https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122219
--- Comment #5 from Pengfei Li <pfustc at gcc dot gnu.org> ---
I might have reduced it too far. The original code was written with x86
intrinsics and compiled with the SIMDe (SIMD everywhere) library. A more
original case (like below) doesn't have undefined data.
#define SIMDE_ENABLE_NATIVE_ALIASES
#include "simde/x86/avx2.h"
void foo(__m256& v, unsigned int n) {
__m128 f0 = {1.0f, 2.0f, 3.0f, 4.0f};
__m128 f1 = {5.0f, 6.0f, 7.0f, 8.0f};
for (int i = 0; i < n; i++) {
f0 = f0 + f0;
f1 = f1 + f1;
v = _mm256_castps128_ps256(f0);
v = _mm256_insertf128_ps(v, f1, 1);
}
}
Even with -fstack-reuse=none, the stores don’t sink.
I acknowledge the workload code isn’t well written. We can either move the
assignments to v out or use _mm256_set_m128 instead of the cast + insert.
However, we also observed that even for this case LLVM can sink the stores. So
perhaps there's still room for optimization.
