cmd/compile: Pow10 freeze the compiler on certain condition on Go 1.24

[etrnal70]

Go version

go version go1.24.0 darwin/arm64

Output of go env in your module/workspace:

AR='ar'
CC='clang'
CGO_CFLAGS='-O2 -g'
CGO_CPPFLAGS=''
CGO_CXXFLAGS='-O2 -g'
CGO_ENABLED='1'
CGO_FFLAGS='-O2 -g'
CGO_LDFLAGS='-O2 -g'
CXX='clang++'
GCCGO='gccgo'
GO111MODULE=''
GOARCH='arm64'
GOARM64='v8.0'
GOAUTH='netrc'
GOBIN=''
GOCACHE='/Users/hanifrmn/Library/Caches/go-build'
GOCACHEPROG=''
GODEBUG=''
GOENV='/Users/hanifrmn/Library/Application Support/go/env'
GOEXE=''
GOEXPERIMENT=''
GOFIPS140='off'
GOFLAGS=''
GOGCCFLAGS='-fPIC -arch arm64 -pthread -fno-caret-diagnostics -Qunused-arguments -fmessage-length=0 -ffile-prefix-map=/var/folders/xd/37pvgts12bg_bk2wx9kp9gbm0000gp/T/go-build2186527099=/tmp/go-build -gno-record-gcc-switches -fno-common'
GOHOSTARCH='arm64'
GOHOSTOS='darwin'
GOINSECURE=''
GOMOD='/Users/hanifrmn/Repo/playground/go.mod'
GOMODCACHE='/Users/hanifrmn/go/pkg/mod'
GONOPROXY=''
GONOSUMDB=''
GOOS='darwin'
GOPATH='/Users/hanifrmn/go'
GOPRIVATE=''
GOPROXY='https://proxy.golang.org,direct'
GOROOT='/Users/hanifrmn/sdk/go1.24.0'
GOSUMDB='sum.golang.org'
GOTELEMETRY='local'
GOTELEMETRYDIR='/Users/hanifrmn/Library/Application Support/go/telemetry'
GOTMPDIR=''
GOTOOLCHAIN='auto'
GOTOOLDIR='/Users/hanifrmn/sdk/go1.24.0/pkg/tool/darwin_arm64'
GOVCS=''
GOVERSION='go1.24.0'
GOWORK=''
PKG_CONFIG='pkg-config'

What did you do?

package main

import (
	"math"
)

func main() {
	test(2)
}

func test(i int) {
	if i <= 0 {
		return
	}

	_ = math.Pow10(i + 2)
}

running go build will freeze the compiler

What did you see happen?

Running go build will freeze the compiler

Running with -gcflags=-l rans fine

What did you expect to see?

It compiles correctly on Go 1.23.6

[gabyhelp]

Related Issues

• cmd/compile: miscompilation of simple arithmetic program targeting darwin/arm64 #62723 (closed)
• cmd/compile: rerwrite cycle detection logic taking too long for large functions #66773 (closed)
• Unable to calculate math.Mod for numbers larger than 1e+16 #36456 (closed)
• math: Pow(negativeZero, number_that_overflows_int64) produces incorrect result on arm64 #57465 (closed)
• cmd/compile: incorrect math reduction #70481 (closed)
• cmd/compile: bad code generation for ARM (bounds test fails) #41780 (closed)
• Test #67906 (closed)
• cmd/compile: internal compiler error: panic during prove while compiling: unexpected induction with too many parents #63955 (closed)
• cmd/compile: generic code compile never end #50204
• compiler: infinite loop and closure assertion failed #69062 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[Jorropo]

Well I am very confused:

