Avoid conflicts when merging branches with mixed normalization

Currently, merging across changes in line ending normalization is
painful since files containing CRLF will conflict with normalized files,
even if the only difference between the two versions is the line
endings.  Additionally, any "real" merge conflicts that exist are
obscured because every line in the file has a conflict.

Assume you start out with a repo that has a lot of text files with CRLF
checked in (A):

      o---C
     /     \
    A---B---D

B: Add "* text=auto" to .gitattributes and normalize all files to
   LF-only

C: Modify some of the text files

D: Try to merge C

You will get a ridiculous number of LF/CRLF conflicts when trying to
merge C into D, since the repository contents for C are "wrong" wrt the
new .gitattributes file.

Fix ll-merge so that the "base", "theirs" and "ours" stages are passed
through convert_to_worktree() and convert_to_git() before a three-way
merge.  This ensures that all three stages are normalized in the same
way, removing from consideration differences that are only due to
normalization.

This feature is optional for now since it changes a low-level mechanism
and is not necessary for the majority of users.  The "merge.renormalize"
config variable enables it.

Signed-off-by: Eyvind Bernhardsen <eyvind.bernhardsen@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

Documentation/gitattributes.txt

@@ -317,6 +317,17 @@ command is "cat").
 	smudge = cat
 ------------------------
 
+For best results, `clean` should not alter its output further if it is
+run twice ("clean->clean" should be equivalent to "clean"), and
+multiple `smudge` commands should not alter `clean`'s output
+("smudge->smudge->clean" should be equivalent to "clean").  See the
+section on merging below.
+
+The "indent" filter is well-behaved in this regard: it will not modify
+input that is already correctly indented.  In this case, the lack of a
+smudge filter means that the clean filter _must_ accept its own output
+without modifying it.
+
 
 Interaction between checkin/checkout attributes
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -331,6 +342,29 @@ In the check-out codepath, the blob content is first converted
 with `text`, and then `ident` and fed to `filter`.
 
 
+Merging branches with differing checkin/checkout attributes
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+If you have added attributes to a file that cause the canonical
+repository format for that file to change, such as adding a
+clean/smudge filter or text/eol/ident attributes, merging anything
+where the attribute is not in place would normally cause merge
+conflicts.
+
+To prevent these unnecessary merge conflicts, git can be told to run a
+virtual check-out and check-in of all three stages of a file when
+resolving a three-way merge by setting the `merge.renormalize`
+configuration variable.  This prevents changes caused by check-in
+conversion from causing spurious merge conflicts when a converted file
+is merged with an unconverted file.
+
+As long as a "smudge->clean" results in the same output as a "clean"
+even on files that are already smudged, this strategy will
+automatically resolve all filter-related conflicts.  Filters that do
+not act in this way may cause additional merge conflicts that must be
+resolved manually.
+
+
 Generating diff text
 ~~~~~~~~~~~~~~~~~~~~
 

Documentation/merge-config.txt

@@ -15,6 +15,16 @@ merge.renameLimit::
 	during a merge; if not specified, defaults to the value of
 	diff.renameLimit.
 
+merge.renormalize::
+	Tell git that canonical representation of files in the
+	repository has changed over time (e.g. earlier commits record
+	text files with CRLF line endings, but recent ones use LF line
+	endings).  In such a repository, git can convert the data
+	recorded in commits to a canonical form before performing a
+	merge to reduce unnecessary conflicts.  For more information,
+	see section "Merging branches with differing checkin/checkout
+	attributes" in linkgit:gitattributes[5].
+
 merge.stat::
 	Whether to print the diffstat between ORIG_HEAD and the merge result
 	at the end of the merge.  True by default.