Re: git merge heuristic

From: Jeff King <peff@peff.net>
Date: Mar 22, 2013, 22:33 UTC
Message-ID: <20130322223330.GB13207@sigill.intra.peff.net>
In-Reply-To: <ECEA5ACFB4B2BC4A81A93F032519ABCAD47F79@minsk.us.stoke.com>

On Fri, Mar 22, 2013 at 11:53:04AM -0700, Senthil Natarajan wrote:

> I want to learn about how Git compares patches while doing a merge. 
> For example, if a patch has been cherry-picked from branch A to branch
> B, and then downstream we do a "git merge" from A to B, how does Git
> know to skip the cherry-picked patch?

It doesn't. Git's 3-way merge only looks at three things: where each side of the merge ended, and what their common ancestor looked like.

So when you cherry-pick a commit, as long as the content in the file ended up the same, there is no conflict. And it doesn't matter if it happened by cherry-picking, or if you just happened to make a sequence of commits that ended in the same state.

However, we do perform such detection during a rebase, in which we try to skip patches that have already been applied upstream.

> It would have a different SHA-1, so what is the comparison
> algorithm/heuristic?  What happens if the comment is different, but the
> actual patch is identical?

Yes, the commit will have a different sha1. For that, we use the "patch-id", which is basically a sha of the contents of the diff of the commit against its parent. See the manual for git-patch-id, git-cherry, and the --cherry-* options of "git log".

It will not find all duplicate commits (e.g., it will miss ones where there was a conflict during cherry-pick, or even where the context is slightly different). However, in many cases, rebase can also realize while applying a patch that it has already been applied, and skip it then.

-Peff