retrospective move

Hi. Awhile ago I copied some files in my local repo then commited. In another commit I deleted the original files, so essentially I did a mv that git doesn't know about. Now I'm trying to merge from another branch where changes have happened to the files in their original locations, and it's making the merge somewhat difficult. Is there a way to ease this?

jemmyw wrote:

Git never "knows" about a move, it only detects them after the fact. A corollary of this is that 'git mv a b' is more or less the same as 'cp a b; git rm a; git add b'.

Now since you have a copy followed by a removal in two separate commits, the merge logic won't treat this as a rename. Some ideas for possible solutions:

* If your cp&rm commits are strictly local, you could squash them into
  a single commit.  See man git-rebase(1), especially the -i flag.
  That way, the merge logic will recognise it as a rename.

* You can use an extra line of history that has a "real" one-step mv.
  Suppose you currently have

     *---*---*---*---*      upstream
      \
       \
        +--C--o--R--o--o    local

  where C is the copy and R is the removal.  Then you can introduce an
  extra "gateway" branch that does just the move:

     *---*---*---*---*      upstream
     |\
     | \
     |  +--M                gateway
     |
      \
       \
        +--C--o--R--o--o    local

  Now merge gateway to local (i.e. git checkout local && git merge
  gateway), and resolve all resulting conflicts.  I think you can
  actually use '-s ours' in that merge, but be sure to check the
  results.

     *---*---*---*---*      upstream
     |\
     | \
     |  +--M                gateway
     |      \
      \      +----------+
       \                 \
        +--C--o--R--o--o--1    local

  Finally, you can use the gateway branch to "do the rename" on all
  incoming new changes, by first merging upstream to gateway and later
  again gateway to local, like so:
  
     *---*---*---*---*.      upstream
     |\                \
     | \                \
     |  +--M-------------2       gateway
     |      \             \
      \      +----------+  \
       \                 \  \
        +--C--o--R--o--o--1--3   local

  It won't be very pretty history, but it should only give (rename
  related) merge conflicts at M.

HTH :-)

2009/2/4 Thomas Rast <trast@student.ethz.ch>:

I don't think it holds. Git only sees the endpoints and the bases, it does not matter if the "move" was actually two commits. Maybe they have changed dramatically and git is not able to detect the rename (or the files are small). Another possible cause could be that you have a lot of renames (see config merge.renamelimit).

You could see if git detects the rename diffing differents commits:

$ git diff --raw -M commit1 commit2

if it does not detect you could use other limits as:

$ git diff --raw -M50 commit1 commit2

while commit1 and 2 are the endpoints or the merge base.

It could hep if you could provide a minimal test case, or at least the commands and output you get.

HTH, Santi

Santi Béjar wrote:

Hmm, you're right.  I stand corrected.

Here's a sequence of commands that makes a test case:

  git init
  seq 1 10 > A
  git add A
  git commit -m initial
  seq 11 15 >> A
  git commit -m second A
  seq 16 20 >> A
  git commit -m third
  git checkout -b side HEAD~2
  cp A B
  git add B
  git commit -m copy
  seq 3 10 > B               # <--
  git commit -m edit B
  git rm A
  git commit -m delete

Then running

  git merge master

goes through cleanly, but changing the line indicated above to 'seq 5 10' makes it less similar, falling through the detection and causing a conflict. The minimum required similarity for merges does not seem to be configurable unfortunately.

2009/2/4 Thomas Rast <trast@student.ethz.ch>:

[...]

Thanks, but that's known, this is why I wrote "or the files are small". I was not clear, but what I wanted was a *real* test case. In which cases this is a problem in practice.

No, but nevertheless it works fine most of the time, except with small files.

HTH, Santi