From: Stefan Sperling <stsp@stsp.name>
Subject: Re: use patience diff for some diff3 merges
To: gameoftrees@openbsd.org
Date: Wed, 16 Jun 2021 11:05:22 +0200

On Thu, Jun 10, 2021 at 06:07:51PM +0200, Stefan Sperling wrote:
> While using 'got histedit' I have encountered another mis-merge where
> a change was dropped from history which shouldn't have been dropped.
> I could reproduce this in 'got cherrypick' as well, by cherry-picking
> the affected commit from my local 'packcreate' branch to 'main'.
> 
> This particular problem disappears when diff3 uses the Patience diff
> algorithm instead of the Myers algorithm to perform the merge.

Thinking about this some more, we will probably want to investigate
the 'recursive' merge Git is using. In addition to providing knobs
for the diff algorithm, Git's recursive merge performs a merge of
common ancestors first if there is no single definite common ancestor.
This is something we could implement in Got, too.
Apparently this avoids merge conflicts and mis-merges. Provided the
merge of the common ancesors is conflict-free, of course. Otherwise,
Git falls back on a regular 3-way merge of the given input files.

For now, I'd still like to go forward with the proposed switch to
Patience diff for cherrypick/backout/rebase/histedit. It will not be
the perfect solution. There might still be other cases where the
merge result contains something unexpected. But so far the only
problem I've seen is that conflict chunks can be a bit large. Which
is not really an issue since more lines of context can also help to
figure out how to resolve the conflict.

Any objections?