Unnecessary conversion of `(a==v)&(b==v)` into `(a!=v)|(b!=v)` results in more assembly

[dzaima]

The code:

#include <immintrin.h>
_Bool both_equal(__m256i a, __m256i b, __m256i v) {
  __m256i a1 = _mm256_cmpeq_epi8(v, a);
  __m256i b1 = _mm256_cmpeq_epi8(v, b);
  __m256i all = _mm256_and_si256(a1, b1);
  return _mm256_movemask_epi8(all) == 0xFFFFFFFF;
}

compiled with -O3 -march=haswell on clang 19.1 or trunk gives:

both_equal:
        vpcmpeqb        ymm0, ymm2, ymm0
        vpcmpeqd        ymm3, ymm3, ymm3
        vpxor   ymm0, ymm0, ymm3
        vpcmpeqb        ymm1, ymm2, ymm1
        vpxor   ymm1, ymm1, ymm3
        vpor    ymm0, ymm0, ymm1
        vpmovmskb       eax, ymm0
        test    eax, eax
        sete    al
        vzeroupper
        ret

whereas 18.1 produced the better:

both_equal:
        vpxor   ymm0, ymm0, ymm2
        vpxor   ymm1, ymm1, ymm2
        vpor    ymm0, ymm0, ymm1
        vptest  ymm0, ymm0
        sete    al
        vzeroupper
        ret

https://godbolt.org/z/coK5qqErr (variant with 4 tests instead of 2: https://godbolt.org/z/96eEK439G)

[RKSimon]

18.1:

define i1 @both_equal(<4 x i64> %a, <4 x i64> %b, <4 x i64> %v) {
entry:
  %0 = bitcast <4 x i64> %v to <32 x i8>
  %1 = bitcast <4 x i64> %a to <32 x i8>
  %cmp.i5 = icmp eq <32 x i8> %0, %1
  %2 = bitcast <4 x i64> %b to <32 x i8>
  %cmp.i = icmp eq <32 x i8> %0, %2
  %and.i78 = and <32 x i1> %cmp.i5, %cmp.i
  %3 = bitcast <32 x i1> %and.i78 to i32
  %cmp = icmp eq i32 %3, -1
  ret i1 %cmp
}

Trunk:

define i1 @both_equal(<4 x i64> %a, <4 x i64> %b, <4 x i64> %v) {
entry:
  %0 = bitcast <4 x i64> %v to <32 x i8>
  %1 = bitcast <4 x i64> %a to <32 x i8>
  %2 = bitcast <4 x i64> %b to <32 x i8>
  %3 = icmp ne <32 x i8> %0, %1
  %4 = icmp ne <32 x i8> %0, %2
  %5 = or <32 x i1> %3, %4
  %6 = bitcast <32 x i1> %5 to i32
  %cmp = icmp eq i32 %6, 0
  ret i1 %cmp
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: dzaima (dzaima)

The code: ```c #include <immintrin.h> _Bool both_equal(__m256i a, __m256i b, __m256i v) { __m256i a1 = _mm256_cmpeq_epi8(v, a); __m256i b1 = _mm256_cmpeq_epi8(v, b); __m256i all = _mm256_and_si256(a1, b1); return _mm256_movemask_epi8(all) == 0xFFFFFFFF; } ``` compiled with `-O3 -march=haswell` on clang 19.1 or trunk gives: ```asm both_equal: vpcmpeqb ymm0, ymm2, ymm0 vpcmpeqd ymm3, ymm3, ymm3 vpxor ymm0, ymm0, ymm3 vpcmpeqb ymm1, ymm2, ymm1 vpxor ymm1, ymm1, ymm3 vpor ymm0, ymm0, ymm1 vpmovmskb eax, ymm0 test eax, eax sete al vzeroupper ret ``` whereas 18.1 produced the better: ```asm both_equal: vpxor ymm0, ymm0, ymm2 vpxor ymm1, ymm1, ymm2 vpor ymm0, ymm0, ymm1 vptest ymm0, ymm0 sete al vzeroupper ret ``` https://godbolt.org/z/coK5qqErr (variant with 4 tests instead of 2: https://godbolt.org/z/96eEK439G)