Re: Two bugs with renaming

Linus Torvalds [off-list ref] writes:

Now, arguably, it should just see them as two independent issues, and then 
the rename detection will notice that the "files/delete.me" file got 
renamed as "files.upstream/delete.me", so *after* rename detection there 
will be no D/F conflict in the end result, but we see the conflict before 
that all even happens.

The common ancestor had files/delete.me.

Our side (test_branch) deleted files/delete.me, created
files.upstream/delete.me and created files.

Their side (master)kept files/delete.me without changing, created foo.

So I would think we should not even trigger D/F conflict at all.

 - files/delete.me is deleted by only one side (our side) and it should go.

 - files and files.upstream/delete.me are created by only one side (ours)
   and it should stay.

 - foo is created by only one side (theirs) and it should come.

However, we wanted to allow policy decision to happen after read-tree, and
we traditionally kept "one side removes" case as an internal conflict in
the index, later to be resolved by merge-one-file (and merge-recursive).

So I think the resulting index should look like this:

        100644 xxxxxxx 0	files.upstream/delete.me
        100644 xxxxxxx 1	files/delete.me
        100644 xxxxxxx 3	files/delete.me
        120000 xxxxxxx 0	files
        100644 xxxxxxx 0	foo

or (if we also want to leave policy decision for "one side adds" case to
merge-one-file and merge-recursive) even:

        100644 xxxxxxx 1	files/delete.me
        100644 xxxxxxx 3	files/delete.me
        120000 xxxxxxx 2	files
        100644 xxxxxxx 2	files.upstream/delete.me
        100644 xxxxxxx 2	foo

But that is not what is happening here.  In fact, if you did not have
"files" in the test branch, here is what you will see:

        100644 xxxxxxx 0	files.upstream/delete.me
        100644 xxxxxxx 1	files/delete.me
        100644 xxxxxxx 3	files/delete.me
        100644 xxxxxxx 0	foo

and merge-recursive knows how to match up the first three entries and if
there are changes between stages #1 and #3 of files/delete.me, that is
carried forward to files.upstream/delete.me

Your unpack_trees() is bug-to-bug compatibile with Daniel's that is in
1.5.4.  Both "read-tree -m -u" bails out with the same error, without
even leaving higher stage entries in the index.