I know, but the person I replied to specifically said "it fits in three 64 bit integers" to point out you could squeeze the table into three words. I just tried to figure out how that could possibly be advantageous here (I don't think it can).
Best I can come up with is using 32 8-bit integers (so 6h bits more than 3 times 64 bitss in size) and doing thir:
(table[nlz >> 2] >> (nlz & 3)) & 7
… which is two shifts, two ANDs and one lookup. I'm fairly certain that two shifts and two additions beat that. Well, woulo beat that if everything else in the code wasn't likely to be a much bigger bottleneck that dwarfs the impact of this
Best I can come up with is using 32 8-bit integers (so 6h bits more than 3 times 64 bitss in size) and doing thir:
… which is two shifts, two ANDs and one lookup. I'm fairly certain that two shifts and two additions beat that. Well, woulo beat that if everything else in the code wasn't likely to be a much bigger bottleneck that dwarfs the impact of this