> The result is quite impressive! LuaJIT continues to pleasantly surprise > the more I use it. > A quick follow-up on this for people who like performance numbers: Mike's fix came in within ~10% of our C implementation. Manually unrolling the loop 4x makes it indistinguishable from GCC -O3 (within <0.5%, and in best-of-10 measurements they each win about 50% of the time). Demetri