Wonder how a manually written vectorized 8 bit version would compare (with some padding inserted at the end of the array to avoid extra checks). Work size/bandw

Wonder how a manually written vectorized 8 bit version would compare (with some padding inserted at the end of the array to avoid extra checks). Work size/bandwidth in a small test like this shouldn't matter, prefetching/write combining should be essentially perfect and bandwidth won't show up until it's a bottleneck/we test throughput. Besides the lacking optimization due to aliasing, could there also be some microcode iffyness be going on? Can an x86/64 cpu just increment an 1 byte value positioned at a random memory location?