[Bug tree-optimization/126471] New: Wrong parity folding with non-even precision

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126471

            Bug ID: 126471
           Summary: Wrong parity folding with non-even precision
           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: ---

typedef unsigned _BitInt(129) u129;
typedef unsigned _BitInt(66) u66;

__attribute__((noipa)) int
p129 (u129 x)
{
  return __builtin_parityg ((u129) ~x);
}

/* Even precision: the identity does hold, so this must keep working.  */
__attribute__((noipa)) int
p66 (u66 x)
{
  return __builtin_parityg ((u66) ~x);
}

int
main (void)
{
  if (p129 (0) != 1)
    __builtin_abort ();
  if (p66 (0) != 0)
    __builtin_abort ();
  return 0;
}

Passes at -O0 and aborts at -O1 and above.

     /* parity(~X) is parity(X).  * /
     (simplify
       (PARITY (bit_not @0))
       (PARITY @0))

Complementing an N bit value flips all N bits, so parity (~X) == parity (X) ^
(N & 1).  The identity should therefore only hold for an even precision.