[Bug tree-optimization/126461] New: [13/14/15/16/17 Regression] Wrong fold of -fma (a, b, c) to fnms (a, b, c) when signed zeros are important

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126461

            Bug ID: 126461
           Summary: [13/14/15/16/17 Regression] Wrong fold of -fma (a, b,
                    c) to fnms (a, b, c) when signed zeros are important
           Product: gcc
           Version: 17.0
            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: ---

__attribute__((noipa)) double
neg_fma (double a, double b, double c)
{
  return -__builtin_fma (a, b, c);
}

__attribute__((noipa)) double
neg_fms (double a, double b, double c)
{
  return -__builtin_fma (a, b, -c);
}

int
main (void)
{
  if (!__builtin_signbit (neg_fma (1.0, 1.0, -1.0)))
    __builtin_abort ();
  if (!__builtin_signbit (neg_fms (2.0, 3.0, 6.0)))
    __builtin_abort ();
  return 0;
}

aborts with GCC and passes with Clang at -O2.
IMO the folds should happen only when !HONOR_SIGNED_ZEROS.
GCC 8 technically got this right (the testcase doesn't abort there)