I also tried unrolling the 256-iteration loop further below, but it actually caused a
slowdown (my guess is either branch prediction stopped kicking in or the instruction
cache was being maculated).