[Bug tree-optimization/93744] [8/9/10 Regression] Different results between gcc-9 and gcc-7

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93744

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
As for the signed zeros, I can't really reproduce it.  Tried with:
__attribute__((noipa)) void
foo (int x)
{
  double a = 0.0;
  double b = -0.0;
  _Bool c = x > 0;
  _Bool d = x > 1;
  double e = c;
  double f = d;
  double g = e * a;
  double h = f * b;
  if (__builtin_copysign (1.0, g) != 1.0
      || __builtin_copysign (1.0, h) != -1.0)
    __builtin_abort ();
}

int
main ()
{
  foo (0);
  return 0;
}

The thing is that convert in (convert (cmp $0 $1)) seems to match just
CASE_CONVERT:, but result of a comparison is always integral (or vector
integral) and so there would need to be a FLOAT_EXPR rather than
{NOP,CONVERT}_EXPR to convert it to floating point.
Allowing other comparisons or any boolean results is certainly reasonable, but
would be an enhancement request for GCC 11.