Hi all,
As mentioned during the standup in #git-devel today, we encountered a
BUG() during certain circumstances of a criss-cross merge. Timing has
prevented me from getting the repro case together to send just yet, but
it appears that the issue is as follows:
- During a merge in a repo with fairly complicated, merge-y history,
the following BUG() is seen:
"BUG: There are unmerged index entries:
BUG: 2 baz/bar.txt
BUG: merge-recursive.c:429: unmerged index entries in
merge-recursive.c
Aborted"
- But when the user then runs `git status` (after clearing
.git/index.lock) the directory is clean.
- The repo does not contain a 'baz/bar.txt' (although 'baz/' exists).
- In the repo's history a 'foo/bar.txt' can be found (that is the only
'bar.txt' to ever exist in the project).
Digging further shows:
- The merge had multiple closest shared ancestors (criss-cross
merge)
- Directory 'foo/' was renamed on one side to 'baz/'...
- ...and 'foo/bar.txt' was deleted on the other side.
- When the virtual ancestor is generated, the directory rename can't be
resolved and leaves a conflict.
- The virtual ancestor being written to disk is in a conflicted state.
- This causes the top-level merge to fail, printing the BUG() above
which references a 'baz/bar.txt' that never really lived there.
Furthermore...
- If merge.directoryrenames is set to any value besides "conflict", the
merge succeeds (no BUG() generated).
This seems to have been introduced in 8c8e5bd6eb331d055aa7fa6345f6dcdadd658979
although I haven't bisected it yet.
It seems like the solution is to watch for whether we're making a
virtual ancestor during a recursive merge, and if so, to treat
merge.directoryrenames = "conflict" as "true" or "false" instead.
I plan to modify the tests added in 8c8e5bd6 to highlight this issue
(just haven't had time, and probably won't til next week), and send a
patch doing the special treatment on merge.directoryrenames.
Happy to hear other ideas folks have.
- Emily
