Re: [PATCH] Fix rebase with file move when diff.renames = copies

On Jonathan Nieder wrote:

David D. Kilzer wrote:

quoted

With diff.renames = copies, a rebase with a  file move will fail with
the following error:

     fatal: mode change for <file>, which is not in current  HEAD
    Repository lacks necessary blobs to fall back on  3-way merge.
    Cannot fall back to three-way  merge.
    Patch failed at 0001.

I would think that the  following works fine:

    git init test-repo  &&
    cd test-repo &&
     echo hello >greeting.txt &&
    git add  greeting.txt &&
    git commit -m base  &&
    git checkout -b move  &&
    git mv greeting.txt moved.txt  &&
    git commit -m move  &&
    git checkout master  &&
    echo hi >greeting.txt  &&
    git add greeting.txt  &&
    git commit -m change  &&
    git checkout move  &&
    echo '[diff] renames = copies'  >>.git/config &&
    git rebase  master

What am I doing wrong?

Given the following tree:

      B' topic
     /
    A---B master

A: New file "F1" is committed.
B: New file "F2" is committed.
B': New file "F2" is committed (identical in content to "F2" on B), and "F1" is 
renamed to "F3".

When the topic branch is rebased onto master with diff.renames=copies, git fails 
when attempting to build a fake ancestor for F1.  The key to reproducing the bug 
is to have an identical new file added on both B and B'.

My original patch in <http://marc.info/?l=git&m=122635667614099&w=2> addressed 
this in builtin-apply.c, but Junio didn't like this approach as noted in 
<http://marc.info/?l=git&m=122636097120953&w=2>.

Patch does not apply to master or maint, due to conflict  with
v1.7.1-rc0~37^2~5 (rebase: support automatic notes  copying,
2010-03-12).  One sneaky way to avoid this kind of thing would  be to
insert new tests at some logical point in the middle of a test  script.

Sorry about that--I forgot to rebase it to maint before sending it.

Test nitpicks:

Thanks!  I'll make the requested changes in the next patch.

This wants to notice that Y  was already added so the top patch can be
simplified to include only a  rename.

Actually, this is the key to reproducing the bug!

Can you explain why this test will fail without your  patch?

Here is a stand-alone script that reproduces the bug:

    git init test-repo &&
    cd test-repo &&
    echo hello > F1 &&
    git add F1 &&
    git commit -m "A" &&
    git checkout -b topic &&
    echo hi > F2 &&
    git add F2 &&
    git mv F1 F3 &&
    git commit -m "B'" &&
    git checkout master &&
    echo hi > F2 &&
    git add F2 &&
    git commit -m "B" &&
    git checkout topic &&
    git config diff.renames copies &&
    GIT_TRACE=1 git rebase master

Note that the test case in my patch depended on "F1" (which was "A") being 
committed by an earlier test.

Dave