[dan-robertson]

The whole blog is excellent.

One issue with looking at instructions retired for small functions is that performance of small functions may be dominated by cache misses (and not having branch predictor data) so two versions may execute a similar number of instructions but have quite different perf due to fewer branches or better memory access patterns. But I guess if you’re optimising that then you’ll know to look at that instead. I guess the moral of ‘get a measurement setup that is good enough to reliably measure the thing you actually care about’ still holds.