[X86] Worse code generation from patch 24780e13e5be1501e34330148137a10fa9965166

[xortator]

We see worse code generation on memcmp-like function caused by

commit 24780e13e5be1501e34330148137a10fa9965166 (HEAD -> main)
Author: Simon Pilgrim <llvm-dev@redking.me.uk>
Date:   Sat Apr 1 15:38:38 2023 +0100

    [X86] MatchVectorAllEqualTest - add support for icmp(reduce_and(X),-1) allof reduction patterns

    Also, improve codegen in LowerVectorAllEqual for X == -1 cases to reduce over sized vector using a AND reduction

Asm on last release: https://godbolt.org/z/Ebcd7P3jP
Current asm: https://godbolt.org/z/nxv1vc16e

Loop block changed from

.LBB0_2:                                # %memcmp.loop
        vmovdqu (%rsi,%rcx), %xmm0
        vmovdqu 16(%rsi,%rcx), %xmm1
        vpsubb  16(%rdi,%rcx), %xmm1, %xmm1
        vpsubb  (%rdi,%rcx), %xmm0, %xmm0
        vpor    %xmm1, %xmm0, %xmm0
        vptest  %xmm0, %xmm0
        jne     .LBB0_4
        addq    $32, %rcx
        cmpq    %rax, %rcx
        jb      .LBB0_2

to

.LBB0_2:                                # %memcmp.loop
        vmovdqu (%rsi,%rcx), %xmm0
        vmovdqu 16(%rsi,%rcx), %xmm1
        vpcmpeqb        (%rdi,%rcx), %xmm0, %xmm0
        vpmovmskb       %xmm0, %edx
        vpcmpeqb        16(%rdi,%rcx), %xmm1, %xmm0
        vpmovmskb       %xmm0, %r8d
        shll    $16, %r8d
        orl     %edx, %r8d
        cmpl    $-1, %r8d
        jne     .LBB0_4
        addq    $32, %rcx
        cmpq    %rax, %rcx
        jb      .LBB0_2

I am not sure into how big performance regression this translates (build & benchmark runs are underway), but the assembly definitely looks worse in the loop block.

[xortator]

@RKSimon can you please take a look?

[xortator]

Reduced the test (links updated).

[RKSimon]

This will be fixed by https://reviews.llvm.org/D147452

[llvmbot]

@llvm/issue-subscribers-backend-x86