Simplified handling of the "final byte", yielding 0.5 cycles/hash improvement.
Rearranged Loop64 memory addresses, yielding 0.5 GB/s large-block hashing improvement.
Note that the output values of the function changed.