Doesn't the unrolled version skip the speculation for three out of every four entries? If so, it seems suspicious that it does so much better when it looks like it's a mix between the first and second version instead of just an unrolled implementation of the second version.
I'd also be interested in seeing how this performs when the data isn't allocated in one chunk (which seems to be the absolute best case scenario for this trick). Will it end up worse than the "naive" version due to all the branch mispredictions?
I'd also be interested in seeing how this performs when the data isn't allocated in one chunk (which seems to be the absolute best case scenario for this trick). Will it end up worse than the "naive" version due to all the branch mispredictions?