git rebase --preserve-merges seems to not work

[Joshua Jensen]

  In order to push changes to our master branch, one has to have pulled
first.  We pull with the --rebase option.  It works great except for
merge commits.

I discovered 'git rebase' has a --preserve-merges option.  I assumed it
would allow me to rebase my merge commit on top of others' changes
pushed to our central master branch.

Unfortunately, I can't seem to make it work.  Does anyone have thoughts
on what I'm doing wrong?

My session is below.

Thanks!

Josh


$ git init

$ echo a > a
$ git add a
$ git commit -m a
$ git tag a

$ echo b > b
$ git add b
$ git commit -m b

$ echo c > c
$ git add c
$ git commit -m c
$ git tag c

# Note the hierarchy is simple: a <- b <- c
$ git lg --all
* d132c87 - (c) c
* 4a88fd1 - b
* b576660 - (a) a


# Create the hierarchy: a <- d
$ git checkout a
Note: checking out 'a'.

$ echo d > d
$ git add d
$ git commit -m d
$ git tag d

$ git lg --all
* 6abf527 - (HEAD, d) d
| * d132c87 - (c) c
| * 4a88fd1 - b
|/
* b576660 - (a) a


# Wanting to merge my branch back to my local master to push to the server.
$ git checkout a
$ git merge --no-ff c
Merge made by recursive.
  b |    1 +
  c |    1 +
  2 files changed, 2 insertions(+), 0 deletions(-)
  create mode 100644 b
  create mode 100644 c
$ git tag bc-merge


# Commit 'd' is something I would pull --rebase from the server before I
push.
$ git lg --all
*   472fd93 - (HEAD, bc-merge) Merge commit 'c' into HEAD (3 seconds ago)
|\
| * d132c87 - (c) c (2 minutes ago)
| * 4a88fd1 - b (2 minutes ago)
|/
| * 6abf527 - (d) d (2 minutes ago)
|/
* b576660 - (a, master) a (3 minutes ago)


# Here is the rebase with the --preserve-merges option.
$ git rebase --preserve-merges d
Rebasing (n/3)
Successfully rebased and updated detached HEAD.

# Of course, bc-merge is an orphaned branch.  Its changes got replayed
on top of 'd'.
# Note the merge commit with the description "Merge commit 'c' into
HEAD" is missing.
$ git lg --all
* 2c93a8a - (HEAD) c (11 seconds ago)
* 2b98bc6 - b (11 seconds ago)
* 6abf527 - (d) d (3 minutes ago)
| *   472fd93 - (bc-merge) Merge commit 'c' into HEAD (74 seconds ago)
| |\
|/ /
| * d132c87 - (c) c (3 minutes ago)
| * 4a88fd1 - b (3 minutes ago)
|/
* b576660 - (a, master) a (4 minutes ago)

--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to [hidden email]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[Björn Steinbrink]

On 2010.09.08 23:17:39 -0600, Joshua Jensen wrote:

> # Commit 'd' is something I would pull --rebase from the server
> before I push.
> $ git lg --all
> *   472fd93 - (HEAD, bc-merge) Merge commit 'c' into HEAD (3 seconds ago)
> |\
> | * d132c87 - (c) c (2 minutes ago)
> | * 4a88fd1 - b (2 minutes ago)
> |/
> | * 6abf527 - (d) d (2 minutes ago)
> |/
> * b576660 - (a, master) a (3 minutes ago)
>
>
> # Here is the rebase with the --preserve-merges option.
> $ git rebase --preserve-merges d
> Rebasing (n/3)
> Successfully rebased and updated detached HEAD.
>
> # Of course, bc-merge is an orphaned branch.  Its changes got
> replayed on top of 'd'.

Yup, and the "problem" is, that both sides of the merge got replayed.

> # Note the merge commit with the description "Merge commit 'c' into
> HEAD" is missing.

Because the merge could be resolved as a fast-forward, as it merges the
replayed commits b' and c', not the originals.

> $ git lg --all
> * 2c93a8a - (HEAD) c (11 seconds ago)
> * 2b98bc6 - b (11 seconds ago)
> * 6abf527 - (d) d (3 minutes ago)
> | *   472fd93 - (bc-merge) Merge commit 'c' into HEAD (74 seconds ago)
> | |\
> |/ /
> | * d132c87 - (c) c (3 minutes ago)
> | * 4a88fd1 - b (3 minutes ago)
> |/
> * b576660 - (a, master) a (4 minutes ago)

What you actually wanted to replay is just the first-parent history,
redoing the merges with their respective original second parent.
Jonathan (Cc'ed) had the same problem about a month ago.

The "first parent" thing isn't that well defined, criss-cross merges
combined with some unfortunate fast-forwards will easily make the
first-parent history become the one that you didn't mean (ask any
git-svn users that dared to use "git merge" without fully understanding
how git-svn uses the history ;-)), but it works in this special case.
And with that special case in mind (and admittedly probably not thinking
much further) I came up with this patch a few years ago:

http://marc.info/?l=git&m=119379735525213&w=2

It was not my itch to scratch though, so I never got around to get it
included, and of course it doesn't apply cleanly anymore. I tried to
give some hints to Jonathan ("hachi") on #git on how to update the patch
to make it apply again:

http://colabti.de/irclogger/irclogger_log/git?date=2010-08-20#l1241

Feel free to pick up from there...

Björn
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to [hidden email]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[Joshua Jensen]

  ----- Original Message -----
From: Björn Steinbrink
Date: 9/8/2010 11:51 PM

>> $ git lg --all
>> * 2c93a8a - (HEAD) c (11 seconds ago)
>> * 2b98bc6 - b (11 seconds ago)
>> * 6abf527 - (d) d (3 minutes ago)
>> | *   472fd93 - (bc-merge) Merge commit 'c' into HEAD (74 seconds ago)
>> | |\
>> |/ /
>> | * d132c87 - (c) c (3 minutes ago)
>> | * 4a88fd1 - b (3 minutes ago)
>> |/
>> * b576660 - (a, master) a (4 minutes ago)
> What you actually wanted to replay is just the first-parent history,
> redoing the merges with their respective original second parent.
> Jonathan (Cc'ed) had the same problem about a month ago.
>
> The "first parent" thing isn't that well defined, criss-cross merges
> combined with some unfortunate fast-forwards will easily make the
> first-parent history become the one that you didn't mean (ask any
> git-svn users that dared to use "git merge" without fully understanding
> how git-svn uses the history ;-)), but it works in this special case.
> And with that special case in mind (and admittedly probably not thinking
> much further) I came up with this patch a few years ago:
>
> http://marc.info/?l=git&m=119379735525213&w=2

I hand merged your change in to my Git installation.  There is one part
of git-rebase--interactive that looks entirely different, so I would
need to look back through the history to determine how to properly apply it.

The good news is the patch does exactly what I want in this simple
case.  I will have to try it on some more complicated hierarchies to see
what the end result is.

It's curious... whenever I ask around about issue X or feature Y, a
previously submitted patch can generally be found in the list archives.  :)

Josh
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to [hidden email]
More majordomo info at  http://vger.kernel.org/majordomo-info.html