https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118949
Bug ID: 118949
Summary: RISC-V: Extra FRM writes since GCC-14.2
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: palmer at gcc dot gnu.org
Target Milestone: ---
GCC 15 is emitting more FRM writes than GCC 14, and as far as I can tell
they're unnecessary. For example, under `-march=rv64gcv -mabi=lp64d -O3
-ffast-math`
#include <cmath>
void func(const float *a, const float *b, float *c)
{
for (long i = 0; i < 1024; ++i) {
float a_round = std::lround(a[i]);
float b_round = std::lround(b[i]);
c[i] = a_round + b_round;
}
}
my main loop for 15 is
.L4:
fsrmi 4
vsetvli a5,a4,e32,mf2,ta,ma
vle32.v v2,0(a0)
vle32.v v3,0(a1)
slli a3,a5,2
sub a4,a4,a5
add a0,a0,a3
add a1,a1,a3
vfwcvt.x.f.v v1,v2
fsrm a6
vfncvt.f.x.w v1,v1
fsrmi 4
vfwcvt.x.f.v v2,v3
fsrm a6
vfncvt.f.x.w v2,v2
vfadd.vv v1,v1,v2
vse32.v v1,0(a2)
add a2,a2,a3
bne a4,zero,.L4
which has those odd back-to-back FRM writes. They're not in 14
.L4:
fsrmi 4
vsetvli a5,a4,e32,mf2,ta,ma
vle32.v v2,0(a0)
vle32.v v3,0(a1)
slli a3,a5,2
sub a4,a4,a5
add a0,a0,a3
add a1,a1,a3
vfwcvt.x.f.v v1,v2
vfwcvt.x.f.v v2,v3
fsrm a6
vfncvt.f.x.w v1,v1
vfncvt.f.x.w v2,v2
vfadd.vv v1,v1,v2
vse32.v v1,0(a2)
add a2,a2,a3
bne a4,zero,.L4
LLVM isn't vectorizing for me, not sure if that's some flag issue or if it's
costing even the pair of FRM writes as bad.
