An out-of-order CPU will interleave things for you. Any modern CPU should have no issues running many iterations of the loop in parallel, as nothing should depend on the previous iteration.
But, given that the task at hand usually deals with very short inputs, there's not much to interleave with anyway.
Hmm, yes, there are only 2 or 3 instructions that can execute concurrently - but your observation made me realise I can eke out another by masking is_ascii more thoroughly. Thanks!
Bulk throughput isn’t really my aim, it’s just a convenient way to get numbers big enough to be easily measurable :-)
You should get much more throughput if you can interleave them with other instructions...
I wonder if that's what the benchmarks did?