https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85466
Drea Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|tree-optimization |target
--- Comment #24 from Drea Pinski <pinskia at gcc dot gnu.org> ---
.L30:
movaps xmm0, xmm2
cmpltss xmm0, DWORD PTR [rdx]
movaps xmm1, xmm0
andnps xmm0, xmm3
andps xmm1, xmm4
orps xmm0, xmm1
movss DWORD PTR [rax-4], xmm0
add rdx, 4
add rax, 4
cmp r12, rdx
jne .L30
vs:
.LBB3_19:
movss xmm1, dword ptr [r14 + rcx]
xor edx, edx
ucomiss xmm1, xmm0
seta dl
movss xmm1, dword ptr [rax + 4*rdx] // dl ? 0.8 : 1.0
movss dword ptr [r15 + rcx], xmm1
add rcx, 4
cmp rcx, 400000
jne .LBB3_19
Looks like this is only done for x86_64 without avx.
With avx you can use vblendvps instead.
.LBB3_19:
vcmpltss xmm3, xmm0, dword ptr [r14 + rax]
vblendvps xmm3, xmm2, xmm1, xmm3
vmovss dword ptr [r15 + rax], xmm3
add rax, 4
cmp rax, 400000
jne .LBB3_19
Seems like you could generate a V2SF memory location with the 2 constants and
load from that when you have a?CST1:CST2; if you don't have AVX.
