http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46943
--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-12-16 15:10:12 UTC --- Hmm. I think at least postreload CSE should handle it as we then know that zero-extends extend to the full HW register. Pre-regalloc would be difficult as we'd need some new form of canonicalization with an associated cost (think of a zero_extend:TI).
