Re: [PATCH] docs: Clarify what git-rebase's "--preserve-merges" does

From: Junio C Hamano <gitster@pobox.com>
Date: Mar 31, 2015, 17:04 UTC
Message-ID: <xmqq8uedrtdb.fsf@gitster.dls.corp.google.com>
In-Reply-To: <xmqqh9t1rv0r.fsf@gitster.dls.corp.google.com>

Junio C Hamano <gitster@pobox.com> writes:

Show 27 quoted lines

> I wonder if it will be the right way to get a correct result to
> apply the difference to go from B to Z on top of an old commit when
> you are side-porting.
>
> Imagine you want to backport the same X-Y history by redoing the
> merge Z on top of another child of O (i.e. A's sibling).  That is,
> you start from this:
>
>
>          X---Y
>         /     \
>    O---A---B---Z---
>     \
>      M---N
>
> and would want to create this:
>     
>     
>    O           X'--Y'
>     \         /     \    
>      M---N---A'--B'--Z'--
>
> As long as everything down to the merge-base of the parents of the
> original merge (in this example, merge-base across Y and B that are
> Z's parents, which is A) is being transplanted, "apply the
> difference going from B to Z, on top of B', to obtain Z'" should
> work, I would think.

And just after I send the message because I needed to catch a bus, I notice that there is a problem.

Actually, "replay diff going from B to Z instead of merging" must be
done very carefully.  Imagine when Y in the original history were a
cherry-pick of M.  What you would be creating would look more like
this instead:
     
    O           X'--.
     \         /     \    
      M---N---A'--B'--Z'--

because Y' becomes a no-op, as the transplanted history already has M applied. But the original "diff going from B to Z" has the effect of M already in there. You would end up adding the same hunk twice without noticing. You somehow need to come up with a way to deal with this.

If you did a real merge between X' and B' to recreate Z', you would not have such a problem.

One way to be careful when recreating Z' out of Z might be:

    - Retry a merge between the original B and Y, with conflict
      markers intact;

    - Compare the result with what is recorded in Z.  The
      differences are textual conflict resolution and evil merge
      changes;

    - Now try a merge between B' and Y', with conflict markers
      intact;

    - Apply the difference you obtained in the second step to the
      result of the third step.

which is essentially the same as what rerere does.