The Universe of Disco

Reordering git commits with git-commit-tree

I know, you want to say “Why didn't you just use git-rebase?” Because git-rebase wouldn't work here, that's why. Let me back up.

Say I have commit A, in which feature X does not exist yet. Then in commit C, I implement feature X.

But I realize what I really wanted was to have A, then B, in which feature X was implemented but disabled, and then C in which feature X was enabled. The C I want is just like the C that I have, but I don't have the intervening B.

I have:

  no X              X on

    A --------------- C

I want:

  no X     X off    X on

    A ------ B ------ C

One way to do this is to use git-rebase in edit mode to split C into B and C. To do this I would pause while rebasing C, edit C to disable feature X, commit the result, which is B, then undo the previous edits to re-enable X, and continue the rebase, creating C. That's two sets of edits. I could backup the files before the first edit and then copy them back for the second edit, but that's the SVN way, so I'm not going to do that.

Now someone wants me to use git-rebase to “reorder the commits”. Their idea is: I have C. Edit C to disable feature X and commit the result as B':

  no X     X on     X off

    A ------ C ------ B'

Now use interactive git-rebase to reorder B and C. But this will not work. git-rebase will construct a patch for turning C into B' and will try to apply it to A. This will fail completely, because a patch for turning C into B' is a patch for turning off feature X once it is implemented. Feature X is not in A and you can't turn something off that isn't there. So the rebase will fail to apply the patch.

What I did instead was rather bizarre, using a plumbing command, but worked well. I wrote the code to disable X, and committed it as B, obtaining this:

  no X     X on     X off

    A ------ C ------ B

Now B and C have the files I want in them, but their parents are wrong. That is, the history is in the wrong order, but if the parent of C was B and the parent of B was A, eveything would be perfect.

But we can't just change the parents; we have to create a new commit, say B', which has the same files as B but whose parent is A instead of C, and we have to create a new commit C' which has the same files as C but whose parent is B' instead of A.

This is what git-commit-tree does. You give it a tree object containing the files you want, a list of parents, and a commit message, and it creates the commit you asked for and prints its SHA1.

When we use git-commit, it first turns the index into a tree, with git-write-tree, then creates the commit, with git-commit-tree, and then moves the current head ref up to the new commit. Here we will use git-commit-tree directly.

So I did:

       % git checkout -b XX A
       Switched to a new branch 'XX'
       % git commit-tree -p HEAD B^{tree}
       10ddf433039fd3cbc5bec0c64970a45add15482e
       % git reset --hard 10ddf433039fd3cbc5bec0c64970a45add15482e
       % git commit-tree -p HEAD C^{tree}
       ce46beb90d4aa4e2c9fe0e2e3d22eea256edceac
       % git reset --hard ce46beb90d4aa4e2c9fe0e2e3d22eea256edceac

The first git-commit-tree

   % git commit-tree -p HEAD B^{tree}

says to make a commit whose tree is the same as B's, and whose parent is the current HEAD, which is A. (B^{tree} is a special notation that means to get the tree from commit B.) Git pauses here to read the commit message from standard input (not shown), and prints the SHA of the new commit on the terminal. I then use git-reset to move the current head ref, XX, up to the new commit. Normally git-commit would do this for us, but we're not using git-commit today.

Then I do the same thing with C:

   % git commit-tree -p HEAD C^{tree}

makes a new commit whose tree is the same as C's, and whose parent is the current head, which looks just like B. Again it reads a commit message from standard input, and prints the SHA of the new commit on the terminal, and again I use git-reset to move XX up to the new commit.

Now I have what I want and I only had to edit the files once. To complete the task I just reset the head of my working branch to wherever XX is now, discarding the old A-C-B branch in favor of the new A-B-C branch. If there's an easier way to do this, I don't know it.

It seems to me that there have been a number of times in the past when I wanted to do something like reordering commits, and git-rebase did not do what I wanted because it reorders patches and not commits. I should keep my eyes open, and see if this comes up again, and if it is worth automating.

[ Thanks to Jeremy Leader for suggesting I write this up and to Jeremy Leader and Rik Signes for advance editing. ]

[ Addendum 20150813: a followup article ]

[ Addendum 20200531: a better way to accomplish the same thing ]

[Other articles in category /prog] permanent link