Re: [PATCH 0/2] Custom low-level merge driver support.

Linus Torvalds [off-list ref] writes:

On Wed, 18 Apr 2007, Junio C Hamano wrote:

quoted

While I agree with all of the above, I am currently fighting
with a back-to-drawing-board design problem.

The series made the low-level 3-way merge machinery
customizable, which is a progress, but I think merge-recursive
needs a hook to affect middle-level merge decision, similar to
what git-merge-one-file does.

Hmm. I think that it's not so much a "hook", as a "recursive merge is a 
separate merge policy altogether, that just *defaults* to the same final 
merge policy".

I suspect I did not state my problem clearly.

For example, for a three-way merge, it makes sense to do the three-way 
merge for the "internal" merge too. But it might be that that isn't true 
for all strategies.

Yes, that is what I alluded to in one of my previous message,
but I think the "foo-recursive" idea is a good one.  I was
planning to give only "3-way" or "pick ancestor" choice to
external low-level merge drivers.

But I think that is a separate issue.

The logic I said I wanted to override in the message you are
responding to is the big toplevel if () else if () ... chain in
process_entry(), and the nested if () else if () ... logic in
merge_file().  They roughly implement what git-merge-one-file
does.

Currently low-level 3-way merge drivers are invoked only in some
cases, exactly where git-merge-one-file used to invoke "RCS
merge".  The specific case of RelNotes symlink I wanted to
override is one of the cases that ll_merge() is not even
invoked.

On Wed, 18 Apr 2007, Junio C Hamano wrote:

But I think that is a separate issue.

The logic I said I wanted to override in the message you are
responding to is the big toplevel if () else if () ... chain in
process_entry(), and the nested if () else if () ... logic in
merge_file().  They roughly implement what git-merge-one-file
does.

Ahh. 

Ok, in that case, I kind of disagree with you. I think the hardcoded 
merging rules (ie when two of the SHA1's match etc) are the only sane 
things to do _regardless_ of any merge strategy.

So yeah, we'll always do one level of merging at the pure SHA1 stage, and 
pick that without actually doing any file-level merging at all. I don't 
think that's wrong. Maybe it means that people cannot pick some really 
strange merge that they want on a file-by-file basis, but I think that's 
simply how "merge-recursive" is defined.

IOW, I don't think you should be able to turn

	git merge -s recursive

into

	git merge -s ours

by making attributes say something like

	*: merge=ours

because I think the attributes are really about what we do WHEN we hit 
file-level conflicts, while the "merge strategy" is a much higher-level 
thing. They are independent.

But maybe there is some real-world and sane usage that shows me wrong. 

		Linus