https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127364
Bug ID: 127364
Summary: __builtin_assoc_barrier does not works in arm neon
Product: gcc
Version: 16.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: z00823823 at outlook dot com
Target Milestone: ---
```c++
#include <arm_neon.h>
float32x4_t round_values(float32x4_t x) {
const float32x4_t big = vdupq_n_f32(0x1p24f);
const float32x4_t sum = __builtin_assoc_barrier(vaddq_f32(x, big));
return vsubq_f32(sum, big);
}
```
compile with `-O2 -ffast-math`, it compiles to nothing, meaning gcc incorrectly
re-associates the add and sub operation.
however, on x86_64, similar code:
```c++
#include <immintrin.h>
__m128 with_barrier(__m128 x) {
const __m128 big = _mm_set1_ps(0x1p24f);
const __m128 sum =
__builtin_assoc_barrier(_mm_add_ps(x, big));
return _mm_sub_ps(sum, big);
}
```
compile with `-O2 -ffast-math`, it compiles to
```asm
"with_barrier(float vector[4])":
movss xmm1, DWORD PTR .LC1[rip]
shufps xmm1, xmm1, 0
addps xmm0, xmm1
movss xmm1, DWORD PTR .LC3[rip]
shufps xmm1, xmm1, 0
addps xmm0, xmm1
ret
```
the add and sub operation is retained, correctly.
https://gcc.godbolt.org/z/YrY3bdaE7
