https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117232
--- Comment #4 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #0)
> This is expansion of PR 113609 which showed when I improved phiopt's factor
> operations to handle more than just 1 operand operations.
>
> New reduced testcase that fails to use kortestw (even without my phiopt
> improvements):
> ```
> #include <immintrin.h>
>
> int
> cmp_vector_je_mask64_t(__m512i a, __m512i b, int c, int d) {
> __mmask64 k = _mm512_cmpeq_epi8_mask (a, b);
> return k == (__mmask64) -1 ? c : d;
> }
> ```
>
> GCC currently produces:
> ```
> vpcmpb $0, %zmm1, %zmm0, %k0
> kmovq %k0, %rax
> cmpq $-1, %rax
> movl %edi, %eax
> cmovne %esi, %eax
> ret
> ```
>
> While LLVM produces:
> ```
> movl %edi, %eax
> vpcmpneqd %zmm1, %zmm0, %k0
> kortestw %k0, %k0
> cmovnel %esi, %eax
> vzeroupper
> retq
> ```
Now GCC generates
.cfi_startproc
vpcmpb $0, %zmm1, %zmm0, %k0
movl %edi, %eax
kortestq %k0, %k0
cmovnc %esi, %eax
ret
.cfi_endproc
