Re: 'git pull' complains that a locally resurrected directory would be overwritten by merge when no pulled changes are affecting that directory

On Thu, Aug 26, 2021 at 6:05 PM Jeff King [off-list ref] wrote:

On Wed, Aug 25, 2021 at 02:19:15PM -0700, Junio C Hamano wrote:

quoted

Elijah Newren [off-list ref] writes:

quoted

... allow if the incoming changes don't touch the same files), but
local *staged* changes.  As per the merge manpage:

"""
To avoid recording unrelated changes in the merge commit, git pull and
git merge will also abort if there are any changes registered in the
index relative to the HEAD commit.
"""

While this particular example could theoretically be handled by the
merge machinery without requiring the index match HEAD,...

While I do not mind seeing a patch that loosens the condition ONLY
when the merge will cleanly auto-resolve without end-user
interaction, when any paths conflict and require editing by the
end-user, it is pretty much essential to require that the index
matches HEAD to keep "git merge" usable.

This is because the final step to conclude such an "automated
procedure cannot cleanly resolve, so the end user helps resolving
with the editor and mark the resolved paths with 'git add' or 'git
rm'" session will become very error prone if we did not have the
requirement.  Not just the user MUST remember not to use "commit -a"
or "git add" a path that was already dirty in the working tree
before the merge started (which is the consequence of the current
requirement, which allows local changes to the unrelated working
tree files), they must MUST remember to somehow EXCLUDE the changes
already registered for unrelted paths from the concluded merge.

Good point.

In theory we could reject the merge only after finding that there were
conflicts. That lets the happy path continue even with unrelated changes
(just like a fast-forward does). I suspect we'd need to change the
interface to the merge-backends, though (to say "do the merge, and bail
on conflicts, but _don't_ write out conflict markers or touch any other
state in that case").

We actually wouldn't need to change the interface.  One of the reasons
there were so many annoying bugs with index != HEAD, was that
builtin/merge.c stated that the merge backends were responsible for
enforcing that condition.  So every single merge backend had to be
individually fixed -- ours, octopus, resolve, recursive.  And then
some of those, like recursive, had multiple affected code paths, and
each one had to be fixed, and I didn't catch them all the first time.

That doesn't mean I'm in favor of the change, just pointing out that
one backend could decide to do this on its own even if the other
backends didn't support it.

(I am just thinking out loud, though. My personal opinion is that if you
have a bunch of staged changes and want to do any non-trivial merging,
you should considering committing or stashing those changes).

Agreed.

Though perhaps I should have pointed out to Yuri that he wouldn't have
run into these problems if he would have used
    git restore --source={hash}~1 -- math/polymake
instead of
    git checkout {hash}~1 -- math/polymake

(since the former defaults to updating just the working directory,
while the later updates both the index and the working directory.)