[a_e_k]

> I would probably appreciate those comments if I ever need to read branchless code, SIMD code, or similar performance-sensitve witchcraft.

I've done this before in production code. One of my only exceptions to my policy against leaving in commented-out code has been things liking leaving the unoptimized serial code commented out at the top of a section of some hand-vectorized code full of SIMD intrinsic. It clearly showed the intended "what", but was also useful to keep around as a baseline for performance testing and reference code to compare against for debugging.