is a simple change, but seems to improve code a little.  For example, on
256.bzip2, we went from 75.0s -> 73.33s (2% speedup).

llvm-svn: 15004