[Bug middle-end/68067] [4.9/5/6 Regression] Wrong constant folding

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68067

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so we transform a - (b - c) to (c - b) + a, that's invalid of course.  This
also started with GCC 4.3.0.

negate_expr_p has

    case MINUS_EXPR:
      /* We can't turn -(A-B) into B-A when we honor signed zeros.  */
      return !HONOR_SIGN_DEPENDENT_ROUNDING (element_mode (type))
             && !HONOR_SIGNED_ZEROS (element_mode (type))
             && reorder_operands_p (TREE_OPERAND (t, 0),
                                    TREE_OPERAND (t, 1));

and fold_binary:

      /* A - B -> A + (-B) if B is easily negatable.  */
      if (negate_expr_p (arg1)
          && !TYPE_OVERFLOW_SANITIZED (type)
          && ((FLOAT_TYPE_P (type)
               /* Avoid this transformation if B is a positive REAL_CST.  */
               && (TREE_CODE (arg1) != REAL_CST
                   ||  REAL_VALUE_NEGATIVE (TREE_REAL_CST (arg1))))
              || INTEGRAL_TYPE_P (type)))
        return fold_build2_loc (loc, PLUS_EXPR, type,
                            fold_convert_loc (loc, type, arg0),
                            fold_convert_loc (loc, type,
                                              negate_expr (arg1)));

we can see "!TYPE_OVERFLOW_SANITIZED" here, a clear sign that sth is wrong.

I think the PLUS_EXPR case of negate_expr_p has the same issue
A + B negated as -B - A or -A - B as as if A == INT_MIN and B == 0 we
again get -0 - INT_MIN instead of INT_MIN + 0.

Testcase:

int main(void)
{
  int a = -1;
  static int b = -2147483647 - 1;
  static int c = 0;
  int t = a - (b + c*-2);
  if (t != 2147483647) { __builtin_abort(); }
  return 0;
}