Thinking about cache-optimality, I'm wondering if anyone is using a hybrid: the first levels are in this way (index K branches to 2K and 2K+1 if 1-based), but the end levels are consecutive to then can be loaded in a single cacheline anyway.