Here's a completed patch.  Bill, if you could regression test it, that'd 
be great! (feel free to augment the testcase)

Richard, rather than copy set_underlying_type's type cloning I added a 
check in it for error_mark_node as the existing name.  If you'd prefer 
that remain unchanged, I can duplicate the type cloning into rs6000_vt 
itself.

nathan

-- 
Nathan Sidwell