Fix another crazy rename assertion

This fixes another special corner case that was being triggered at GitHub; the error being triggered is the same as what I submitted a fix for a few months ago, but the way it was triggered and the fix needed are different in this case. See the final commit message for details. The first two patches are just tiny cleanups I noticed while investigating the problem.

I will also note that I first came up with an alternative fix -- checking in use_cached_pairs() whether new_name was contained in the opt->priv->paths strmap, and if not, skipping to the next cached rename instead of adding it to pairs. That would also work, but it would mean that if a yet-subsequent commit after that did modify the old/file path, I think we'd have to re-detect the rename, which would hurt the effectiveness of the cached renames optimization. Simply avoiding using it in process_renames() allows it to avoid being forgotten (and since old/file is NOT modified, the upstream rename remains valid). Besides, this fix is nicely symmetrical to the check on !oldinfo, so it seems more aesthetic to me as well as helping us preserve performance.

Elijah Newren (3):
t6429: update comment to mention correct tool
merge-ort: remove debugging crud
merge-ort: fix failing merges in special corner case

 merge-ort.c                              | 31 +++++++-
 t/t6429-merge-sequence-rename-caching.sh | 93 ++++++++++++++++++++++--
 2 files changed, 114 insertions(+), 10 deletions(-)

base-commit: 4253630c6f07a4bdcc9aa62a50e26a4d466219d1
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1992%2Fnewren%2Ffix-another-crazy-rename-assertion-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1992/newren/fix-another-crazy-rename-assertion-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1992

gitgitgadget

From: Elijah Newren <[email protected]>

While developing commit a16e8efe5c2b (merge-ort: fix merge.directoryRenames=false, 2025-03-13), I was testing things out and had an extra condition on one of the if-blocks that I occasionally swapped between '&& 0' and '&& 1' to see the effects of the changes. I forgot to remove it before submitting and it wasn't caught in review. Remove it now.

---
 merge-ort.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 29858074f9..23b55c5b92 100644
@@ -3438,7 +3438,7 @@ static int collect_renames(struct merge_options *opt,
 			continue;
 		}
 		if (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_NONE &&
-		    p->status == 'R' && 1) {
+		    p->status == 'R') {
 			possibly_cache_new_pair(renames, p, side_index, NULL);
 			goto skip_directory_renames;
 		}

gitgitgadget

From: Elijah Newren <[email protected]>

At GitHub, we had a repository that was triggering
git: merge-ort.c:3032: process_renames: Assertion newinfo && !newinfo->merged.clean failed.
during git replay.

This sounds similar to the somewhat recent f6ecb603ff8a (merge-ort: fix directory rename on top of source of other rename/delete, 2025-08-06), but the cause is different. Unlike that case, there are no rename-to-self situations arising in this case, and new to this case it can only be triggered during a replay operation on the 2nd or later commit being replayed, never on the first merge in the sequence.

To trigger, the repository needs:
* an upstream which:
* renames a file to a different directory, e.g. old/file -> new/file
* leaves other files remaining in the original directory (so that e.g. "old/" still exists upstream even though file has been removed from it and placed elsewhere)
* a topic branch being rebased where:
* a commit in the sequence:
* modifies old/file
* a subsequent commit in the sequence being replayed:
* does NOT touch *anything* under new/
* does NOT touch old/file
* DOES modify other paths under old/
* does NOT have any relevant renames that we need to detect _anywhere_ elsewhere in the tree (meaning this interacts interestingly with both directory renames and cached renames)

In such a case, the assertion will trigger. The fix turns out to be surprisingly simple. I have a very vague recollection that I actually considered whether to add such an if-check years ago when I added the very similar one for oldinfo in 1b6b902d95a5 (merge-ort:
process_renames() now needs more defensiveness, 2021-01-19), but I think I couldn't figure out a possible way to trigger it and was worried at the time that if I didn't know how to trigger it then I wasn't so sure that simply skipping it was correct. Waiting did give me a chance to put more thorough tests and checks into place for the rename-to-self cases a few months back, which I might not have found as easily otherwise. Anyway, put the check in place now and add a test that demonstrates the fix.

Note that this bug, as demonstrated by the conditions listed above, runs at the intersection of relevant renames, trivial directory resolutions, and cached renames. All three of those optimizations are ones that unfortunately make the code (and testcases!) a bit more complex, and threading all three makes it a bit more so. However, the testcase isn't crazy enough that I'd expect no one to ever hit it in practice, and was confused why we didn't see it before. After some digging, I discovered that merge.directoryRenames=false is a workaround to this bug, and GitHub used that setting until recently (it was a "temporary" match-what-libgit2-does piece of code that lasted years longer than intended). Since the conditions I gave above for triggering this bug rule out the possibility of there being directory renames, one might assume that it shouldn't matter whether you try to detect such renames if there aren't any. However, due to commit a16e8efe5c2b (merge-ort: fix merge.directoryRenames=false, 2025-03-13), the heavy hammer used there means that merge.directoryRenames=false ALSO turns off rename caching, which is critical to triggering the bug. This becomes a bit more than an aside since...

