BUG: git subtree split - revert commit followed by a merge

Hello:

I'm running into the problem described in this mailing list post:
	http://thread.gmane.org/gmane.comp.version-control.git/202645

 'git subtree split' fails to reconstruct the history when a revert commit is followed by a merge commit. I have slightly adjusted the test script provided by Fabien in his mailing list post:

    git init
    
    # create a directory that is going to be split
    mkdir doc
    echo "TEST" > doc/README
    git add doc
    # commit A
    git commit -a -m"first version"
    
    # create a branch with a new commit (Z)
    git checkout -b test
    echo "TEST" > doc/README1
    git add doc/README1
    git commit -a -m"added README1"
    git checkout master
    
    # modify the README file (commit B)
    echo "TEST_" > doc/README
    git commit -a -m"second version"
    
    # revert the change (commit C)
    echo "TEST" > doc/README
    git commit -a -m"revert second version"
    # or use git revert HEAD^
    
    # split
    git subtree split --prefix="doc" --branch=TARGET
    
    # add another commit (to a file *not* in the subtree dir)
    echo "BLA" > BLA
    git add BLA
    git commit -a -m"third version"
    
    # adding another commit to a file in the subtree dir will "fix" things
    #echo "MEH" > doc/MEH
    #git add doc
    #git commit -a -m"fourth version"

    # the log will show the 3 commits as expected (including B and C)
    GIT_PAGER= git log --oneline TARGET
    
    # merge the test branch
    git merge -m"merged test" test
    
    # attempt to re-split; this will fail
    git subtree split --prefix="doc" --branch=TARGET
    
    # see what history split generates
    git subtree split --prefix="doc" --branch=TARGET2

I have discovered that if the revert commit is followed by another commit that makes changes in the subtree directory, the split will work as expected (see "fourth version" above).

See also this related SO question where I ask for a workaround: http://stackoverflow.com/questions/18465867

Best regards, Brecht

Hi,

I think I understand what's going on. If there are no net changes in one of the two branches, the id of the merge commit will be the same as that of the last commit on the other branch (pass the -d option to 'git subtree split' to see the new commit id's of the split out commits). In that case, git-subtree will drop the merge commit. I believe this is done because for regular commits, this means that the commit does not make changes to files in the subtree directory. I'm not sure I'm right about this, as I would expect the merge commit's hash to depend on the hashes of its parent commits.

I was thinking the solution would be to simply adjust the copy_or_skip function to always copy merge commits. However, this may result in the produced commit history to differ from before (unnecessary merge commits were previously left out), leading to other problems.

A solution might be to check for the presence of such commit + revert-commit branches. This can be done by checking for each parent of a merge commit if it is the ancestor of any of the other parents (using git merge-base --is-ancestor).

Best regards, Brecht

On Tue, 27 Aug 2013 15:09:30 +0200, Machiels, Brecht <Brecht.Machiels@kla-tencor.com> wrote: