[PATCH/RFC] Allow empty commits during rebase -i

[PATCH/RFC] Allow empty commits during rebase -i

[Pete Harlan]

If you squash two commits into a previous commit, where the first
squash reverts the previous commit and the second redoes the change
correctly, rebase -i would fail during the first squash because it
generates an empty commit.  This patch allows the rebase to succeed.

This also introduces the possibility that you might accidentally
create an empty commit with a squash, but I expect that will happen
less often than the scenario this is intended to address.

Signed-off-by: Pete Harlan <pgit@pcharlan.com>
---

This arose for me recently; I used "git revert" to undo a commit
several changes back, and then reworked and committed anew.  The first
commit that I was redoing had a thorough commit message, while my new
commit had a message like "do it right this time".  I squashed the
three commits into one with rebase -i, but git choked on the
intermediate empty commit.

I could have simply removed the first two commits I was squashing (the
initial version and its revert), but then would have lost the
well-written commit message that went with the first version.

I imagine an ideal version of this fix would make it so the use case I
presented here would work, but rebase -i would still prevent
introducing a new empty commit, or at least warn when it was
introducing one.  In the absence of that ideal fix, I think this
behavior is better than failing to handle this case.

 git-rebase--interactive.sh    |    2 +-
 t/t3404-rebase-interactive.sh |    9 +++++++++
 2 files changed, 10 insertions(+), 1 deletions(-)

diff --git a/git-rebase--interactive.sh b/git-rebase--interactive.sh
index 1560e84..81db5cf 100755
--- a/git-rebase--interactive.sh
+++ b/git-rebase--interactive.sh
@@ -403,7 +403,7 @@ do_next () {
 			GIT_AUTHOR_NAME="$GIT_AUTHOR_NAME" \
 			GIT_AUTHOR_EMAIL="$GIT_AUTHOR_EMAIL" \
 			GIT_AUTHOR_DATE="$GIT_AUTHOR_DATE" \
-			$USE_OUTPUT git commit --no-verify \
+			$USE_OUTPUT git commit --no-verify --allow-empty \
 				$MSG_OPT "$EDIT_OR_FILE" || failed=t
 		fi
 		if test $failed = t
diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
index 3a37793..5eb9f7e 100755
--- a/t/t3404-rebase-interactive.sh
+++ b/t/t3404-rebase-interactive.sh
@@ -484,4 +484,13 @@ test_expect_success 'reword' '
 	git show HEAD~2 | grep "C changed"
 '

+test_expect_success 'squash including empty' '
+	test_commit Initial_emptysquash emptysquash abc &&
+	test_commit first_mod emptysquash abd &&
+	test_tick &&
+	git revert --no-edit HEAD &&
+	test_commit second_mod emptysquash abe &&
+	FAKE_LINES="1 squash 2 squash 3" git rebase -i Initial_emptysquash
+'
+
 test_done
-- 
1.6.6.196.g1f735

Re: [PATCH/RFC] Allow empty commits during rebase -i

[Junio C Hamano]

Pete Harlan <pgit@pcharlan.com> writes:

> I imagine an ideal version of this fix would make it so the use case I
> presented here would work, but rebase -i would still prevent
> introducing a new empty commit, or at least warn when it was
> introducing one.  In the absence of that ideal fix, I think this
> behavior is better than failing to handle this case.

Sorry, I actually tend to think that in the absense of that fix, your
version introduces risky behaviour that only a corner-case use case
benefits, and pros-and-cons doesn't look attractive enough.

Why not do something like:

    pick X a crap tree with a good message
    pick Y revert X
    pick Z a good tree with a crap message

-->

    # drop X
    # drop Y
    edit Z

and then run "git commit --amend -C X" when it is Z's turn to be
processed?

Re: [PATCH/RFC] Allow empty commits during rebase -i

[Pete Harlan]

On 01/17/2010 05:29 PM, Junio C Hamano wrote:
> Pete Harlan <pgit@pcharlan.com> writes:
> 
>> I imagine an ideal version of this fix would make it so the use case I
>> presented here would work, but rebase -i would still prevent
>> introducing a new empty commit, or at least warn when it was
>> introducing one.  In the absence of that ideal fix, I think this
>> behavior is better than failing to handle this case.
> 
> Sorry, I actually tend to think that in the absense of that fix, your
> version introduces risky behaviour that only a corner-case use case
> benefits, and pros-and-cons doesn't look attractive enough.
> 
> Why not do something like:
> 
>     pick X a crap tree with a good message
>     pick Y revert X
>     pick Z a good tree with a crap message
> 
> -->
> 
>     # drop X
>     # drop Y
>     edit Z
> 
> and then run "git commit --amend -C X" when it is Z's turn to be
> processed?

That is another way to accomplish the same thing, but doesn't prevent
the current behavior from being confusing.

Part of the problem is that with the current behavior the user is sent
to the command line with:

  # Not currently on any branch.
  nothing to commit (working directory clean)

  Could not apply a0b17c5... Revert "Crap tree good message"

with HEAD pointed to X^.  Unsure of how to proceed from here, I
--aborted the rebase and copy/pasted the commit message I wanted and
resolved to track this down and fix it when I got a chance.

As it happens, "git rebase --continue" does exactly what I would have
wanted to happen, including putting me in an editor with all three
commit messages and succeeding when I exit the editor.  But without a
better message from git I don't expect a user to discover that.  And,
when rebase can continue just by being told so it would be nice if it
didn't require that user intervention.

If the introduction of empty commits that the user has asked for
(perhaps inadvertently) is considered too undesirable, then perhaps my
fix is too simple.  I'll think about how to do something more
sophisticated.

Thanks for your feedback,

--Pete

Re: [PATCH/RFC] Allow empty commits during rebase -i

[Junio C Hamano]

Pete Harlan <pgit@pcharlan.com> writes:

> ..., "git rebase --continue" does exactly what I would have
> wanted to happen, including putting me in an editor with all three
> commit messages and succeeding when I exit the editor.  But without a
> better message from git I don't expect a user to discover that.

There seems to be an idea for a good improvement ;-)  CC'ing Michael as he
has been most active in this area for the past few weeks.

Re: [PATCH/RFC] Allow empty commits during rebase -i

[Johannes Schindelin]

Hi,

On Sun, 17 Jan 2010, Pete Harlan wrote:

> If the introduction of empty commits that the user has asked for 
> (perhaps inadvertently) is considered too undesirable, then perhaps my 
> fix is too simple.  I'll think about how to do something more 
> sophisticated.

How about something less sophisticated instead?  Namely, check for the 
condition that nothing was changed, and tell the user that the commit 
blablabla seems to introduce changes that are already present in HEAD.  
Maybe even mention that this may be due to the commit being applied 
already and saying that --continue is safe in that case, but please check.

Hmm?

Ciao,
Dscho