Re: [PATCH v3 2/3] merge: Add merge.renames config setting

From: Eckhard Maaß <eckhard.s.maass@googlemail.com>
Date: Apr 30, 2018, 08:03 UTC
Message-ID: <20180430080341.GA28348@esm>
In-Reply-To: <CABPp-BHwM1jx2+VTxt7hga7v-E6gvHuxVNPqm-MPRXYe5CDVtA@mail.gmail.com>

On Fri, Apr 27, 2018 at 01:23:20PM -0700, Elijah Newren wrote:

> I doubt it has ever been discussed before this thread.  But, if you're
> curious, I'll try to dump a few thoughts.

Thank you, I try to dump some of mine, too. Maybe let me first stress that for me copy detection without --find-copies-harder is much more a "find content extracted" (like methods being factored out). In a way this is nearer to a rename than to a real copy.

Show 8 quoted lines

> [...] Let's say we have branches
> A and B, and:
>    A: modifies file z
>    B: copies z to y
> 
> Should the modifications to z done in A propagate to both z and y?  If
> not, what good is copy detection?  If so, then there are several
> ramifications...

If one just assumes the most likely outcome is that something from z wad factored out to y, it might just be sufficient to see whether the modifications of the two branches apply cleanly - if A touched the parts of B that have been factored out there would be a normal merge conflict (where one could be nice and give a hint that some content was copied to y on the B branch), if A did not touched the parts touched (or moved) by B, then there is no problem. If A exactly deleted the content moved by B, there will be no conflict - but this is seems to be strange anyway.

I admit that a "real" copy would get unnoticed that way. But the semantics of such a copy isn't too clear for me either - did I copy the other part to make it independent of the other or did I just employ a copy and paste tactic? The former does not want the changes, the later does. But I am happy catering to the former here.

To sum up:
- fail as before for conflicting merges, but give a hint that one has
  copied to quicken up resolution.

> - If B not only copied z but also first modified it, then do we have
>   potential conflicts with both z and y -- possibly the exact same
>   conflicts, making the user resolve them repeatedly?

With the above suggestion, if there are conflicts, you fail and give a hint.

> - What if A copied z to x?  Do changes to z propagate to all three of
>   z and x and y?  Do changes to either x or y affect z?  Do they
>   affect each other?

A copy on branch to x and one another to y seems strange even if z merges cleanly. Did both sides try to factor the same thing out to different files? Or did they try to make something independent, but managed to make it to different files? For this I would be inclined to just suggest fail with a copy/copy(somewhere else). But this is a real corner case after all. Has anyone seen just thing in practice?

> - If A deleted z, does that give us a copy/delete conflict for y?  Do
>   we also have to worry about copy/add conflicts?  copy/add/delete?
>   rename/copy (multiple variants)?  copy/copy?

We do have the modified/deleted conflict where we could hint that content also has been copied and then not try to do more.

Show 5 quoted lines

> - Extra degrees of freedom may mean new conflict types:
> 
>   - The extra degrees of freedom from renames introduced multiple new
>     conflict types (e.g. rename/add, rename/rename(1to2),
>     rename/rename(2to1)).

For renaming one side and coping the other, I would think doing the same as above is sensible enough: if there are conflicts one can give an additional hint of the one part having been copied, but not change the kind of conflicts much.

> The more I think about it, the more I think that attempting to detect
> copies in a merge algorithm just doesn't make sense.  Anything I can
> think of that someone might attempt to use detected copies for would
> just surprise users in a bad way...

Hm, it didn't sound like that. Would you think that users would be surprised by my suggestions? Or are they all too corner casey to be worth implementing anyway?

Greetings, Eckhard