add benchmark for slice vectorization... expected it to be little or zero benefit... turns out to be 20x speedup. Something is wrong.