There are probably much larger speedups to be had by renumbering locally instead
of looping over the whole function. For now, the greedy register allocator is
25% faster.

llvm-svn: 126926