Re-reading that old commit, a16e8efe5c2b (merge-ort: fix merge.directoryRenames=false, 2025-03-13), it appears that the solution to this latest bug might have been at least a partial alternative solution to that old commit. And it may have been an improved alternative (or at least help implement one), since it may be able to avoid the heavy-handed disabling of rename cache. That might be an interesting future thing to investigate, but is not critical for the current fix. However, since I spent time digging it all up, at least leave a small comment tweak breadcrumb to help some future reader (myself or others) who wants to dig further to connect the dots a little quicker.

---
 merge-ort.c                              | 29 ++++++++-
 t/t6429-merge-sequence-rename-caching.sh | 78 ++++++++++++++++++++++++
 2 files changed, 106 insertions(+), 1 deletion(-)

index 23b55c5b92..a1f3333e44 100644
@@ -2912,6 +2912,32 @@ static int process_renames(struct merge_options *opt,
 		if (!oldinfo || oldinfo->merged.clean)
 			continue;
 
+		/*
+		 * Rename caching from a previous commit might give us an
+		 * irrelevant rename for the current commit.
+		 *
+		 * Imagine:
+		 *     foo/A -> bar/A
+		 * was a cached rename for the upstream side from the
+		 * previous commit (without the directories being renamed),
+		 * but the next commit being replayed
+		 *     * does NOT add or delete files
+		 *     * does NOT have directory renames
+		 *     * does NOT modify any files under bar/
+		 *     * does NOT modify foo/A
+		 *     * DOES modify other files under foo/ (otherwise the
+		 *       !oldinfo check above would have already exited for
+		 *       us)
+		 * In such a case, our trivial directory resolution will
+		 * have already merged bar/, and our attempt to process
+		 * the cached
+		 *     foo/A -> bar/A
+		 * would be counterproductive, and lack the necessary
+		 * information anyway.  Skip such renames.
+		 */
+		if (!newinfo)
+			continue;
+
 		/*
 		 * diff_filepairs have copies of pathnames, thus we have to
 		 * use standard 'strcmp()' (negated) instead of '=='.

@@ -5118,7 +5144,8 @@ static void merge_check_renames_reusable(struct merge_options *opt,
 	 * optimization" comment near that case).
 	 *
 	 * This could be revisited in the future; see the commit message
-	 * where this comment was added for some possible pointers.
+	 * where this comment was added for some possible pointers, or the
+	 * later commit where this comment was added.
 	 */
 	if (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_NONE) {
 		renames->cached_pairs_valid_side = 0; /* neither side valid */

gitgitgadget

From: Elijah Newren <[email protected]>

A comment at the top of t6429 mentions why the test doesn't exercise git rebase or git cherry-pick. However, it claims that it uses test-tool fast-rebase. That was true when the comment was written, but commit f920b0289ba3 (replay: introduce new builtin, 2023-11-24) changed it to use git replay without updating this comment.

We could potentially just strike this second comment, since git replay is a bonified built-in, but perhaps the explanation about why it focuses on git replay is still useful. Update the comment to make it accurate again.

---
 t/t6429-merge-sequence-rename-caching.sh | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)

index 0f39ed0d08..dcb734b10b 100755
@@ -11,14 +11,13 @@ test_description="remember regular & dir renames in sequence of merges"
 #         sure that we are triggering rename caching rather than rename
 #         bypassing.
 #
-# NOTE 2: this testfile uses 'test-tool fast-rebase' instead of either
-#         cherry-pick or rebase.  sequencer.c is only superficially
-#         integrated with merge-ort; it calls merge_switch_to_result()
-#         after EACH merge, which updates the index and working copy AND
-#         throws away the cached results (because merge_switch_to_result()
-#         is only supposed to be called at the end of the sequence).
-#         Integrating them more deeply is a big task, so for now the tests
-#         use 'test-tool fast-rebase'.
+# NOTE 2: this testfile uses replay instead of either cherry-pick or rebase.
+#         sequencer.c is only superficially integrated with merge-ort; it
+#         calls merge_switch_to_result() after EACH merge, which updates the
+#         index and working copy AND throws away the cached results (because
+#         merge_switch_to_result() is only supposed to be called at the end
+#         of the sequence).  Integrating them more deeply is a big task, so
+#         for now the tests use 'git replay'.
 #
 
 

gitgitgadget

s/bonified/bona fide/ ?

Yep, good catch. Got it fixed locally; will wait to see if any other feedback comes in.

And nothing seems to have happened since then. I can amend the typo away if you want after the release before starting to merge topics down to 'next' again.

If it's easier for you to amend locally, that's great, but if it's easier for you to have me send a re-rolled series, I've got it all queued up and ready to go -- it's just this one typofix. Sorry for not getting it sent out a little sooner.

I grew very fond of "git commit --fixup amend:<that-commit>" followed by "git rebase --keep-base --autosquash", so it is not a problem for me.

Thanks.