https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126789
--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'm testing a patch for the V2SImode case which we'll code-generate as the
following (with -fno-vect-cost-model):
foo:
.LFB0:
.cfi_startproc
vmovupd (%rsi), %xmm2
vmovddup %xmm0, %xmm1
vmovd %edi, %xmm3
movabsq $4294967297, %rax
vpshufd $0xe0, %xmm3, %xmm0
vcmppd $1, %xmm1, %xmm2, %k1
vmovq (%rdx), %xmm1
vpcmpgtd %xmm1, %xmm0, %xmm0
vmovdqa32 %xmm0, %xmm1{%k1}{z} (A)
vmovq %rax, %xmm0
vpand %xmm1, %xmm0, %xmm0
vmovd %xmm0, %eax
vpextrd $1, %xmm0, %edx
addl %edx, %eax
ret
it's nice that we are able to combine the %kN -> SSE mask representation
with the bitwise and using (A). The V4SImode variant manages the
-1 -> 1 transition with
vpsrld $31, %xmm1, %xmm0{%k1}{z}
not sure why we chose
movabsq $4294967297, %rax
vmovq %rax, %xmm0
vpand %xmm1, %xmm0, %xmm0
here over
vpsrld $31, %xmm0, %xmm0
it's both from a .VCOND_MASK (mask, {1, 1}, {0, 0}) expansion. With
AVX2 we'll get earlier folding to & { 1, 1, 1, 1 } though.
