https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115161

--- Comment #21 from Sergei Trofimovich <slyfox at gcc dot gnu.org> ---
Shrunk the example down to a single simpler function while preserving the
original masking intent:

```c
cat bug.cc
#include <stdint.h>
#include <string.h>
#include <emmintrin.h>

__attribute__((noipa))
static void assert_eq_p(void * l, void * r) {
    char lb[16];
    char rb[16];

    __builtin_memcpy(lb, l, 16);
    __builtin_memcpy(rb, r, 16);

    if (__builtin_memcmp(lb, rb, 16) != 0) __builtin_trap();
}

__attribute__((noipa))
static void assert_eq(__m128i l, __m128i r) { assert_eq_p(&l, &r); }

int main() {
  const __m128i su = _mm_set1_epi32(0x4f800000);
  const __m128  sf = _mm_castsi128_ps(su);

  const __m128  overflow_mask_f32 = _mm_cmpge_ps(sf,
_mm_set1_ps(2147483648.0f));
  const __m128i overflow_mask = _mm_castps_si128(overflow_mask_f32);

  const __m128i conv = _mm_cvttps_epi32(sf); // overflows
  const __m128i yes = _mm_set1_epi32(INT32_MAX);

  const __m128i a = _mm_and_si128(overflow_mask, yes);
  const __m128i na = _mm_andnot_si128(overflow_mask, conv);

  const __m128i conv_masked = _mm_or_si128(a, na);

  const __m128i actual = _mm_cmpeq_epi32(conv_masked,
_mm_set1_epi32(INT32_MAX));
  const __m128i expected = _mm_set1_epi32(-1);

  assert_eq(expected, actual);
}
```

The discrepancy:

Ok:
    $ /tmp/gb/gcc/xg++ -Wall -B/tmp/gb/gcc bug.cc -o bug -O0 && ./bug
Bad:
    $ /tmp/gb/gcc/xg++ -Wall -B/tmp/gb/gcc bug.cc -o bug -O2 && ./bug
    Illegal instruction (core dumped)

Reply via email to