I was one of the people who asked you about bigger benchmarks at the
summit (from the back row). I do think that micro-benchmarks can
sometimes lead you to misoptimize. But I have to admit that I don't
have any bigger benchmarks myself.
And I certainly agree that the work you've been doing has in general
been very appropriate and effective. Thanks for doing it.
Eventually we are going to hit the point where the choices between
optimizations are going to be harder to make--where the function runs
somewhat faster, but gets noticeably bigger, so you have to consider
whether the effect on the instruction cache outweighs the other
performance gain. Etc. I know you know this already.