Merging changes from branches where file paths have changed

Merging changes from branches where file paths have changed

[Thomas Adam]

Hi all,

I've got two git repositories -- for the purposes of illustration,
I'll call them RepoA, and RepoB.  RepoA is a large repository
originally from a CVS conversion.  Repo B is a repository which has
been formed from git-filter-branch (should it matter) from RepoA.

In doing the conversion -- RepoB contains the same history as in
RepoA, except that RepoB is now a limited subset in terms of its files
that RepoA has (repoB contains commits out of repoA for a specific
subdirectory).   However, the problem I'm wanting to solve now, which
is causing me to scratch my head and go "hmm", is how I go about
merging commits from RepoA and have them in RepoB -- where the commits
from repoA are in the same *directory* as the one which formed RepoB.

In RepoA, for the directory (and all its sub-directories) which is
contained in RepoB, I've got a list of commits -- roughly like the
following:

% cd ./distrib/build/sources/coretree/tree
% git log --oneline master..my/topic/branch -- .

Now in RepoB, I've got all of the directories and files from
./distrib/build/sources/coretree/tree/ in the root, having obtained
this from git-filter-branch from RepoA.  Despite the change of file
locations, the history of the files between the two repositories are
the same.

The commits though coming from "git log" in RepoA touch files in
distrib/build/sources/coretree/tree, yet the files to which I want the
changes for in RepoB are not in "distrib/build/sources/coretree/tree"
-- there all in "./" at the top-level directory for RepoB.  The files
in both once shared the same history -- it's just that now the
locations of those files in RepoB which have changed.

I want to cherry-pick the commits I obtained from the "git log"
command in RepoA, and have them contained in RepoB.  So I thought
something like this would work:

[ In RepoB.... ]:
% git remote add repoA file:///path/to/repoA/
% git fetch repoA
% git checkout specific/branch
% git cherry-pick $SOME_COMMIT_FROM_REPOA

However, this won't work, and doesn't.  Not only do the two
repositories not have any common ancestry, but even if they did, the
file paths from RepoA and RepoB do not match for the commit(s) I would
want to cherry-pick.  As far as I can tell, were I able to convince
Git the file paths did work, then I could cherry-pick them.

Can anyone suggest a good way round this?  I hope I've managed to
explain this adequately; just ask if not.

Thanks in Advance.

-- Thomas Adam

Re: Merging changes from branches where file paths have changed

[Johannes Sixt]

Am 14.04.2012 15:15, schrieb Thomas Adam:
> The commits though coming from "git log" in RepoA touch files in
> distrib/build/sources/coretree/tree, yet the files to which I want the
> changes for in RepoB are not in "distrib/build/sources/coretree/tree"
> -- there all in "./" at the top-level directory for RepoB.  The files
> in both once shared the same history -- it's just that now the
> locations of those files in RepoB which have changed.
> 
> I want to cherry-pick the commits I obtained from the "git log"
> command in RepoA, and have them contained in RepoB.  So I thought
> something like this would work:
> 
> [ In RepoB.... ]:
> % git remote add repoA file:///path/to/repoA/
> % git fetch repoA
> % git checkout specific/branch
> % git cherry-pick $SOME_COMMIT_FROM_REPOA

You have these histories


  A--o--B--C--o--o--D--o--E   <- RepoA

  A'----B'-C'                 <- RepoB

and you want to cherry-pick D and E to RepoB. Right?

I suggest that you install a graft in RepoB, then cherry-pick D and E:

  echo $(git rev-parse C' C) >> .git/info/grafts
  git cherry-pick D
  git cherry-pick E

Then it will look as if everything outside the directory was removed,
and everything else was moved from the directory to the root with 100%
similarity, which makes the merge process easy. (This works only if
there are no files with the same names in C's and C''s root; a Makefile
in each would pose a problem.)

(Remove the graft when you're done.)

-- Hannes

Re: Merging changes from branches where file paths have changed

[Thomas Adam]

Hello,

On 14 April 2012 20:21, Johannes Sixt <j6t@kdbg.org> wrote:
> You have these histories
>
>
>  A--o--B--C--o--o--D--o--E   <- RepoA
>
>  A'----B'-C'                 <- RepoB
>
> and you want to cherry-pick D and E to RepoB. Right?

Yes, that's about right.

> I suggest that you install a graft in RepoB, then cherry-pick D and E:
>
>  echo $(git rev-parse C' C) >> .git/info/grafts

Can you elaborate on this a little more please?  Or point me towards
some documentation.  Note that due to the way git-filter-branch works,
although the history between RepoA and RepoB are common, the commit
IDs between the two are completely different; presumably
git-filter-branch uses patch-id at some level to reform commits.
Nevertheless, if I'm understanding superficially what this graft is
supposed to do, I do not see how getting a list of commits from C' to
C is going to help in repoB when neither RepoA or RepoB have any
ancestry in common?

Kindly,

-- Thomas Adam

Re: Merging changes from branches where file paths have changed

[Johannes Sixt]

Am 14.04.2012 21:36, schrieb Thomas Adam:
> Hello,
> 
> On 14 April 2012 20:21, Johannes Sixt <j6t@kdbg.org> wrote:
>> You have these histories
>>
>>
>>  A--o--B--C--o--o--D--o--E   <- RepoA
>>
>>  A'----B'-C'                 <- RepoB
>>
>> and you want to cherry-pick D and E to RepoB. Right?
> 
> Yes, that's about right.
> 
>> I suggest that you install a graft in RepoB, then cherry-pick D and E:
>>
>>  echo $(git rev-parse C' C) >> .git/info/grafts
> 
> Can you elaborate on this a little more please?  Or point me towards
> some documentation.  Note that due to the way git-filter-branch works,
> although the history between RepoA and RepoB are common, the commit
> IDs between the two are completely different; presumably
> git-filter-branch uses patch-id at some level to reform commits.
> Nevertheless, if I'm understanding superficially what this graft is
> supposed to do, I do not see how getting a list of commits from C' to
> C is going to help in repoB when neither RepoA or RepoB have any
> ancestry in common?

There is no "list of commits from C' to C". Look carefully: it's
rev-parse, not rev-list. The command above only states that the (only)
parent of C' is C, i.e., it fakes this history:

  A--o--B--C--o--o--D--o--E   <- RepoA
            \
             C'               <- RepoB

and then the two repositories have history in common.

And BTW:

1. C' must be the top-most commit that was derived from RepoA.

2. If D or E add new paths, they will not be moved to the root by the
cherry-pick. Be sure to check the result.

3. If you have additional commits in RepoB that were not derived from
RepoA, it may not work out as smoothly, if the changes are so large that
rename detection does not kick in anymore.

-- Hannes