[Bug tree-optimization/93681] Wrong optimization: instability of x87 floating-point results leads to nonsense

[vincent-gcc at vinc17 dot net via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93681

--- Comment #5 from Vincent Lefèvre <vincent-gcc at vinc17 dot net> ---
Note that this is also reproducible with:

----------------------------------------------------------------------
#include <stdio.h>

__attribute__((noipa,optnone)) // imagine it in a separate TU
static double opaque(double i) { return i; }

int main()
{
  int z = opaque(1.0) + 0x1p-60 == 1.0;

  printf("z = %d\n", z);
  if (z)
    puts("z is one");
}
----------------------------------------------------------------------

i.e. where everything has the double type in the tests.

But with a similar program using _Float16 under x86_64 (where one also has
excess precision by default, the evaluation type being float) instead of double
with x87 math, i.e. with

static _Float16 opaque(_Float16 i) { return i; }

and

  int z = opaque(1.0f16) + 0x1p-17f16 == 1.0f16;

the issue does not occur (with or without -fexcess-precision=16).

So this seems specific to the x87 excess precision, not a general bug involving
excess precision.