> The theory is sound, but it is going to take a good deal of effort to isolate the effects (either good or bad) in realistic benchmarks.

Oh, it was just asking for a microbenchmark on set(). I don't care of
a "real" benchmark (realistic use case) for such issue :-)

I asked because I don't understand all these complex things of the CPU
(cache size, data dependency, prefetch, ...), whereas I can compare
number of a microbenchmark :-) And I can try to reproduce it on a
different CPU and a different dataset. Because I don't understand
low-level CPU things, I always fear that a patch works well on a
specific CPU whereas it would behave badly on a completly different
CPU architecture (ex: RISC vs CISC).

All your arguments sound good and I'm always happy to see people
involved to optimize Python!