2.5x faster for 8-bit mode without autovectorization in GCC, 2x
slower with it on x86. However, since the platforms we enable GCC
autovectorization on most probably has support for SSE2
optimization (added in the subsequent commit), this commit should
in general do good.