[Bug lto/47924] [4.6 Regression] Missed optimization with LTO

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47924

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-02-28 15:54:55 UTC ---
Hmm.  TYPE_CANONICAL seems to be broken (NULL) and we use alias-set 0 for both
int and short.  Probably because of

      /* Type merging will get confused by the canonical types as they
         are set by the middle-end.  */
      if (in_lto_p)
        TYPE_CANONICAL (node) = NULL_TREE;
      *nodep = node = gimple_register_type (node);

we are missing to register the canonical type here.

I have a patch.