Re: [PATCH 1/3] commit-reach: deduplicate queue entries in paint_down_to_common
- From
Kristofer Karlsson <krka@spotify.com>
- Date
- May 25, 2026, 07:53 UTC
- Message-ID
- <CAL71e4ODJeCJctKg=3o9PKD6Rw3_xHnrjc+zT_MYFc=CdNc59A@mail.gmail.com>
- In-Reply-To
- <xmqqse7gez5l.fsf@gitster.g>
Good catch Jeff! I think it's possible that I missed the flag cleanup case here, but it's also possible that I got lucky and it worked anyway. That said, I think the observation in the other email thread/commit is key here. I will reply back in that one, but it seems like this can all be simplified using Jeff's idea with an amortized O(1) solution by caching a known non-stale entry in the queue, and thus becomes obsolete. I will post a new patchset when the discussion slows down.
As for general flag management, I will spend some more time thinking about it. I don't fully trust static code analysis to work, but some cheap assertion based model might give a nice trade-off.
Thanks for all the feedback! - Kristofer
On Mon, 25 May 2026 at 09:17, Junio C Hamano <gitster@pobox.com> wrote:
Show 18 quoted lines
> > Kristofer Karlsson <krka@spotify.com> writes: > > > While doing the audit I noticed that reasoning about flag safety is > > currently entirely manual. Would there be interest in something more > > systematic (e.g. runtime registration/assertion, dynamic allocation or static > > analysis of flag usage)? I have some local work on that already, but I was > > not sure if this was something worth spending time on or not. > > If there weren't existing code that are so tied to their current > uses of fixed flag bits and assumption that nobody else uses these > bits outside their intended use, I'd love to have any of these. > Uncolliding and unbounded number of usable bits per object that are > *fast* to access would be good (and commit-slab was an attempt to > introduce a framework that can be used as the basis for such a > system). Independent of that, if we can statically analyze the uses > of these bits to prove that the same flag bits are never used at the > same time for colliding purposes, that would really be valuable.