https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121591
Manuel López-Ibáñez <manu at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |manu at gcc dot gnu.org
--- Comment #4 from Manuel López-Ibáñez <manu at gcc dot gnu.org> ---
And even smaller example:
int dominates(const double * restrict a, const double * restrict b)
{
int a_le_b_0=(a[0] <= b[0]);
int a_eq_b_0=(a[0] == b[0]);
int a_le_b_1=(a[1] <= b[1]);
int a_eq_b_1=(a[1] == b[1]);
return a_le_b_0 & a_le_b_1 & !(a_eq_b_0 & a_eq_b_1);
}
x86-64 gcc (trunk) -O3 -march=x86-64-v4 -ffast-math
"dominates":
vmovsd xmm2, QWORD PTR [rdi]
vmovsd xmm3, QWORD PTR [rsi]
vmovsd xmm0, QWORD PTR [rdi+8]
vmovsd xmm1, QWORD PTR [rsi+8]
vcomisd xmm2, xmm3
sete al
vcomisd xmm0, xmm1
sete dl
and eax, edx
vcomisd xmm3, xmm2
setnb dl
vcomisd xmm1, xmm0
setnb cl
and edx, ecx
andn eax, eax, edx
and eax, 1
ret
x86-64 clang (trunk) -O3 -march=x86-64-v4 -ffast-math
dominates:
vmovsd xmm0, qword ptr [rdi]
vmovsd xmm1, qword ptr [rdi + 8]
vmovsd xmm2, qword ptr [rsi]
vmovsd xmm3, qword ptr [rsi + 8]
vcmplesd k0, xmm1, xmm3
vcmplesd k1, xmm0, xmm2
kandw k0, k1, k0
kmovd eax, k0
vcmpneqsd k0, xmm1, xmm3
vcmpneqsd k1, xmm0, xmm2
korw k0, k1, k0
kmovd ecx, k0
and cl, al
movzx eax, cl
and eax, 1
ret
x86-64 icc 19.0.1
dominates:
..B1.1: # Preds ..B1.0
xor r9d, r9d #7.47
movsd xmm3, QWORD PTR [rdi] #3.19
movsd xmm5, QWORD PTR [8+rdi] #5.19
movaps xmm0, xmm3 #7.47
movsd xmm2, QWORD PTR [rsi] #3.27
movaps xmm1, xmm5 #7.47
movsd xmm4, QWORD PTR [8+rsi] #5.27
cmpeqsd xmm3, xmm2 #7.47
cmpeqsd xmm5, xmm4 #7.47
cmplesd xmm0, xmm2 #7.47
cmplesd xmm1, xmm4 #7.47
movd ecx, xmm3 #7.47
movd r8d, xmm5 #7.47
movd eax, xmm0 #7.47
movd edx, xmm1 #7.47
neg ecx #7.47
neg r8d #7.47
test ecx, r8d #7.47
sete r9b #7.47
neg eax #7.47
neg edx #7.47
and eax, edx #7.23
and eax, r9d #7.47
ret #7.47
