Register allocation can be quite tricky, and sometimes it can only explore a small part of the problem space, so if you don't start the algorithm with exactly the same seed you might end up with significantly different code in certain functions.