[Bug target/107093] AVX512 mask operations not simplified in fully masked loop

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107093

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Hongtao.liu from comment #4)
> change "*k, CBC" to "?k, CBC", in *mov{qi,hi,si,di}_internal.
> then RA works good to choose kxnor for setting constm1_rtx to mask register,
> and i got below with your attached patch(change #if 0 to #if 1), seems
> better than orginal patch.
> 
>  6foo:
>  7.LFB0:
>  8        .cfi_startproc
>  9        testl   %edi, %edi
> 10        jle     .L9
> 11        kxnorb  %k1, %k1, %k1
> 12        cmpl    $4, %edi
> 13        jl      .L11
> 14.L3:
> 15        vbroadcastsd    .LC2(%rip), %ymm3
> 16        vmovdqa .LC0(%rip), %xmm2
> 17        xorl    %eax, %eax
> 18        xorl    %ecx, %ecx
> 19        .p2align 4,,10
> 20        .p2align 3
> 21.L7:
> 22        vmovapd b(%rax), %ymm0{%k1}
> 23        addl    $4, %ecx
> 24        movl    %edi, %edx
> 25        vmulpd  %ymm3, %ymm0, %ymm1
> 26        subl    %ecx, %edx
> 27        cmpl    $4, %edx
> 28        vmovapd %ymm1, a(%rax){%k1}
> 29        vpbroadcastd    %edx, %xmm1
> 30        movl    $-1, %edx
> 31        vpcmpd  $1, %xmm1, %xmm2, %k1
> 32        kmovb   %k1, %esi
> 33        cmovge  %edx, %esi

not sure if the round-trip to GPRs for the sake
of a cmovge is worth, I guess a branch would be
better.