This is a bit faster at -O2 because memcpy()/memmove()/memset() are
vectorized. The code is also cleaner.