[Bug target/84251] [8 Regression] Performance regression in gcc 8 when comparing floating point numbers

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84251

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Testcase without headers:

int
cmp (double a, double b)
{
  if (__builtin_isnan (a) || __builtin_isnan (b))
    __builtin_abort ();
  return a == b;
}

In *.optimized we have:
  if (a_2(D) unord b_3(D))
    goto <bb 3>; [0.04%]
  else
    goto <bb 4>; [99.96%]

  <bb 3> [0.04%] [count: 0]:
  __builtin_abort ();

  <bb 4> [99.96%] [count: INV]:
  _1 = a_2(D) == b_3(D);
  _4 = (int) _1;
  return _4;

For int cmp (double a, double b) { return a == b; } we emit the same code as
before though, but aren't able to figure out that we've already performed the
unordered comparison before anymore.