[Bug tree-optimization/45034] [4.3/4.4/4.5/4.6 Regression] "safe" conversion from unsigned to signed char gives broken code

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

------- Comment #11 from rguenth at gcc dot gnu dot org  2010-07-28 09:32 -------
I suppose that

    uy = -ux;

is really

    uy = (unsigned char) -(int) ux;

as unsigned char promotes to int.  We then incorrectly
narrow this to -(signed char) ux.

We do this again in convert_to_integer.  Our bag of premature (and bogus)
optimizations.

        case NEGATE_EXPR:
        case BIT_NOT_EXPR:
          /* This is not correct for ABS_EXPR,
             since we must test the sign before truncation.  */
          {
            tree typex;

            /* Don't do unsigned arithmetic where signed was wanted,
               or vice versa.  */
            if (TYPE_UNSIGNED (TREE_TYPE (expr)))
              typex = unsigned_type_for (type);
            else
              typex = signed_type_for (type);
            return convert (type,
                            fold_build1 (ex_form, typex,
                                         convert (typex,
                                                  TREE_OPERAND (expr, 0))));

well - I have no idea why we can't always choose an unsigned type here
(testing that now).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45034