https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106952
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
With the proposed patches for PR88540 and PR105715 I get with -O3 -msse4.1
intersection:
.LFB2:
.cfi_startproc
movss .LC0(%rip), %xmm5
pxor %xmm2, %xmm2
movss (%rdi), %xmm4
movss 12(%rsi), %xmm1
movss 12(%rdi), %xmm0
divss %xmm2, %xmm5
movss (%rsi), %xmm3
subss %xmm4, %xmm1
subss %xmm4, %xmm3
pxor %xmm4, %xmm4
mulss %xmm0, %xmm1
mulss %xmm0, %xmm3
movaps %xmm1, %xmm0
cmpnltss %xmm2, %xmm0
blendvps %xmm0, %xmm1, %xmm4
movaps %xmm3, %xmm0
cmpnltss %xmm2, %xmm0
pxor %xmm2, %xmm2
blendvps %xmm0, %xmm3, %xmm2
movss 16(%rsi), %xmm0
minss %xmm5, %xmm3
minss %xmm5, %xmm1
movss 4(%rdi), %xmm5
minss %xmm4, %xmm2
movss 16(%rdi), %xmm4
subss %xmm5, %xmm0
maxss %xmm3, %xmm1
movss 4(%rsi), %xmm3
subss %xmm5, %xmm3
mulss %xmm4, %xmm0
movss 8(%rdi), %xmm5
mulss %xmm4, %xmm3
movaps %xmm2, %xmm4
maxss %xmm0, %xmm4
minss %xmm1, %xmm0
maxss %xmm3, %xmm2
minss %xmm1, %xmm3
movss 8(%rsi), %xmm1
subss %xmm5, %xmm1
maxss %xmm3, %xmm0
movss 20(%rsi), %xmm3
minss %xmm4, %xmm2
movss 20(%rdi), %xmm4
subss %xmm5, %xmm3
mulss %xmm4, %xmm1
movaps %xmm2, %xmm5
mulss %xmm4, %xmm3
movaps %xmm2, %xmm4
maxss %xmm1, %xmm4
minss %xmm0, %xmm1
movaps %xmm3, %xmm2
maxss %xmm3, %xmm5
minss %xmm0, %xmm2
minss %xmm5, %xmm4
maxss %xmm1, %xmm2
comiss %xmm4, %xmm2
seta %al
ret
there's the existing issue that RTL conditional move expansion doesn't
preserve the equality of constants for
_33 = t2_34 < 0.0;
_12 = _33 ? 0.0 : t2_34;
but it emits two loads from the constant pool for 0.0 here which in the x86
backend fail to be recognized as min/max.
