On Thu, Jul 24, 2025 at 03:03:27PM +0200, Patrick Steinhardt wrote:
> I've quickly hacked something together now, see the work-in-progress
> patch below. The patch does not yet handle reflogs, but that isn't too
> hard to implement.
>
> And these changes indeed speed up things by quite a lot: instead of
> hours it now takes 7 seconds :) I'll polish this patch series and will
> likely send it in tomorrow.
Cool. I agree with all of the pain points you outlined, and the general
direction. There was one other sub-optimal thing I noticed, which was...
> - refs_for_each_ref(get_main_ref_store(the_repository),
> - read_remote_branches, &rename);
> [...]
> + result = refs_for_each_rawref(get_main_ref_store(the_repository),
> + queue_one_rename, &rename);
Both before and after your patch, we're iterating over _all_ refs and
skipping ones that aren't in "refs/remotes/<remote>/". If we just ask to
iterate over that subset of refs, then we save the effort of iterating
over the others that we don't care about.
But:
1. We have refs_for_each_ref_in() and refs_for_each_rawref(), but no
refs_for_each_rawref_in(). Feels like it should be easy to add it,
though.
2. It's an obvious small optimization, but it doesn't help us in a
big-O way. Iterating the refs is obviously O(n), and in the worst
case rewriting the packed-refs file is likewise O(n). So I wouldn't
expect to see the dramatic improvements you found by removing the
quadratic bits. But I'd bet it's still measurable in a repo with a
lot of refs (and maybe with reftables it actually would be bigger,
since the goal there is to amortize the rewrites).
-Peff
