[AArch64] Peephole pass incorrectly removes comparison

[nikic]

Bugzilla Link	51476
Resolution	FIXED
Resolved on	Oct 11, 2021 20:29
Version	trunk
OS	Linux
Blocks	#51489
CC	@Arnaud-de-Grandmaison-ARM,@DMG862,@smithp35
Fixed by commit(s)	81b1065 b643ee1

Extended Description

target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-unknown-linux-gnu"

define void @​test(i8 %arg) {
%tmp = alloca i8
%cmp1 = icmp ne i8 %arg, 1
%zext = zext i1 %cmp1 to i8
store i8 %zext, i8* %tmp
%zext2 = load i8, i8* %tmp
%cmp2 = icmp eq i8 %zext2, 3
br i1 %cmp2, label %exit, label %do_call

do_call:
call void @​unknown(i8 %zext2)
ret void

exit:
ret void
}

declare void @​unknown(i8)

Running through llc produces:

// %bb.0:
str x30, [sp, #-16]! // 8-byte Folded Spill
.cfi_def_cfa_offset 16
.cfi_offset w30, -16
and w8, w0, #​0xff
cmp w8, #​1
cset w0, ne
strb w0, [sp, #​12]
b.ne .LBB0_2
// %bb.1: // %do_call
bl unknown
.LBB0_2: // %common.ret
ldr x30, [sp], #​16 // 8-byte Folded Reload
ret

Note how the final comparison to #​3 has now reuses the result of the comparison to #​1.

This is caused by the new optimization introduced in https://reviews.llvm.org/D98564. From a cursory reading, this optimization was written under the assumption that CmpValue==1 means that the comparison is against #​1, while analyzeCompare() canonicalizes the comparison such that CmpValue==1 indicates a comparison against any non-zero value.

[nikic]

Candidate patch: https://reviews.llvm.org/D108076

[nikic]

Fixed by 81b1065. Keeping this open to track release/13.x backport.

[tstellar]

Merged: b643ee1

[tstellar]

mentioned in issue #51489