On Sun, Jan 16, 2022 at 1:51 PM Mark Dickinson <dicki...@gmail.com> wrote:
> On Sun, Jan 16, 2022 at 9:28 PM Guido van Rossum <gu...@python.org> wrote: > >> Does the optimization for //10 actually help in the real world? [...] >> > > Yep, I don't know. If 10 is *not* the most common small divisor in real > world code, it must at least rank in the top five. I might hazard a guess > that division by 2 would be more common, but I've no idea how one would go > about establishing that. > All of the int constants relating to time and date calculations show up frequently as well. But I'd assume -fprofile-values isn't likely to pick many to specialize on to avoid adding branches so maybe 10 is ironically it. --enable-optimizations with clang doesn't trigger value specialization (I'm pretty sure they support the concept, but I've never looked at how). > > The reason that the divisor of 10 is turning up from the PGO isn't a > particularly convincing one - it looks as though it's a result of our > testing the builtin int-to-decimal-string conversion by comparing with an > obviously-correct repeated-division-by-10 algorithm. > > Then again I'm not sure what's *lost* even if this optimization is >> pointless -- surely it doesn't slow other divisions down enough to be >> measurable. >> > > Agreed. That at least is testable. I can run some timings (but not > tonight). > BTW, I am able to convince clang 11 and higher to produce a 64:32 divide instruction with a modified version of the code. Basically just taking your assembly divl variant as an example and writing that explicitly as the operations in C: https://godbolt.org/z/63eWPczjx Taking that code and turning it into an actual test within CPython itself, it appears to deliver the desired speedup in gcc9 as well. https://github.com/python/cpython/pull/30626 for https://bugs.python.org/issue46406. 20% faster microbenchmarking with x//1 or x//17 or other non-specialized divide values. similar speedup even in --enable-optimizations builds. with both gcc9 and clang13. The compilers seem happier optimizing that code. -gps
_______________________________________________ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-le...@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/FJRKRXSXPF24C3NHGYZMVPB3ZZPCBI6A/ Code of Conduct: http://python.org/psf/codeofconduct/
