[Bug tree-optimization/126549] New: [16/17 Regression] Wrong code with ranger and int->float casts

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126549

            Bug ID: 126549
           Summary: [16/17 Regression] Wrong code with ranger and
                    int->float casts
           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: ---

/* operator_cast::fold_range (frange, tree, const irange &): the forward
   INT -> FLOAT cast.  It computes the bounds with real_from_integer in the
   destination mode, which correctly overflows to +Inf for _Float16, and then
   throws that away with frange_drop_infs, so the range of (_Float16) u for
   u in [0, 100000] comes out as [0, 65504] and the comparison h > 65504.0f16
   folds to 0.  For u = 70000 the real value is +Inf.  */
__attribute__((noipa)) int
g (unsigned a)
{
  unsigned u = a % 100001u;        /* u in [0, 100000] */
  _Float16 h = (_Float16) u;
  return h > 65504.0f16;
}

int
main (void)
{
  volatile unsigned v = 70000u;
  if (g (v) != 1)
    __builtin_abort ();
  return 0;
}

aborts on aarch64 at -O2 and passes at -O0