https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97137
Bug ID: 97137
Summary: Missed add-with-carry after FP compare
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Extracted from 508.namd_r
void foo (double *r2i, double r2_0, double cutoff2_delta, int jout)
{
bool test0 = (r2_0 < cutoff2_delta);
jout += test0;
r2i[jout] = r2_0;
}
produces
xorl %eax, %eax
comisd %xmm0, %xmm1
seta %al
addl %esi, %eax
movsd %xmm0, (%rdi,%rax,8)
while llvm optimizes this to
ucomisd %xmm1, %xmm0
adcl $0, %esi
movslq %esi, %rax
movsd %xmm0, (%rdi,%rax,8)
possibly saving one register. Note that clang needs -ffinite-math-only for
this.
I'm not sure how r2_0 < cutoff2_delta behaves with NaNs, but with
-ffinite-math-only we'd be sure there are none and thus even comisd set
of CF should be enough to optimize this.
