https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048
--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
The recent patch improved this to avoid some of the compares. We still have
the three-argument PHI and thus three VEC_CONDs.
.L10:
vmovups (%rdi,%rdx), %ymm0
vcmpltps %ymm6, %ymm0, %ymm3
vcmpltps %ymm2, %ymm0, %ymm1
vpandn %ymm1, %ymm3, %ymm1
vblendvps %ymm1, %ymm5, %ymm4, %ymm1
vblendvps %ymm3, %ymm7, %ymm1, %ymm1
vaddps %ymm1, %ymm0, %ymm0
vaddps (%rax,%rdx), %ymm0, %ymm0
vmovups %ymm0, (%rax,%rdx)
addq $32, %rdx
cmpq $1024, %rdx
jne .L10
vs. GCC 12
.L6:
vmovups (%rdi,%rdx), %ymm1
vcmpltps %ymm5, %ymm1, %ymm0
vcmpltps %ymm6, %ymm1, %ymm4
vblendvps %ymm0, %ymm3, %ymm2, %ymm0
vandps %ymm3, %ymm4, %ymm4
vaddps %ymm4, %ymm0, %ymm0
vaddps %ymm1, %ymm0, %ymm0
vaddps (%rax,%rdx), %ymm0, %ymm0
vmovups %ymm0, (%rax,%rdx)
addq $32, %rdx
cmpq $1024, %rdx
jne .L6
which at least overall looks comparable.
