Bad base selection / three-way-text-merge issues

[Aaron Bentley]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Aaron Bentley wrote:
> The merge base selection in bzr is usually pretty good, but here's an
> example of how it can go wrong:
> http://panoramicfeedback.com/opensource/bzr/wrong-base.svgz

Heh, maybe bzr knows best after all.  You get different conflicts if you
force R1318.  With the original base, you get a conflict in
bzrlib/transport/__init__.  With R1318, you get a conflict in
bzrlib/merge.py

This is another case where diff3 gets fewer conflicts than merge3--
merge3 gets additional conflicts in bzrlib/config.py.

What's somewhat frustrating about even the diff3 output is that you get
many lines reported as conflicts that clearly are the same:

<<<<<<< TREE
            base_rev_id = base_branch.get_rev_id(base_revision[1])
        fetch(from_branch=base_branch, to_branch=this_branch)
        base_is_ancestor = is_ancestor(this_rev_id, base_rev_id,
                                       this_branch)
...
=======
            base_rev_id = base_branch.get_rev_id(base_revision[1])
        fetch(from_branch=base_branch, to_branch=this_branch)
        base_is_ancestor = is_ancestor(this_rev_id, base_rev_id,
                                       this_branch)

Algorithmically, I can understand why diff3 does this:
- - THIS and OTHER both differ from BASE, which starts a new comparision
section.
- - The point where THIS and OTHER start to agree with BASE is the end of
the comparison section.
- - In the comparison section, THIS, OTHER, and BASE have different texts.

Still, I think we would be better served by dividing the text into
smaller sections, such that:

- - The point where THIS and OTHER start to disagree with BASE begins a
section.
- - The point where THIS and OTHER start to disagree with each other as
well as BASE starts another section.

I guess what I really want from an algorithm is:

- - compare THIS and OTHER
- - where they disagree with one another, find the equivalent lines in BASE
- - do a three-way merge; if THIS matches BASE, OTHER wins, if OTHER
matches base, THIS wins, otherwise conflict.

Aaron
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (GNU/Linux)
Comment: Using GnuPG with Thunderbird - http://enigmail.mozdev.org

iD8DBQFDV6hY0F+nu1YWqI0RAgj0AJ9hfQYN2aQl2dSCpuvBw4lq+YaCswCdGOnZ
E+s5HE+iyuyl4+mIBdxKiOI=
=/RFK
-----END PGP SIGNATURE-----