[Bug middle-end/27260] [4.1 Regression] ICE in expand_expr_real_1, at expr.c:6750

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2006-04-22 13:08 -------
The problem is we expand the value in QI mode because of

      val = builtin_save_expr (val);
      len = builtin_save_expr (len);

      len_rtx = expand_expr (len, NULL_RTX, VOIDmode, 0);
      dest_mem = get_memory_rtx (dest, len);

      if (TREE_CODE (val) != INTEGER_CST)
        {
          tree cval;
          rtx val_rtx;

          cval = fold_build1 (CONVERT_EXPR, unsigned_char_type_node, val);
          val_rtx = expand_expr (cval, NULL_RTX, VOIDmode, 0);

The "easiest" solution is something like

Index: gcc/builtins.c
===================================================================
*** gcc/builtins.c      (revision 113168)
--- gcc/builtins.c      (working copy)
*************** expand_builtin_memset (tree arglist, rtx
*** 3437,3442 ****
--- 3437,3443 ----
        /* Stabilize the arguments in case we fail.  */
        dest = builtin_save_expr (dest);
        val = builtin_save_expr (val);
+       expand_expr (val, NULL_RTX, VOIDmode, 0);
        len = builtin_save_expr (len);

        len_rtx = expand_expr (len, NULL_RTX, VOIDmode, 0);


which forces expansion in the original mode.  This may be not the best fix
though, as we then get (subreg:QI (reg:SI)) for the expansion of cval.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27260