On Tue, Jun 21, 2022 at 08:11:49AM +0200, Rasmus Villemoes wrote: > On 20/06/2022 02.41, Kent Overstreet wrote: > > > Rasmus pointed out that -fno-strict-aliasing is going to cause gcc to generate > > nasty code, and indeed it unfortunately does but according to worst case > > scenario microbenchmarks it's not a problem for actual performance. > > Well, that's not how I interpreted those numbers, but, except if they > showed an improvement, how much is acceptable is of course always a > matter of judgment. > > However, what's really annoying and somewhat dishonest is that you're > not including those numbers, nor the methodology, in either the cover > letter or commit itself. There's nothing dishonest about it, and I wasn't claiming an improvement; merely no regressions (some were a 5-10% percent up, some down by around the same amount, overall it was a wash). My priority simply isn't microoptimizing everything. I find that programmers who chase optimizing every loop and are constantly trying to shave instructions everywhere they can end up with code where the large scale structure is a mess, and that's where you miss out on the _real_ performance opportunities. My priority is clean, readable, simple, easy to work on code, because _that_ is the code that becomes fast in the long run. Premature optimization really is the root of all evil, and I am _absolutely_ going to try to drive the discussion away from shaving cycles when there's new APIs to get right and messy refactorings to complete.
