teach codegen to turn trunc(zextload) into load when possible.

This doesn't occur much at all, it only seems to formed in the case
when the trunc optimization kicks in due to phase ordering.  In that
case it is saves a few bytes on x86-32.

llvm-svn: 101350