This implementation is probably slightly inefficient than before,
but that should be negligible because this is not a performance-
critical pass.

llvm-svn: 296570