the same allocation size but different primitive sizes(e.g., <3xi32> and
<4xi32>).  When ScalarRepl promotes them, it can't use a bit cast but
should use a shuffle vector instead.

llvm-svn: 129472