[Bug rtl-optimization/67218] New: Combine incorrectly folds (double) (float) (unsigned)

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67218

            Bug ID: 67218
           Summary: Combine incorrectly folds (double) (float) (unsigned)
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rsandifo at gcc dot gnu.org
  Target Milestone: ---

simplify-rtx.c folds (double) (float) (some-int) into
(double) (some-int) if it can prove that the intermediate
(float) doesn't have any effect.  This check is based on
the number of known "sign bit copies", which inherently assumes
a signed interpretation of (some-int) and is therefore only
valid for FLOAT.  For UNSIGNED_FLOAT we can end up treating a
large unsigned value as fitting in a float.

The following test case fails at -O and above on aarch64-linux-gnu
because of this:

extern void abort (void) __attribute__ ((noreturn));

double __attribute__ ((noinline, noclone))
foo (unsigned int x)
{
  return (double) (float) (x | 0xffff0000);
}

int
main ()
{
  if (foo (1) != 0x1.fffep31)
    abort ();
  return 0;
}