Now I think I am done with diff...

JCH> Now I think I am done with diff,...

I think I have fixed the diff-raw output routine enough to address the rename and copy distinction issues (although I chose not to give individual entries "this is a copy" or "this is a rename" indicators).

And I thought I was really done with diff this time, but here are some leftover issues I would like to further address, not necessarily in this order, and not necessarily promising to do all of them.

 - Make "diff-raw human readable" output quite different from
   "diff-raw machine readable" format.  I think human readable
   format should be more like "concise version of diff-patch
   format for human consumption", not the current "LF and TAB
   separated version of machine readable output", to enhance
   readability.  This includes:

   (1) not repeating paths when src and dst are the same
       (i.e. most of the time);

   (2) substituting src/dst with /dev/null when talking about
       creation or deletion;

   (3) prefixing the src and dst with '+' or '*' or '-'.

   Currently I do not intend to add _parsing_ of such output to
   diff-helper because you can always use -z on both ends of the
   pipe (and as yourself said "Sure. Although I doubt people use
   the raw diff output except to (a) feed to diff-helper or (b)
   check that it's non-empty."), but I have to sleep on this.

 - Add an option to cause diff-tree three-brothers not to call
   diffcore_prune() at all.  This will essentially produce
   something like a merge of two output from ls-files or
   ls-trees commands, but with renames/copies resolved.  I do
   not know how useful this would be, especially now the
   immediate problem of diff_flush() pruning all the
   "non-change" entries is (hopefully) fixed.

 - Give pickaxe an option to cause it to produce diffs for the
   entire commit, not just the found file.

 - Experiment with the patch reordering I talked about in the
   previous message.

 - Add the support of "copied from post-edit image" detection in
   rename/copy logic.  Currently I compare only with pre-edit
   image for modified files.

 - Read and understand the patch-delta.c, to make rename/copy
   similarity estimator more accurate.  As you pointed out, it
   does very bad things on big consecutive deletes.

 - Think about performance issues around "diff-tree -C" stuff.

 - Think more about the problem with the patch output between
   two symbolic links, and how to fix it (I've been quietly
   thinking about this for some time ;-)).  I think the current
   output of just comparing the two readlink results and
   generating a normal diff between them is broken and hazardous
   if such a thing is mistakenly fed to patch -p1.

 - Teach JIT about the diff-raw format changes.