cmd/compile: internal compiler error: "schedule does not include all values in block" on linux-s390x-ibm

[bcmills]

See also #37246.

https://build.golang.org/log/37189dee0a49af1e6e538c7e05d0e6a402f5a210

# golang.org/x/perf/internal/stats
internal/stats/utest.go:127:13: internal compiler error: 'MannWhitneyUTest': schedule does not include all values in block b80

goroutine 21 [running]:
runtime/debug.Stack(0x9c51a0, 0xc00000e018, 0x0)
	/data/golang/workdir/go/src/runtime/debug/stack.go:24 +0x98
cmd/compile/internal/gc.Fatalf(0xc000b96080, 0x36, 0xc000a7dba0, 0x2, 0x2)
	/data/golang/workdir/go/src/cmd/compile/internal/gc/subr.go:193 +0x170
cmd/compile/internal/gc.(*ssafn).Fatalf(0xc000a667e0, 0xd0007f0d1, 0x870026, 0x30, 0xc000a5b3e0, 0x1, 0x1)
	/data/golang/workdir/go/src/cmd/compile/internal/gc/ssa.go:6857 +0x184
cmd/compile/internal/ssa.(*Func).Fatalf(...)
	/data/golang/workdir/go/src/cmd/compile/internal/ssa/func.go:625
cmd/compile/internal/ssa.schedule(0xc0004d8420)
	/data/golang/workdir/go/src/cmd/compile/internal/ssa/schedule.go:293 +0x1194
cmd/compile/internal/ssa.Compile(0xc0004d8420)
	/data/golang/workdir/go/src/cmd/compile/internal/ssa/compile.go:93 +0x946
cmd/compile/internal/gc.buildssa(0xc0000d8420, 0x3, 0x0)
	/data/golang/workdir/go/src/cmd/compile/internal/gc/ssa.go:460 +0x93c
cmd/compile/internal/gc.compileSSA(0xc0000d8420, 0x3)
	/data/golang/workdir/go/src/cmd/compile/internal/gc/pgen.go:296 +0x40
cmd/compile/internal/gc.compileFunctions.func2(0xc000a83140, 0xc000498240, 0x3)
	/data/golang/workdir/go/src/cmd/compile/internal/gc/pgen.go:361 +0x40
created by cmd/compile/internal/gc.compileFunctions
	/data/golang/workdir/go/src/cmd/compile/internal/gc/pgen.go:359 +0x116

CC @mundaym @randall77

[mundaym]

Steps to reproduce (I used the compiler at revision 83bfe3b and perf at revision 9c9101d):

git clone https://go.googlesource.com/perf
cd perf/internal/stats
GOOS=linux GOARCH=s390x go test -c

Output:

# golang.org/x/perf/internal/stats [golang.org/x/perf/internal/stats.test]
./utest.go:127:13: internal compiler error: 'MannWhitneyUTest': schedule does not include all values in block b80

goroutine 29 [running]:
runtime/debug.Stack(0xd6ff40, 0xc00000e018, 0x0)
	/home/michael/Development/go/master/src/runtime/debug/stack.go:24 +0x9d
cmd/compile/internal/gc.Fatalf(0xc0003f9d40, 0x36, 0xc000a98380, 0x2, 0x2)
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/subr.go:193 +0x1aa
cmd/compile/internal/gc.(*ssafn).Fatalf(0xc0009cf680, 0x7f0d10000000d, 0xc1c4a1, 0x30, 0xc0003f6d60, 0x1, 0x1)
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/ssa.go:6843 +0x19b
cmd/compile/internal/ssa.(*Func).Fatalf(...)
	/home/michael/Development/go/master/src/cmd/compile/internal/ssa/func.go:625
cmd/compile/internal/ssa.schedule(0xc000b634a0)
	/home/michael/Development/go/master/src/cmd/compile/internal/ssa/schedule.go:293 +0x1abd
cmd/compile/internal/ssa.Compile(0xc000b634a0)
	/home/michael/Development/go/master/src/cmd/compile/internal/ssa/compile.go:93 +0x9bb
cmd/compile/internal/gc.buildssa(0xc0000e89a0, 0x3, 0x0)
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/ssa.go:460 +0xcca
cmd/compile/internal/gc.compileSSA(0xc0000e89a0, 0x3)
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/pgen.go:296 +0x5d
cmd/compile/internal/gc.compileFunctions.func2(0xc000b97260, 0xc000017da0, 0x3)
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/pgen.go:361 +0x49
created by cmd/compile/internal/gc.compileFunctions
	/home/michael/Development/go/master/src/cmd/compile/internal/gc/pgen.go:359 +0x128

[mundaym]

I printed a bit more information. The scheduler seems to assume the Select0 and Select1 ops are always in the same block as their argument, which is probably necessary for register allocation, however that invariant appears to be broken in this case:

schedule does not include all values in block b80
  value not found: v724 = Select1 <flags> v616 (block b80)
    argument 0: v616 = FSUB <float64,flags> v282 v403 (block b70)

Floating point ops generating flags was added in CL 209160. I think the oversight in that CL is that the new rules do not force the Select1 to be in the same block as the original value:

(LTDBR (Select0 x:(F(ADD|SUB) _ _)))   -> (Select1 x)
(LTEBR (Select0 x:(F(ADDS|SUBS) _ _))) -> (Select1 x)

The solution is probably to replace these rules with:

(LTDBR (Select0 x:(F(ADD|SUB) _ _)))   && b == x.Block -> (Select1 x)
(LTEBR (Select0 x:(F(ADDS|SUBS) _ _))) && b == x.Block -> (Select1 x)

Alternatively we could use @x.Block, but we probably don't want to move the block that the flags are generated in for this optimization.

cc @ruixin-bao

[gopherbot]

Change https://golang.org/cl/227879 mentions this issue: cmd/compile: fix incorrect block for s390x Select1 op

[ruixin-bao]

Thank you for looking into this issue.

Steps to reproduce (I used the compiler at revision 83bfe3b and perf at revision 9c9101d):

I am able to reproduce it as well.

The solution is probably to replace these rules with:
(LTDBR (Select0 x:(F(ADD|SUB) _ _))) && b == x.Block -> (Select1 x)
(LTEBR (Select0 x:(F(ADDS|SUBS) _ _))) && b == x.Block -> (Select1 x)

It looks like so. FWIW, I have tested the CL you sent and can confirm it fixes the problem.

[gopherbot]

Change https://golang.org/cl/237118 mentions this issue: cmd/compile: always tighten and de-duplicate tuple selectors