Re: git-cherry-pick(1) -- path

[Phillip Wood <phillip.wood123@gmail.com>]

Hi Alex and Junio

On 11/05/2024 12:46, Alejandro Colomar wrote:
> Hi Phillip, Junio,
> 
> On Fri, May 10, 2024 at 10:03:31AM GMT, Junio C Hamano wrote:
>> Phillip Wood <phillip.wood123@gmail.com> writes:
>>
>>>>>       $ git format-patch --stdout -1 $ThatCommit -- $ThisPath |
>>>>>         git am -3
>>>> Hmmm, I hadn't thought of that; very interesting!
>>>> Although I have some concerns with git-am(1); basically that it's
>>>> almost
>>>> clueless when there's a conflict.
>>>
>>> "git am -3" should be fine here as you're guaranteed to have the
>>> necessary blobs available to create conflicts - this is what "git
>>> rebase --apply" does.
>>
>> Good thing to point out.  "am -3" is just as good for this purpose
>> and "almost clueless" is a derogatory comment that requires an
>> apology ;-)
> 
> Huh, I am quite surprised by `git am -3`.  I've tried it just now, and
> it's amazing.  I certainly must apologize.  :-)
> 
> I tried it yesterday, but the patches were from a different repo, and
> not available locally, so it really didn't do anything.  But now I tried
> it within the same repo, and it's really nice!
> 
>>> As far as the implementation goes I haven't thought too deeply but I
>>> suspect we'd want to create a couple of trees based on the commit we
>>> want to cherry-pick and its parent filtered by the pathspec and use
>>> those in the tree-way merge with HEAD.
>>
>> If we were to use the ort machinery, it may not be too bad to use
>> the pathspec only at the final writeout phase.

That would be tempting if cherry-pick didn't have a "--strategy" option. 
I think we need something more generic to cope with that.

>>  That is, perform a
>> full cherry-pick in the in-core index, reset all the entries in the
>> in-core index back to HEAD that are outside the given pathspec, and
>> then write out the result to the working tree.  That way, an old
>> change that was made to paths at the original location can be cherry
>> picked to a much newer tree after these paths have been moved to a
>> different location.  Doing the same with the recursive machinery
>> would be missier but perhaps the more recent merge-tree that uses
>> the ort machinery to work purely in-core might be a good way to go.

My hope was that the changes required to create a couple of new trees 
that are then used instead of the original commits when a pathspec is 
given would be fairly localized.

> I didn't understand the last part well, but I guess I may do when I
> start researching into it.  :)
> 
> Apart from <builtin/revert.c>, do you recommend I look into any
> particular files?

sequencer.c. If we go for the "write new trees and use those in the 
merge" approach then we'd need to change do_pick_commit() to create the 
trees and we'd probably want to change do_recursive_merge() to take 
trees rather than commits. We'd also need to add a new pathspec member 
to struct replay_opts to pass the pathspec around.

Best Wishes

Phillip