On Tue, Oct 28, 2008 at 01:11:16PM +0000, Pierre Habouzit wrote: > On Tue, Oct 28, 2008 at 05:45:40AM +0000, Jeff King wrote: > > On Mon, Oct 27, 2008 at 01:07:10PM -0700, Linus Torvalds wrote: > > > Of course, if the commit is reachable through multiple sources (which is > > > common), our particular choice of "first" reachable is entirely random > > > and depends on the particular path we happened to follow. > > > > Hmm. It would be nice to keep even a simple counter to get a "distance" > > from the ref and choose the one with the smallest distance (I think we > > can get away without the complex rules that git-describe uses, since we > > are not interested in describing the full commit, but rather finding a > > "likely" branch). > > > > However, that would require making multiple passes over the commit > > graph, which might impact the startup speed. > > Actually I tried to do that (and you meant name-rev --contains rather > than describe actually ;p), and I stopped because it's too slow. I > believe the proper way to do that is to help git-log knowing which are > the short (topic) branches, and to crawl incrementally using a > date-based hack. This would basically work a bit like this. Let's > imaging you want to crawl "next" in git and know which topics from pu > are in it. You would say e.g.: > > git-log --topics=*/* next (as pretty much every */* is a topic branch > for git.git). > > > Then one has to know which are the heads of every topic branches first, > then crawl next the usual way, except that when you arrive to a point > that is a topic branch head, you don't look that way. You remember the > date of that point, and continue to crawl "next" a bit further so that > you can start annotating the topic's commits you've stumbled upon. And > you do that, you look at jd/topic (as in John Doe topic branch) and mark > all the commits as being from jd/topic, until you either go back to some > commit from next, or your current commit date is younger than your > current "next" crawling front. In the former case, you're done with that > topic, in the latter you must continue to preprocess "next" a bit more. My description is clumsy, as we probably sometimes want to crawl topics that aren't merged either like in git-log --topics=*/* --all. But we just have to mark the current "front" of the walk, and have it sorted between things that are part of topic heads and the thing that are not. We don't even need to decorate the things that are _not_ part of the topics at all I'd say, because that's where you definitely need distance based algorithms and multiple passes to yield correct results. Whereas for the specific case of topic branches, which is what people want (see Ingo's mails on the subject a week or so ago), the algorithm I propose is, I guess, mostly O(n + m) in the number of walked commits (n) and references (m). IOW fast. -- ·O· Pierre Habouzit ··O madcoder@xxxxxxxxxx OOO http://www.madism.org
Attachment:
pgpdRxyjbzn9W.pgp
Description: PGP signature
