[PATCH v5 0/2] Let merge-file write out conflict markers with correct EOLs

The original patch was sent by Beat Bolli in http://thread.gmane.org/gmane.comp.version-control.git/281600

My suggestion to extend it to respect gitattributes led to changes that broke the original patch. And they were misguided to begin with (see below).

Since there have been a couple of "What's cooking" mails containing unheard calls for updates on this patch, I took it on myself to fix things.

Junio's comment that we might look at blobs containing CR/LF line endings made me rethink the entire approach, and I am now convinced that we need to abandon the entire idea to make the conflict markers depend on config settings or attributes: after all, I introduced `git merge-file` as a replacement for GNU merge that can be used *outside* of any repository, by design.

The crucial idea hit me yesterday when I took a step back: all we need to do is to ensure that the end-of-line style is matched when *all* input files are LF-only, or when they are all CR/LF. In all other cases, we have mixed line endings anyway.

And to do that, it is sufficient to look at *one single line ending* in the context. Technically, it does not even have to be the context, but the first line endings of the first file would be enough, however it is so much more pleasant if the conflict marker's eol matches the one of the preceding line.

To prevent my future self from repeating mistakes, I added a little bit of background to the first commit message.

Triggered by a comment by Junio, I realized that a second patch is needed: we need to make sure that the conflicting sections are also augmented by the appropriate line endings if they lack any.

The change relative to v4 is that I now test the correct conflict markers: the merge-file call we add to t6023 actually produces four conflict markers because a previous test switched the conflict style to `diff3` mode. This is only a side effect, though, therefore I really wanted to avoid testing for it. However, I managed to test an *incorrect* subset of three, but this is now fixed. Thanks go to Ramsay Jones for poking my nose onto that issue.

Johannes Schindelin (2):
  merge-file: let conflict markers match end-of-line style of the
    context
  merge-file: ensure that conflict sections match eol style

 t/t6023-merge-file.sh | 13 +++++++
 xdiff/xmerge.c        | 98 +++++++++++++++++++++++++++++++++++++++++----------
 2 files changed, 93 insertions(+), 18 deletions(-)

Interdiff vs v4:

 diff --git a/t/t6023-merge-file.sh b/t/t6023-merge-file.sh
 index 1390548..20aee43 100755
 --- a/t/t6023-merge-file.sh
 +++ b/t/t6023-merge-file.sh
 @@ -352,11 +352,11 @@ test_expect_success 'conflict sections match existing line endings' '
  	printf "1\\r\\n2\\r\\n5" >crlf-diff2.txt &&
  	test_must_fail git -c core.eol=crlf merge-file -p \
  		crlf-diff1.txt crlf-orig.txt crlf-diff2.txt >crlf.txt &&
 -	test $(tr "\015" Q <crlf.txt | grep "\\.txtQ$" | wc -l) = 3 &&
 +	test $(tr "\015" Q <crlf.txt | grep "^[<=>].*Q$" | wc -l) = 3 &&
  	test $(tr "\015" Q <crlf.txt | grep "[345]Q$" | wc -l) = 3 &&
  	test_must_fail git -c core.eol=crlf merge-file -p \
  		nolf-diff1.txt nolf-orig.txt nolf-diff2.txt >nolf.txt &&
 -	test $(tr "\015" Q <nolf.txt | grep "\\.txtQ$" | wc -l) = 0
 +	test $(tr "\015" Q <nolf.txt | grep "^[<=>].*Q$" | wc -l) = 0
  '
  
  test_done