git bisect start
# status: waiting for both good and bad commits
# bad: [4700bc41750c1a80c21b00baa0a93a64eb79e7ac] test: trying to reproduce 71852
git bisect bad 4700bc41750c1a80c21b00baa0a93a64eb79e7ac
# status: waiting for good commit(s), bad commit known
# good: [a991f9c34d454d3d844f21dc08f2d05df35a8c60] [release-branch.go1.23] go1.23.6
git bisect good a991f9c34d454d3d844f21dc08f2d05df35a8c60
# good: [5e8a7316658c2f300a375041b6e0a606fec4c5f2] README: fix CC BY license name
git bisect good 5e8a7316658c2f300a375041b6e0a606fec4c5f2
# bad: [2fd2718f6b64000fa25e0c5d1ee48aa1426d5a6f] bufio: add example for ReadFrom and remove unused code
git bisect bad 2fd2718f6b64000fa25e0c5d1ee48aa1426d5a6f
# bad: [123594d3863b0a4b9094a569957d1bd94ebe7512] runtime: remove cloudwego/frugal unused linkname from comment
git bisect bad 123594d3863b0a4b9094a569957d1bd94ebe7512
# good: [206df8e7ad8a736b0dcbf5d98ab237b6784e5007] net/http: rename server receiver for consistency
git bisect good 206df8e7ad8a736b0dcbf5d98ab237b6784e5007
# good: [e126129d7612349874828685c2bcd49de498a1a0] cmd/compile/internal/ssa: combine shift and addition for riscv64 rva22u64
git bisect good e126129d7612349874828685c2bcd49de498a1a0
# bad: [a00195d304e6858406c6c9c961d253eeb8cb0aec] all: use t.Chdir in tests
git bisect bad a00195d304e6858406c6c9c961d253eeb8cb0aec
# good: [9b6efc25cddad4bebfa64fd8543afd07fbb0ab3e] go/ast: remove unused code
git bisect good 9b6efc25cddad4bebfa64fd8543afd07fbb0ab3e
# good: [c3f346a485f2fa97a7bdee82d587419b3823a1ba] math,os,os/*: use testenv.Executable
git bisect good c3f346a485f2fa97a7bdee82d587419b3823a1ba
# good: [d91a2e5b11cc1badd7c760a911218a64d91ac8b0] cmd: replace many sort.Interface with slices.Sort and SortFunc
git bisect good d91a2e5b11cc1badd7c760a911218a64d91ac8b0
# bad: [820f58a27f7f64f21353db6a071bf5dbf658924c] cmd/compile: compute Negation's limits from argument's limits
git bisect bad 820f58a27f7f64f21353db6a071bf5dbf658924c
# good: [4f2c0e5d0806a2f6fbe2d0704683d2b71d8191be] cmd/compile: compute Trunc's limits from argument's limits
git bisect good 4f2c0e5d0806a2f6fbe2d0704683d2b71d8191be
# good: [2f3165973f9bf1e1a95d108c57e5257f4655ee59] cmd/compile: compute Complement's limits from argument's limits
git bisect good 2f3165973f9bf1e1a95d108c57e5257f4655ee59
# first bad commit: [820f58a27f7f64f21353db6a071bf5dbf658924c] cmd/compile: compute Negation's limits from argument's limits

820f58a is very simple, copying patterns and code from other patches.

I don't know when I'll be able to spend time investigating this,
I'll revert that for 1.24.1 but I think there is a latent bug hiding somewhere.

[gopherbot]

Change https://go.dev/cl/650975 mentions this issue: Revert "cmd/compile: compute Negation's limits from argument's limits"

[Jorropo]

cc @golang/compiler

[Jorropo]

@gopherbot Please open backport issue for 1.24

This is a regression where non esoteric code that used to build now does not.
https://gophers.slack.com/archives/C6CM8NUEB/p1740060033459149

[gopherbot]

Backport issue(s) opened: #71855 (for 1.24).

Remember to create the cherry-pick CL(s) as soon as the patch is submitted to master, according to https://go.dev/wiki/MinorReleases.

[Jorropo]

Thanks to @etrnal70's reproducer I was able to preliminarily narrow down the issue to a bug in newLimit.
There is a loop permanently doing this operation in the flowLimit loop:

flowLimit: v51 = Rsh64Ux64 <uint> [false] v49 v46: sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 -> sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 false
flowLimit: v49 = Neg64 <int> v12: sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 -> sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 true
flowLimit: v51 = Rsh64Ux64 <uint> [false] v49 v46: sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 -> sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 false
flowLimit: v49 = Neg64 <int> v12: sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 -> sm,SM,um,UM=-9223372036854775808,9223372036854775807,0,18446744073709551615 true

Altho you might note that the limits do not change YET newLimit returns true which cause the loop to never break since it thinks the values are being forever narrowed down when they are not.

[gopherbot]

Change https://go.dev/cl/650580 mentions this issue: cmd/compile: stop flowLimit search when we proved unsat

[Jorropo]

I've submitted two different proper fixes:

• https://go-review.googlesource.com/c/go/+/650975
• https://go-review.googlesource.com/c/go/+/650580

My original plan to revert 820f58a would not solve all regression cases, just this particular one.
I think the only correct revert would involve reverting most of @randall77's prove constants handling improvements, so I think we should consider backporting the fix altho this is more risky.