git and appending merge commits

git and appending merge commits

[Adam Majer]

Hi all,

I've recently started using Git and it is a very nice tool. During some
merging, I've come upon one type of issue that seems to be a bug
somewhere. To reproduce,


  1. branch master
  2. branch upstream
  3. add new version to upstream
  4. merge it to master
  5. append master merge changeset

#5 results in "messsed up" diffs

For an example of such a mess up please see,

http://git.debian.org/?p=collab-maint/mrtg.git;a=commitdiff;h=a8bc78ffd6d51ab09a791fa97e25f57b60eecd06

It appears that the tool generates a patch to a patch and displays that
instead of the appended merge diff which is what I would have expected..

Is current output by design?

- Adam

PS. I should have probably used a new commit to do that changelog
addition instead of appending the merge!

Re: git and appending merge commits

[Shawn O. Pearce]

Adam Majer <adamm@zombino.com> wrote:
...
>   5. append master merge changeset
> 
> #5 results in "messsed up" diffs
> 
> For an example of such a mess up please see,
> 
> http://git.debian.org/?p=collab-maint/mrtg.git;a=commitdiff;h=a8bc78ffd6d51ab09a791fa97e25f57b60eecd06
> 
> It appears that the tool generates a patch to a patch and displays that
> instead of the appended merge diff which is what I would have expected..
> 
> Is current output by design?

Yes, it is by design.  This is a feature of Git that very few
(if any) other systems have.

What you are seeing here is a combined diff.  It is a diff of the
final output of the merge against its two parents (the two branches
that were merged together).  Places where both a "++" or "--" prefix
a line indicate a place where the merge commit adds something that
is not in either parent.  This is stuff that the merge author edited
himself/herself during the merge.

If you had not amended the merge, you would have seen a less
interesting diff here, as the combined diff output tries to
avoid showing trivial changes.

-- 
Shawn.

Re: git and appending merge commits

[Junio C Hamano]

Adam Majer <adamm@zombino.com> writes:

> Is current output by design?

Yes, it is _very much_ by design and it not "messed up" at all.  Look at
the output of URL you quoted and notice that you are asking for condensed
combined diff ("diff --cc").

For a merge commit with N parents, you can define "a diff for the
resulting merge commit" in various ways:

 * You could have a set of diffs between each parent to the merge result.
   You can get this with "git diff-tree -p -m $commit".

 * You could say the first parent is special, and define it as the diff
   between its first parent and the merge result.  You can get this by
   asking "git diff $commit^1 $commit".

 * You can say what non-trivial changes the merge itself did.  This is the
   combined merge and what "git show $commit" (and "git log -p") gives.

Depending on what you want, you can ask these various forms of diffs.

When you are reviewing the changes series of commits introduced to the
history (iow, "git log -p"), a single diff between parent and the merge
result is not so interesting, primarily because each individual change
from commits on each side branch is (or should be) much easier to read and
understand, but if a merge needed to do something non-trivial to resolve
conficts, you would want to know about it, and that is the reason why --cc
format is the default.

In a combined diff, unlike the single column to show "+/-/ " at the
beginning for one-parent diff, each parent gets one column.  And in a
condensed combined diff, a hunk whose result is inherited solely from one
parent (iow, a trivial conflict resolution) is removed from the output, so
that only a non-trivial conflict resolution that leaves result that is
different from all parents are shown.

Re: git and appending merge commits

[Adam Majer]

Junio C Hamano wrote:
> In a combined diff, unlike the single column to show "+/-/ " at the
> beginning for one-parent diff, each parent gets one column.  And in a
> condensed combined diff, a hunk whose result is inherited solely from one
> parent (iow, a trivial conflict resolution) is removed from the output, so
> that only a non-trivial conflict resolution that leaves result that is
> different from all parents are shown.

Thank you Junio and Shawn for the very clear explanation! Now it makes
very good sense why the patch looks so "weird" :) For cases where both
branches touch the same file and there is a conflict, this would be very
useful indeed.

Cheers,
Adam