[Bug tree-optimization/126455] New: Wrong folding of (a < b) ^ (a > b) to a != b for FP vectors

[ktkachov at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126455

            Bug ID: 126455
           Summary: Wrong folding of (a < b) ^ (a > b) to a != b for FP
                    vectors
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64

Consider the testcase:
#include <stdio.h>

typedef float v4sf __attribute__((vector_size (16)));
typedef int   v4si __attribute__((vector_size (16)));

__attribute__((noipa)) v4si vec_xor (v4sf a, v4sf b) { return (a < b) ^ (a >
b); }
__attribute__((noipa)) v4si vec_ior (v4sf a, v4sf b) { return (a < b) | (a >
b); }
__attribute__((noipa)) int  scl_xor (float a, float b) { return (a < b) ^ (a >
b); }

int
main (void)
{
  float n = __builtin_nanf ("");
  v4sf a = { n, n, n, n }, b = { 1, 1, 1, 1 };
  v4si x = vec_xor (a, b);
  v4si o = vec_ior (a, b);
  int s = scl_xor (n, 1.0f);
  int bad = 0;

  printf ("vector ^ : %d  (must be 0)\n", x[0]);
  printf ("vector | : %d  (must be 0)\n", o[0]);
  printf ("scalar ^ : %d  (must be 0)\n", s);

  if (x[0] != 0) { printf ("FAIL: vector exclusive or\n"); bad = 1; }
  if (o[0] != 0) { printf ("FAIL: vector inclusive or\n"); bad = 1; }
  if (s != 0)    { printf ("FAIL: scalar\n"); bad = 1; }
  return bad;
}

with -O2 this gives different results between GCC and LLVM.
Expected, and what the scalar version gives:

```
vector ^ : 0  (must be 0)
vector | : 0  (must be 0)
scalar ^ : 0  (must be 0)
```

Observed:

```
vector ^ : -1  (must be 0)
FAIL: vector exclusive or
```

This fold should only happen under -ffinite-math-only IMO