On Tue, 2024-12-17 at 16:03 -0800, Andrii Nakryiko wrote: [...] > I agree with Ihor. I think he invested a lot of time into these > improvements, and asking him to re-do the series just to shuffle a few > patches around is just an unnecessary overhead (which also delays the > ultimate outcome: faster BTF generation with pahole). Patches #1-4 are good refactorings. Patches #5-7 introduce a global state and complication where this could be avoided. (These were necessary before Ihor figured out the trick with patch #10). E.g. 'elf_functions_list': - there is no reason for it to be global, it could be a part of the encoder, as it is now; - there is no reason for it to be a list, encoding works with a single function table and keeping it as a list just confuses reader. Same for btf_encoder_list_lock / btf_encoder_list. > And as Ihor mentioned, we might improve upon this series by > parallelizing encoders to gain some further improvements, so I think > all the internal refactoring and preparations are setting up a good > base for further work. Not really, the main insight found by Ihor is that parallel BTF encoding doesn't make much sense: constructing BTF is as cheap as copying it. I don't see much room for improvement here. [...]
