Related paper: Context threading

http://www.cs.toronto.edu/syslab/pubs/demkea_context.ps

The gist is you can structure your interpreter to improve branch prediction by keeping your vm & hardware in similar states.