https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104485
Bug ID: 104485
Summary: x378 fmod inline code is slow
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
In 526.blender_r one can see us expanding fmod as
fld1
fldl (%rsi)
.L2:
fprem
fnstsw %ax
testb $4, %ah
jne .L2
fstp %st(1)
...
which is quite a bit slower than just calling into libm. The case in
question is actually special and can be approximated by
void foo (double * __restrict s, double *d)
{
s[0] = fmod(d[0], 1.0f);
s[1] = fmod(d[1], 1.0f);
}
where obtaining the fractional part of {d[0], d[1]} might even be vectorizable.
Building 526.blender_r with -fno-builtin-fmod (-mno-fancy-math-387 doesn't do
the trick here) speeds it up by 1% on Zen2.
