%        428ms ± 1%  -3.53%  (p=0.000 n=19+20)
Compiler         48.9ms ± 3%       45.6ms ± 3%  -6.81%  (p=0.000 n=20+20)
SSA               3.25s ± 2%        3.09s ± 1%  -5.03%  (p=0.000 n=19+20)
Flate            44.0ms ± 2%       42.3ms ± 2%  -3.72%  (p=0.000 n=19+20)
GoParser         76.7ms ± 1%       73.5ms ± 1%  -4.15%  (p=0.000 n=18+19)
Reflect           172ms ± 1%        165ms ± 1%  -4.13%  (p=0.000 n=20+19)
Tar              63.1ms ± 1%       60.4ms ± 2%  -4.24%  (p=0.000 n=19+20)
XML              83.2ms ± 2%       79.2ms ± 2%  -4.79%  (p=0.000 n=20+20)
[Geo mean]        127ms             121ms       -4.73%

Change-Id: Ifbe35e48b3ea3b29430249b4667d2df8a2515aeb
Reviewed-on: https://go-review.googlesource.com/c/go/+/495596
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>

diff --git a/src/cmd/compile/default.pgo b/src/cmd/compile/default.pgo
new file mode 100644 (file)
index 0000000..2ba7968
Binary files /dev/null and b/src/cmd/compile/default.pgo differ

diff --git a/src/cmd/compile/profile.sh b/src/cmd/compile/profile.sh
new file mode 100644 (file)
index 0000000..37d65d8
--- /dev/null
+++ b/src/cmd/compile/profile.sh
@@ -0,0 +1,21 @@
+# Copyright 2023 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.
+
+# This script collects a CPU profile of the compiler
+# for building all targets in std and cmd, and puts
+# the profile at cmd/compile/default.pgo.
+
+dir=$(mktemp -d)
+cd $dir
+seed=$(date)
+
+for p in $(go list std cmd); do
+       h=$(echo $seed $p | md5sum | cut -d ' ' -f 1)
+       echo $p $h
+       go build -o /dev/null -gcflags=-cpuprofile=$PWD/prof.$h $p
+done
+
+go tool pprof -proto prof.* > $(go env GOROOT)/src/cmd/compile/default.pgo
+
+rm -r $dir

diff --git a/src/cmd/dist/buildtool.go b/src/cmd/dist/buildtool.go
index af3db5f59023bd9e99fdef6884cac49ca2c3de8d..7ecfb3a811b0e83896accae9bc51ddd832ff909b 100644 (file)
--- a/src/cmd/dist/buildtool.go
+++ b/src/cmd/dist/buildtool.go
@@ -100,6 +100,10 @@ var ignorePrefixes = []string{
 var ignoreSuffixes = []string{
        "_test.s",
        "_test.go",
+       // Skip PGO profile. No need to build toolchain1 compiler
+       // with PGO. And as it is not a text file the import path
+       // rewrite will break it.
+       ".pgo",
 }
 
 var tryDirs = []string{