In an aside he mentions you should use bubblesort instead of quicksort for small arrays, due to cache locality, etc. I'd recommend using insertion sort instead of bubblesort - it does much better in both cache locality and branch performance (one branch prediction miss per key).
That seems a strange assertion - I was under the impression that insertion sort beats bubble sort on small arrays, pretty much always, and even in 'nearly sorted' arrays, which is often touted as a strength of bubble sort.
