Re: [PATCH] combine-diff: use textconv for combined diff format

Michael J Gruber [off-list ref] writes:

Junio C Hamano venit, vidit, dixit 15.04.2011 20:34:

quoted

Michael J Gruber [off-list ref] writes:

quoted

git diff -m produces a combined diff!

Hmm, what is the rest of your command line?  I thought -m was a way to ask
pairwise diff with each parent.

Sure, but it does not always work like that. Just look at the test from
my patch, or do any "git merge --no-commit" and then "git diff -m". I
would expect that to compare the worktree to each parent, but in fact it
runs "diff --cc".

Thanks; it wasn't clear you are comparing stages with the working tree.
Asking for the rest of the command line paid off ;-)

And yes, comparing multiple entries with the worktree files is done in
run_diff_files() defined in diff-lib.c; it is unaware of the -m option
that was originally defined for diff-tree to show pairwise diff.  That
codepath never cared about -m before nor after -c/--cc was invented for
diff-tree, and it only learned about -c/--cc when it was introduced.  I
think diff-index is unaware of the -m option from the same historical
background (read: not "for the same reason or justification").

We may want to change that, but I am personally not very interested.  We
can ask to diff against a specific stage, I know that is what I do, and I
think that is what most people do [*1*].

Yes this is the case for diff-tree running pair-wise comparison.

so I'm wondering whether we shouldn't stay closer to that with "--cc
also", e.g.:

aa01ae1 Merge branch 'master' into somebranch
diff --cc a
index 72594ed,86e041d..d8323da
Binary files a/a and b/a differ

The -c/--cc options are about presenting the pairwise -m output in a
different way by combining and condensing.  So in that sense, if we really
want to combine and condense information, one possibility is to do:

 Binary files a/a and c/a differ, b/a and c/a differ.

naming each parent as 'a', 'b', ... and giving the highest letter (in the
two-parent merge case, 'c') to the final result.  By doing so, you can
express where in its tree each parent had the content when you are viewing
a renaming merge.

Bu that opens an old can of worms we should have opened and closed four
years ago.

The header shows "diff --cc a" followed by "--- a/a" followed by "+++ b/a"
before the hunk for a two-way merge.  But if we are to "combine and
condense", another possibility is to show:

    diff --cc a/a b/a c/a
    index bf7c788,fa9d23a,5d24d9f..cc69134
    --- a/a
    --- b/a
    +++ c/a
    @@@@ -74,26 -74,6 -74,29 +74,50 @@@@
    ...

to keep the paths information.  I do not think anybody cared so far, and
perhaps we should have done it when we introduced -c/--cc, but it is not
at all worth changing now.

That means that we are not all that worried about losing the rename
information when showing such a diff in --cc/-c form.  After all, the
"diff --cc a" header is not "diff --cc a/a b/a c/a" that mentions all
paths, and "--- a/a" lines are not repeated for each parent.  So while
showing the names like you suggested may be a possibility, I think an
approach that is more in line with the current output would be:

 "Binary files a in different versions differ"

or something without naming them with a/, b/, ...

In short, it all depends on how much we condense when running -c/--cc.  We
are inconsistent by showing both "--- a/a" and "+++ b/a" lines, but modulo
that we condense away the renamed path information in our current output.
And the final alternative in the previous paragraph would be more in line
with that design.


[Footnote]

*1* I also often use 

 diff HEAD...MERGE_HEAD $path
 diff HEAD $path
 diff MERGE_HEAD $path

during a conflicted merge when it is hard to read in the --cc form.