Re: Changing branches in supermodule does not affect submodule?

Peter Krefting <peter@softwolves.pp.se> writes:

My understanding is that the primary reason "git checkout $another_branch" does not touch submodules (other than updating the commit object name bound in the index in the superproject) is because we did not know what people would want to happen when submodule support was introduced.

Now people have used submodules for a while, we can update this. I think modelling the behaviour after how normal blobs are updated upon branch switching would make sense.

So let's step back a bit and review what happens to a normal blob upon branch switching. There are three cases for a path while switching branches from A to B:

 #1 The path is the same between A and B.  In this case, whatever you have
    for path in the index and in the work tree stay the same.  Your local
    modification is carried across branch switching.

    This is already what we do with submodules.  In cases #2 and #3 below,
    path is different between A and B.

 #2 You do not have any change to the path in either the index nor in the
    work tree.  The path is updated to B's version in the index and also
    what you have in the work tree matches it.

    Note that if B lacks the path, it is _removed_ from both the index and
    the work tree.

 #3 You have a change to the path in the index or in the work tree.  The
    branch switch is refused if the index does not match what is in B.

    I think this also is already what we do with submodules.

With a reasonable definition of "have a change in the work tree" for a submodule path, we can define a reasonable behaviour for updating the submodules when a different branch is checked out in the superproject.

For a regular blob, having a change in the work tree is that the object recorded in HEAD (in the superproject) is different from what you have in the work tree. What should the definition be for a submodule?

The HEAD (in the submodule), the index (in the submodule) and the work tree (in the submodule) form the "state" of what you have in the path, and that corresponds to the "work tree state". So I think we can say that there is no change in the work tree for a submodule at path P if all of the following holds:

    - the HEAD in the submodule matches the commit recorded for the path P
      in the HEAD in the superproject;

    - the index in the submodule matches the HEAD in the submodule; and

    - the work tree in the submodule matches the index in the submodule.

With this definition, you should be able to patch "git checkout" to recurse into the submodule directory for case #2.

As to the implementation, I'd suggest:

 - the check between the submodule HEAD and what the superproject records
   in the HEAD (the first condition above) be done inside the main process
   (because we already have a working check for #3 to refuse branch
   switching, I think it would be easier to implement this check by
   comparison between the submodule HEAD and the superproject index).

 - the check between the index and the HEAD in the submodule and the check
   between the index and the work tree in the submodule (the second and
   the third conditions above) be done in a subprocess that runs something
   like "git status" via the run_command() interface; and

 - the actual recursive checkout to happen in a subprocess that runs
   another instance of "git checkout" via the run_command() interface.

There is one caveat. The potential removal of the work tree path in #2 may not be desirable for a submodule path, even if you do not have any change to it (i.e. both the index and the work tree in the submodule match the HEAD in the submodule), as you may have untracked files you would rather want to keep in the submodule work tree.