https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119405
Bug ID: 119405
Summary: Missed FMA optimization for C code present in C++
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: alexander.gr...@tu-dresden.de
Target Milestone: ---
In a third-party library a test is failing on AMD Rome (Zen2) that I traced
ultimately to a function doing a simple quaternion multiplication.
Similar code is present in a C and a C++ file that are later linked together,
reducing that to exactly the same code I see an FMA in the C++ code but not in
C.
The reduced code is:
```
void mjuu_mulquat(double* res, const double* qa, const double* qb) {
double tmp[4] = {
qa[0]*qb[0] - qa[1]*qb[1] - qa[2]*qb[2] - qa[3]*qb[3],
qa[0]*qb[1] + qa[1]*qb[0] + qa[2]*qb[3] - qa[3]*qb[2],
qa[0]*qb[2] - qa[1]*qb[3] + qa[2]*qb[0] + qa[3]*qb[1],
qa[0]*qb[3] + qa[1]*qb[2] - qa[2]*qb[1] + qa[3]*qb[0]
};
res[0] = tmp[0];
res[1] = tmp[1];
res[2] = tmp[2];
res[3] = tmp[3];
}
```
Compiling it with `-O3 -std=c++11 -mavx2 -mfma` and `-O3 -std=c11 -mavx2
-mfma` shows vfmadd132pd vfnmadd132pd in C++ not present in C. Comparison:
https://godbolt.org/z/TsYqcssnM
I don't see why the code would optimize differently in those cases as I'd
expect the optimizer gets the same input which is seemingly not the case.
