[Bug tree-optimization/98028] [8/9/10/11 Regression] __builtin_sub_overflow_p not folded to const when some constraints are known

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98028

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Note, for
unsigned f1(unsigned i, unsigned j) {
  if (j != i) __builtin_unreachable();
  return i - j;
}
this is already optimized through:
  if (vr->varying_p ()
      && (code == PLUS_EXPR || code == MINUS_EXPR)
      && TREE_CODE (op1) == SSA_NAME
      && vr0.kind () == VR_RANGE
      && symbolic_range_based_on_p (&vr0, op1))
    {
      const bool minus_p = (code == MINUS_EXPR);
      value_range n_vr1;

      /* Try with VR0 and [-INF, OP1].  */
      if (is_gimple_min_invariant (minus_p ? vr0.max () : vr0.min ()))
        n_vr1.set (vrp_val_min (expr_type), op1);

      /* Try with VR0 and [OP1, +INF].  */
      else if (is_gimple_min_invariant (minus_p ? vr0.min () : vr0.max ()))
        n_vr1.set (op1, vrp_val_max (expr_type));

      /* Try with VR0 and [OP1, OP1].  */
      else
        n_vr1.set (op1, op1);

      range_fold_binary_expr (vr, code, expr_type, &vr0, &n_vr1);
    }
(and matching if below for the other range/operand pair) + the symbolic
handling
in range_fold_binary_symbolics_p -> extract_range_from_plus_minus_expr.
And that is also able to optimize
unsigned f1(unsigned i, unsigned j, unsigned *r) {
  if (j >= i) __builtin_unreachable();
  return __builtin_sub_overflow (i, j, r);
}
*r setting to 0.
But we don't have any such support for symbolics in
check_for_binary_op_overflow.

I guess it could be added even for GCC 11, the question is if it would be then
usable even for GCC 12 with smbolic ranges in ranger, or not.