patchcontrib/subtree bugfix: Can't `add` annotated tag

cmd_add_commit() is passed FETCH_HEAD by cmd_add_repository, which is then rev-parsed into an object ID. However, if the user is fetching a tag rather than a branch HEAD, such as by executing:

$ git subtree add -P oldGit https://github.com/git/git.git tags/v1.8.0

The object ID is a tag and is never peeled, and the git commit-tree call (line 561) slaps us in the face because it doesn't handle tag IDs.

Because peeling a committish ID doesn't do anything if it's already a commit, fix by peeling[1] the object ID before assigning it to $rev, as per the patch.

[*1*]: Via peel_committish(), from git:git-sh-setup.sh

Reported-by: Kevin Cagle <kcagle@micron.com>
Diagnosed-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: James Denholm <nod.helm@gmail.com>
---
NB: This bug doesn't surface when using --squash, as $rev is reassigned
to the squash commit via new_squash_commit before git commit-tree sees
it (though for simplicity, new_squash_commit now also sees the peeled
ID).

Also doesn't surface when using "git subtree merge", as git merge can handle tag objects.

On a side note, if merging a tag without --squash, git merge recognises that it's a tag and adds a note to the merge commit body. It may be worth mimicking this when using "subtree merge --squash" or "subtree add".

 contrib/subtree/git-subtree.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

James Denholm <nod.helm@gmail.com> writes:

The "rev" (not "revs") seems to be used by more things than the final commit-tree state. Are we losing some useful information by peeling it too early like this patch does? The reason why we stopped peeling when writing FETCH_HEAD was because we wanted to record the fact that we merged a tag (and use the GPG signature if found in it) when constructing the log message for the merge, and peeling the tag too early and recording the commit in FETCH_HEAD would make it impossible to do, and I am wondering if this change is making the same kind of mistake here.

I see that add_msg does not use anything useful from latest_new, so with the current state of the code, it does not make that much difference (except that it says "from commit '$latest_new'", and by peeling, the fact that the user wanted to use a tag is lost from the result).

Yes, and this change makes such a change harder to implement on top, I suspect.

Would it be sufficient to do

	git commit-tree $tree $headp -p "$rev^0"

in that "not squashing" codepath instead?

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

You're not wrong, actually, peeling at the last minute (or at least later) would be a better choice. I'd suggest that we aren't losing currently-useful information (as it'd be rare-if-ever that a user would look at a hash in their commit logs and think "Oh, that's that tag!"), but certainly with future development in mind it's more ideal.

Yeah, that might be a worthy thing to porcelain-up in the future with logging the tag name rather than, or in addition to, the hash, as well as a similar change in add_squashed_msg.

On line 561, sure. Do you want me to do a re-roll?

On Fri, May 09, 2014 at 05:36:15PM +1000, James Denholm wrote:

Sorry to bump, but do you want a reroll on this?

--- Regards, James Denholm.