cmd/compile: unnecessary slice literal zeroing

[icholy]

Keyless slice literals zero the backing array and then immediately overwrite it. https://godbolt.org/z/5bx9TPo7n

What operating system and processor architecture are you using?

linux amd64

What did you do?

package main

func main() {
    a := []int{1, 2, 3, 4}
    println(a)
}

What did you expect to see?

movq    $1, ""..autotmp_2+24(SP)
movq    $2, ""..autotmp_2+32(SP)
movq    $3, ""..autotmp_2+40(SP)
movq    $4, ""..autotmp_2+48(SP)

What did you see instead?

leaq    ""..autotmp_2+24(SP), AX
xorps   X15, X15
movups  X15, (AX)
leaq    ""..autotmp_2+40(SP), CX
xorps   X15, X15
movups  X15, (CX)
movq    $1, ""..autotmp_2+24(SP)
movq    $2, ""..autotmp_2+32(SP)
movq    $3, ""..autotmp_2+40(SP)
movq    $4, ""..autotmp_2+48(SP)

[randall77]

Yes, the deadstore pass needs to be upgraded to handle cases like this. It's been on the back burner for a while.

I think there are 2 things that are needed to make this work:

1. Instead of just recording the size of shadowed memory at each pointer, record an offset+size. Then we can fold OffPtr arguments into the analysis easily.
2. Handle LocalAddr ops somehow - these are currently treated as reads and mess up the analysis. The code conservatively assumes taking the address makes the zeroed temp visible, when in fact that address is only used for storing.

I think 1 is easy. 2 might be a bit harder.

[icholy]

My initial though was to handle this at the ir/walk stages. Once in SSA form, the deadstore pass would need to re-discover the fact that the zeroed memory is shadowed.

[networkimprov]

cc @josharian

[gopherbot]

Change https://go.dev/cl/538595 mentions this issue: cmd/compile: handle constant pointer offsets in dead store elimination