Keep in mind that the microbenchmark uses exactly fixed cardinalities, with no randomness, so the branches in the original algorithm will be perfectly predicted

Keep in mind that the microbenchmark uses exactly fixed cardinalities, with no randomness, so the branches in the original algorithm will be perfectly predicted and hence will tend to inflate the reported performance versus a case where the number of spaces has an expected cardinality but with enough variance that the loop mispredicts once for every block. So I suspect the cutover point for the pext approach to be better than the microbenchmark suggestions for typical real-world inputs. It would also be interesting to see a lookup based approach, using multiple lookups (probably with gather) to build the shuffle control for a single block. 0.7 cycles per element means ~45 cycles per 64 byte block, so that's still quite a large budget per block, and the load ports have little pressure.