https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79725
Bug ID: 79725
Summary: Sinking opportunity missed if complex type is changed
Product: gcc
Version: 7.0.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: drraph at gmail dot com
Target Milestone: ---
Consider:
#include <complex.h>
complex double f(complex double x[]) {
complex float p = 1.0;
for (int i = 0; i < 1000000; i++)
p = x[i];
return p;
}
This compiles using -O3 -march=core-avx2 -ffast-math to:
f:
lea rax, [rdi+16000000]
.L2:
vmovupd ymm0, YMMWORD PTR [rdi]
vmovupd ymm5, YMMWORD PTR [rdi+32]
sub rdi, -128
vmovupd ymm1, YMMWORD PTR [rdi-64]
vmovupd ymm4, YMMWORD PTR [rdi-32]
vunpcklpd ymm2, ymm0, ymm5
vunpckhpd ymm0, ymm0, ymm5
vunpcklpd ymm3, ymm1, ymm4
vunpckhpd ymm1, ymm1, ymm4
vpermpd ymm2, ymm2, 216
vpermpd ymm3, ymm3, 216
vpermpd ymm0, ymm0, 216
vpermpd ymm1, ymm1, 216
vcvtpd2ps xmm2, ymm2
vcvtpd2ps xmm3, ymm3
vcvtpd2ps xmm0, ymm0
vcvtpd2ps xmm1, ymm1
vinsertf128 ymm2, ymm2, xmm3, 0x1
vinsertf128 ymm1, ymm0, xmm1, 0x1
cmp rax, rdi
jne .L2
vextractf128 xmm2, ymm2, 0x1
vextractf128 xmm1, ymm1, 0x1
vshufps xmm0, xmm2, xmm2, 255
vshufps xmm1, xmm1, xmm1, 255
vcvtss2sd xmm0, xmm0, xmm0
vzeroupper
vcvtss2sd xmm1, xmm1, xmm1
ret
More efficient would be:
f: # @f
vmovsd xmm0, qword ptr [rdi + 15999984] # xmm0 = mem[0],zero
vmovsd xmm1, qword ptr [rdi + 15999992] # xmm1 = mem[0],zero
vcvtsd2ss xmm0, xmm0, xmm0
vcvtsd2ss xmm1, xmm1, xmm1
vcvtss2sd xmm0, xmm0, xmm0
vcvtss2sd xmm1, xmm1, xmm1
ret
If we change the line complex float p = 1.0; to complex double p = 1.0; then
the sinking happens correctly.