Some of these implementations also reduce the number of branches, which is generally a good performance idea in modern pipelined, branch-predicting CPUs, but not possible in all cases or even a guaranteed win for performance.

FWIW I checked what I did with the last ring buffer I wrote: I used position + length to eke out that last cell. As the original article notes, it's not a concurrency-friendly solution, but I wasn't writing for that in this case.