Two extra instructions in RC4 character loop give 80% performance
improvement on Core2. I still need to detect Core2 and choose this
path...