Re: mergetool: what to do about deleting precious files?

From: "Junio C Hamano" <redacted>

"Philip Oakley" [off-list ref] writes:

quoted

quoted

So I do not think this is not limited to "new file".  Anything that
a tree-level three-way merge would resolve cleanly without having to
consult the content-level three-way merge will complete without
consulting the merge.ours.driver; per-file content-level three-way
merge driver (which is what merge=<drivername> mechanism lets you
specify via the attributes mechanism) is not something you would
want to use for this kind of thing.  It is purely for resolving the
actual content-level conflicts.

That (that Git knows best) sounds just wrong.

Don't twist my words.  I never said Git knows best.

The part I was responding to was "would resolve cleanly without having to
consult the content-level three-way merge will complete without
consulting the merge.ours.driver".

It was that lack of consultation (by git) of the putative merge-driver that 
was being noted.

The general misunderstanding, as I now see it, is the (false) expectation 
that a merge-driver would do the whole merge process.

It took a bit of digging through the documentation for me to find out just 
what the merge process appears to be. I'm sure that it obvious to those who 
have worked with git from the beginning and the previous patch flow process, 
but the merge process wasn't obvious to me, and various blogs and SO Q&A on 
the issue suggest the same for many others.

If I now understand correctly, the merge process flow is:

* canonicalise content (eol, smudge-clean, $id, renormalise, etc)
* diff the content (internal, or GIT_EXTERNAL_DIFF)
* apply the diff
* if conflicts, only then use merge-driver/tool

Would that be a correct interpretation?

The user-level merge driver is a mechanism to affect conflict level
three-way merges.  The interface to the content level three-way
merge driver feeds three versions of blobs and the driver is
expected to give a merged result.  The interface as designed is
incapable of passing "here is the common ancestor", "our side is
missing" and "their side is this content".

So if we want a mechanism that can affect the outcome of tree-level
three-way merge, we need a _new_ mechanism.  The existing merge
drivers that are written by end users (at least the ones written
correctly to the spec, anyway) are not expecting to be called with
"in our tree, there is no blob here", and trying to piggyback on it
will break existing users.

Is an alternative to use the GIT_EXTERNAL_DIFF to create a nul diff, so no 
changes are applied (precious/sensitive file is left behind)? This would 
have no conflicts and no requirement for a merge-conflict driver.

--

Philip