The graphs are extremely interesting - there's a significant qualitative difference between what works best: in-cache, "least branches, least work" works best.

The graphs are extremely interesting - there's a significant qualitative difference between what works best: in-cache, "least branches, least work" works best. out-of-cache, "least fetches" works best. Not at all surprising, of course, but a reminder that there's no point discussing these issues independent of the cache size. Also, no point optimizing without knowing the approximate cache size and memory footprint of your real data.