Ah, that's better. Actually would work correctly on big-endian platforms (should we ever support one) and with way less code.