Custom merge drivers: accessing the pathnames and revisions of the files being merged

Hello,

Custom merge drivers [1] provide a convenient way to extend Git's merge algorithms. But as far as I can tell, there is no way to access the pathnames and revisions of the files to merge. Those would be very useful to generate informative conflict markers, like the built-in merge strategies do.

For instance, when merging with the ort strategy and diff3 conflict style, I get markers such as:

<<<<<<< HEAD:main/my_file.txt
…
||||||| 194c4190b:main/my_file.txt
…
=======
…
 >>>>>>> origin/branch:main/my_renamed_file.txt

Those strings at the end of conflict markers, with the revision and the pathname, are very useful to understand which parts are coming from where.

When implementing a custom merge driver, I don't see how to access this information to include it in the conflict markers. Custom merge drivers are typically invoked on temporary files generated by Git with uninformative paths, such as ".merge_file_NgiKjJ".

Therefore, my merge driver is only able to generate conflicts which look like this:

<<<<<<< .merge_file_NgiKjJ
…
||||||| .merge_file_D1XtCW
…
=======
…
 >>>>>>> .merge_file_WbmrBA

Of course, in a given rebase/merge session, the order in which the conflicting parts will be presented will remain consistent, so I will generally be able to remember which revision each part corresponds to, but it's still a mental load I would ideally try to avoid. Also, if the rename detection heuristics have false positives, then I can get merge conflicts which come from unrelated files: in that case it is very useful to see the pathnames, to understand this situation better.

So, I wonder: would people be open to exposing additional parameters to merge drivers? For instance we could add parameters "%X", "%Y" "%Z" to expose those "revision:pathname" strings for each part. (I think colons cannot be part of revision names, so this can be parsed unambiguously by the custom merge driver to recover the revision and pathname independently, if needed.)

I would be happy to submit a patch for this if you think this makes sense. If it is already possible to access this information in another way, I would like to work on documenting that.

Best wishes, Antonin

[1]: https://git-scm.com/docs/gitattributes#_defining_a_custom_merge_driver

Antonin Delpeuch <antonin@delpeuch.eu> writes:

The last time this changed was in ef45bb1f (ll-merge: pass the original path to external drivers, 2015-06-04).

I may not necessarily endorse the choice of XYZ [*], but I do not fundamentally oppose to such a new feature existing. The mechanism to define a custom merge driver is designed to be future-proof in that only the parameters it uses is given as the value of merge.*.driver variable, so it is not a problem that existing merge drivers will not know what to do with "pathname in the common ancestor", "pathname on our side", and "pathname on their side".

[Footnote]

 * Whatever letters we choose, they must have mnemonic value that
   signals two of them are the both sides of the merge that are
   equal participants, and the other one is the old-file, their
   common ancestor that plays quite a different from these two in
   the merge.  I cannot tell which one of the XYZ would be the more
   special than other two, which is the primary reason why I do not
   know if XYZ is a good idea.

Thanks a lot for your insights! Yes I also very much appreciate that the extension point is built in such a way that introducing new parameters is non-breaking.

On 16/01/2024 18:51, Junio C Hamano wrote:

That makes perfect sense. How about:

- %S for the "source" pathname (corresponding to the %O file)

- %X for the first side of the merge (corresponding to the %A file)

- %Y for the second side of the merge (corresponding to the %B file)

Anyway, I'll try to work on a patch: it should be easy to adapt the letters to any other choice.

Best,

Antonin