[Bug rtl-optimization/126521] New: Wrong RTL ifcvt with abs

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126521

            Bug ID: 126521
           Summary: Wrong RTL ifcvt with abs
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64

__attribute__((noipa)) double
g (double a)
{
  double x;
  if (a < 0.0)
    x = a;
  else
    x = -a;
  return x;
}

static unsigned long long
bits (double d)
{
  unsigned long long u;
  __builtin_memcpy (&u, &d, 8);
  return u;
}

static double
from_bits (unsigned long long u)
{
  double d;
  __builtin_memcpy (&d, &u, 8);
  return d;
}

int
main (void)
{
  double mnan = from_bits (0xfff8000000000000ULL);  /* -NaN */
  double pnan = from_bits (0x7ff8000000000000ULL);  /* +NaN */

  /* a < 0.0 is false for either NaN, so both take the else arm and the
     result is the operand with its sign bit flipped.  */
  if (bits (g (mnan)) != 0x7ff8000000000000ULL)
    __builtin_abort ();
  if (bits (g (pnan)) != 0xfff8000000000000ULL)
    __builtin_abort ();

  /* Ordinary values must still work.  */
  if (g (-2.0) != -2.0 || g (3.0) != -3.0)
    __builtin_abort ();

  return 0;
}

aborts on aarch64 with -O2  -fno-signed-zeros -fno-ssa-phiopt but passes at -O0