cmd/compile: arm64 ssa check compilation failure

[josharian]

Reproduce:

• Patch in CL 216418
• GOOS=linux GOARCH=arm64 go build -gcflags=-d=ssa/check/seed=21431 cmd/internal/obj/s390x

Result:

# cmd/internal/obj/s390x
cmd/internal/obj/s390x/asmz.go:2709:7: internal compiler error: '(*ctxtz).asmout': bad int32 AuxInt value for v8033

goroutine 1 [running]:
runtime/debug.Stack(0x1c3bda0, 0xc0000ac008, 0x0)
	/Users/josh/go/tip/src/runtime/debug/stack.go:24 +0x9d
cmd/compile/internal/gc.Fatalf(0xc000d211a0, 0x23, 0xc0007132a0, 0x2, 0x2)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/subr.go:193 +0x291
cmd/compile/internal/gc.(*ssafn).Fatalf(0xc0004005a0, 0xa9507100000005, 0x1a874ef, 0x1d, 0xc00081d620, 0x1, 0x1)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/ssa.go:6820 +0x1b0
cmd/compile/internal/ssa.(*Func).Fatalf(...)
	/Users/josh/go/tip/src/cmd/compile/internal/ssa/func.go:625
cmd/compile/internal/ssa.checkFunc(0xc002281600)
	/Users/josh/go/tip/src/cmd/compile/internal/ssa/check.go:141 +0x1ece
cmd/compile/internal/ssa.Compile(0xc002281600)
	/Users/josh/go/tip/src/cmd/compile/internal/ssa/compile.go:135 +0xa60
cmd/compile/internal/gc.buildssa(0xc0000ffe40, 0x0, 0x0)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/ssa.go:444 +0xcd8
cmd/compile/internal/gc.compileSSA(0xc0000ffe40, 0x0)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/pgen.go:298 +0x5d
cmd/compile/internal/gc.compile(0xc0000ffe40)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/pgen.go:277 +0x33b
cmd/compile/internal/gc.funccompile(0xc0000ffe40)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/pgen.go:222 +0xc1
cmd/compile/internal/gc.Main(0x1a9b3b8)
	/Users/josh/go/tip/src/cmd/compile/internal/gc/main.go:714 +0x3299
main.main()
	/Users/josh/go/tip/src/cmd/compile/main.go:52 +0xac

cc @cherrymui

[toothrot]

also cc'ing @randall77 @griesemer

Is this concerning enough to be a release blocker? linux/arm64 is a first-class port. If not, I'll change the milestone.

[josharian]

@toothrot hard to say for sure without diagnosing a bit. Let's wait to see what @cherrymui has to say.

[cherrymui]

It's not very concerning. It is just that the Value's AuxInt doesn't fit in int32 while the AuxInt type is specified as int32. I think the high bits are never used and so don't actually matter. So it shouldn't affect anything when SSA check is not on.

[randall77]

I can't reproduce now.
(CL 216418 is in, so just trying at tip.)
Was this fixed? By accident, or on purpose?

[josharian]

It wouldn’t be surprising if the seed required to reproduce changed. I found this by running in a loop with randomly generated seeds. I’ll try that again at some point if no one beats me to it. (It’d still be better to have a builder doing this, not me: #36756)

[josharian]

Poking through zsh history, looks like I was testing with this:

for i in `seq 10000`; do rnd=$RANDOM; echo seed=$rnd; echo js/wasm; GOOS=js GOARCH=wasm go build -a -gcflags=all=-d=ssa/check/on=$rnd std cmd; for arch in 386 amd64 arm arm64 mips mips64 mips64le mipsle ppc64 ppc64le s390x; do echo linux/$arch; GOOS=linux GOARCH=$arch go build -a -gcflags=all=-d=ssa/check/seed=$rnd std cmd; done; done

Running now.

[josharian]

Looks fixed, I’d guess by accident as part of strongly typing arm64 rewrite rules.