On Mon, 2017-12-11 at 12:27 -0800, Stefan Beller wrote:
> On Mon, Dec 4, 2017 at 1:39 PM, David Turner <[email protected]>
> wrote:
> > When merging with a submodule modify/delete conflict (i.e. I've
> > deleted
> > the submodule, and I'm merging in a branch that modified it), git
> > lies
> > about what it is doing:
> >
> > "CONFLICT (modify/delete): submodule deleted in HEAD and modified
> > in
> > submodules.
>
> Up to here the error message sounds correct, still?
Yep!
> > Version submodules of submodule left in tree at
> > submodule~submodules.
> > Automatic merge failed; fix conflicts and then commit the result."
>
> This sounds as if the code assumed to handle only files.
This assumption is unfortunately very common -- I just filed a PR to
fix an instance of this in libgit2.
> > In fact, the working tree does not contain anything named
> > 'submodule~submodules'.
> >
> > In addition, I would ordinarily resolve a conflict like this by
> > using
> > 'git rm'. Here, this gives a warning:
> >
> > $ git rm submodule
> > submodule: needs merge
>
> (Regarding submodule merges in general:)
>
> Uh. We cannot add merge markers to a submodule or such.
> More importantly we'd have to ask the question if the merge conflict
> is on the superproject level (Choose one of the commits of the
> submodule)
> or on the submodule level (perform a merge in the submodule between
> the
> two commits) or some hybrid approach thereof.
Yeah, that's a tricky thing in general. In this case, tho, the
submodule is being removed, so git rm should work.
> > rm 'submodule'
> > warning: Could not find section in .gitmodules where path=submodule
>
> The deletion of the submodule removed the .gitmodules entry, and the
> merge of the .gitmodules file presumably went fine. :/
Indeed.
> I assume we need a special merge driver for the .gitmodules file to
> keep
> the submodule around when it is in at least one side.
>
> > Git's behavior here is significantly better than liggit2's (which
> > tries
> > to check out 'submodule' as if it were a blob, and fails to do so),
> > but
> > it's still confusing.
> >
> > It's not clear to me what the correct behavior is here. Maybe it's
> > sufficient to just fix the message?
>
> I think the first step is to fix the message to reflect reality.
>
> As alluded to above, I don't know what the correct merge
> behavior is (and where to put 'conflict markers').
The only case we need 'conflict markers' is in the
{add,modify}/{add,modify} case (with different versions on each side).
The delete/* and */delete case can be handled more simply. We might
not want to do this until we can handle all cases -- but personally, I
think a half-solution is better than a non-solution.
