https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119876
Hongtao Liu <liuhongt at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Known to work| |16.0
Resolution|--- |FIXED
Status|WAITING |RESOLVED
--- Comment #8 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #7)
> I think this is now fixed after r16-2298-g4d6c3f3b4fbf8c , can you confirm?
Yes, it's fixed, now it generates
test:
.LFB0:
.cfi_startproc
movl $2, %eax
vpxor %xmm5, %xmm5, %xmm5
vpbroadcastd %eax, %zmm4
vpternlogd $0xFF, %zmm5, %zmm5, %zmm5
xorl %eax, %eax
vpxor %xmm6, %xmm6, %xmm6
vmovdqa32 %zmm5, %zmm2
vmovdqa32 %zmm6, %zmm3
vmovdqa32 %zmm4, %zmm1
.p2align 6
.p2align 4
.p2align 3
.L2:
vmovdqa32 b(%rax), %zmm0
addq $64, %rax
vpcmpd $6, %zmm3, %zmm0, %k1
vmovdqa32 %zmm1, %zmm0
vpsrld $31, %zmm2, %zmm0{%k1}
vpaddd c-64(%rax), %zmm0, %zmm0
vmovdqa32 %zmm0, a-64(%rax)
cmpq $3968, %rax
jne .L2
I think it's better than clang, gcc uses vpsrld instead of vpblendmd to select
between 1 and 2.
