Re: Rename of file is causing changes to be lost
On Thu, Mar 8, 2018 at 8:01 AM, Robert Dailey <rcdailey.lists@gmail.com> wrote:
Show 53 quoted lines
> I'm on Windows and core.ignorecase is set to 'true' when I clone/init
> a repository. I've got a branch where I started making changes to a
> file AND renamed it only to change its case. The changes I've made
> were significant enough that git no longer detects a rename, instead
> the files show up as "D" and "A" in git status (deleted then added).
> To correct this, I do an interactive rebase to add an additional
> commit before the first one to rename the file without changing it,
> and *then* allow the second commit to change the file. The goal is
> that rebase should detect the rename and automatically move the
> changes in the (now) second commit to the newly named file. Here's a
> MCVE (treat this as a script):
>
> #/bin/bash
> git init testgitrepo
> cd testgitrepo/
> git config core.ignorecase true # This is set by Windows for me, but
> hopefully will allow this to repro on linux. Didn't test linux though.
> echo "first change" > foo.txt
> git add . && git commit -m 'first change'
> git checkout -b topic
> echo "second change" > foo.txt
> git mv foo.txt FOO.txt
> git add . && git commit -m 'second change'
> git rebase -i master # Move line 1 to line 2, and put "x false" in line 1
> git mv foo.txt FOO.txt && git commit -m 'rename foo'
> git rebase --continue
> git mergetool
>
> After the rebase continue, you will get a conflict like so:
>
> error: could not apply 527d208... second change
>
> When you have resolved this problem, run "git rebase --continue".
> If you prefer to skip this patch, run "git rebase --skip" instead.
> To check out the original branch and stop rebasing, run "git rebase --abort".
>
> Could not apply 527d208... second change
> CONFLICT (rename/delete): foo.txt deleted in 527d208... second change
> and renamed to FOO.txt in HEAD. Version HEAD of FOO.txt left in tree.
>
> The last command, `git mergetool` runs, giving you the option to pick
> the Created (left) or Deleted (right) version of the file:
>
> Left: The file is created, but selecting this erases the changes from
> the "added" version on the remote (which is topic). Basically the
> rename of only case confused git, and we lost the changes on the
> remote version of the file
> Right: File is deleted. Changes are still lost.
>
> The ideal outcome is that the changes from the "added" version of the
> file in the 2nd commit get carried over to the "renamed" version of
> the file, which when you compare the two are named exactly the same
> after the 1st commit is introduced. How can I solve this issue?
Cool, thanks for the testcase. I don't have a good workaround for you, but this is clearly a bug in the merge-recursive logic in git. I guess it's what might be called a rename/add/delete conflict, which git just doesn't handle. Your testcase triggers the bug just fine on linux, though you can trigger the exact same bug without case sensitivity using a slightly different setup (and no need for an interactive rebase):
------ git init foobar cd foobar echo "original file" >foo git add foo git commit -m "original" git branch L git branch R
git checkout L git rm foo echo "different file" >bar git add bar git commit -m "Remove foo, add bar"
git checkout R git mv foo bar git commit -m "rename foo to bar"
git merge L -------
git has code to handle rename/delete conflicts and rename/add conflicts, but since one side of history BOTH deleted foo and added an unrelated bar, that means both types of changes are relevant to the same path (bar) -- essentially, a rename/delete/add conflict. Sadly, git only goes down a codepath that handles one of those two (the rename/delete), and incorrectly throws away the separate add:
----- $ git ls-files -s 100644 78fa0f415ae2bdb5c0182c067eacaaf843699b39 2 bar
$ git ls-tree -r master 100644 blob 78fa0f415ae2bdb5c0182c067eacaaf843699b39 foo $ git ls-tree -r L 100644 blob f286e5cdd97ac6895438ea4548638bb98ac9bd6b bar $ git ls-tree -r R 100644 blob 78fa0f415ae2bdb5c0182c067eacaaf843699b39 bar -----
But the problem is actually a bit bigger than shown here; there are higher order corner cases here too. I realized in the past that e.g. rename/rename(1to2) could also have rename/add conflicts for each rename (thus making a rename/rename/add/add conflict possible), but I also felt there were probably some other bad interactions out there. I figured they were likely theoretical only, so I didn't bother investigating. But, combining your example with that other one, we should also be able to come up with a rename/rename/add/add/delete/delete conflict. I wonder if there are others...
Anyway, I recorded this at https://bugs.chromium.org/p/git/issues/detail?id=11. Sorry I don't have a workaround, but I'll try to eventually get back to this and fix it.
