Also, the benchmark of exactly as written there is missing some overhead, as it will always be cached. Memory accesses aren't likely to be an issue here, but I

Also, the benchmark of exactly as written there is missing some overhead, as it will always be cached. Memory accesses aren't likely to be an issue here, but I always try to run over a couple dozen or hundred of megs of input in benchmarks to ward against that.