I think the gcc.c-torture/execute/20010114-2.c test is invalid, because it will fail if you have excess precision in intermediate calculations, or if you just have a float implementation with more than the minimal precision required.