cmd/compile: eliminate more dead stores

[TocarIP]

What version of Go are you using (go version)?

Trunk
go version devel +3470321 Tue Apr 24 15:26:21 2018 -0500 linux/amd64

What did you do?

go build following program:

package test

//go:noinline
func f(x ...interface{}) {
}

func g() {
        var x string
        f(&x)
}

When looking into generated code I see that autotmp_1 is zeroed and than immediately overwritten

v27    00003 (8)       XORPS   X0, X0
 v7     00004 (8)       MOVUPS  X0, "".x-32(SP)
 v13    00005 (9)       MOVUPS  X0, ""..autotmp_1-16(SP) // zeroing
 v23    00006 (9)       LEAQ    type.*string(SB), AX
 v35    00007 (9)       MOVQ    AX, ""..autotmp_1-16(SP) // followed by 2  stores
 v34    00008 (9)       LEAQ    "".x-32(SP), AX
 v21    00009 (9)       MOVQ    AX, ""..autotmp_1-8(SP)
 v36    00010 (9)       LEAQ    ""..autotmp_1-16(SP), AX

dse pass already eliminates dead stores, including zeroing, so it should also handle this case.

[josharian]

cc @mvdan, who recently poked at the deadstore pass.

[mvdan]

@josharian thank you for the ping! Quite obviously, I'm not going to submit anything for DSE in 1.11 :) Will try to get something working during the freeze. Michael Munday lives really close, so I might bribe him with a beer to help me hack on DSE.

[josharian]

I poked at this briefly yesterday. I think (but am not sure) that there are two issues here.

1. Clearing a [1]T vs clearing a T. I think Michael’s recent struct CL might help. We should check.

2. DSE needs identical—like same v.ID—store dests. But decompose introduces new OffPtr Values for each decomposition. Doing CSE again before DSE would help but that is way too big a hammer. Unwrapping OffPtrs during DSE might do it, or having decompose somehow do in-flight CSE.

I leave this in your hands. :)

[gopherbot]

Change https://golang.org/cl/110121 mentions this issue: cmd/compile: unwrap OpOffPtr during DSE

[josharian]

cc @mundaym

[mundaym]

Just a quick note: while I was looking at reproducing #29892 I noticed that the LocalAddr and InlMark ops aren't specially handled by the dead store pass, so it will treat them as loads (preventing dead stores being eliminated if their mem arg is used as an input to one of the new ops). Possibly worth fixing if we revisit dead store elimination.