[dev.ssa] cmd/compile/internal/gc: wrong code generation

[brtzsnr]

package main
import "fmt"
func main() {
fmt.Println(f1_ssa())
}
func f1_ssa() int8 {
switch {} // prevent inlining
v4 := uint(21) // uint
v7 := v4 & v4 * v4 >> 2 - 3 + v4 << 1 ^ v4 >> 2 + v4 * v4 + v4 >> uint32(0) * 3 | v4 & v4 << 1 - 2 ^ v4 | v4 >> (3 + 1 << 2 * 2 ^ 3 >> uint32(0) * 3 << (uint32(0) + uint32(0)) - 1 & 2) // uint
v8 := uint32(0) >> ((v7 & v4 * v4) ^ v7 * v4 * v7)
v10 := 3 << v8 | 1 * 3 + 3 >> 2 * (3) - int8(3) - (int8(3) + int8(3) >> v4) // int8
return 3 & v10 + int8(3) >> v8 - int8(3) << 2 << 1 - v10 ^ int8(3) * int8(3) - 3 | int8(3) * v10 & int8(3) & v10 << 1 >> 12
}

go1.4: -29
master: -29
dev.ssa: -30

This is a bug in dev.ssa at tip. I'll try to minimize this later if I have time.

[randall77]

This appears to be the result of wrong flags processing. The scheduled code does:

v88 = CMPLconst <flags> [8] v48
v86 = SBBLcarrymask <int8> v88
... do things that clobber flags...
v87 = SBBLcarrymask <uint32> v88

v88 is not spilled and restored properly. There are 3 issues:

1. Regalloc doesn't know that the intervening instructions clobber flags.
2. We can't spill/restore flags.
3. The two SBBLs aren't CSEd because their types are different.

1 is about to be fixed with the new regalloc.
2 is to be done, but no timeline yet.
3 Josh was thinking about. Maybe remove types from (most?) values, they aren't used in many places.

With just 1 we'll at least get a "can't be regalloced" error instead of silently generating wrong code.

[randall77]

Closed via https://go-review.googlesource.com/#/c/13622/
1 is done, so we get "spill of flags not implemented yet" and fallback to the old compiler.