Skip to contentSkip to conversation

BUG() during criss-cross merge with directory rename and deleted file

PATCH In review Emily Shaffer 4 participants 11 messages Jul 26, 2019 – Aug 8, 2019

lore
Share & tools
lore
#0 lore raw

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

#1 lore raw

PATCH 1/1 merge-recursive: avoid directory rename detection in recursive case

Ever since commit 8c8e5bd6eb33 ("merge-recursive: switch directory rename detection default", 2019-04-05), the default handling with directory rename detection was to report a conflict and leave unstaged entries in the index. However, when creating a virtual merge base in the recursive case, we absolutely need a tree, and the only way a tree can be written is if we have no unstaged entries -- otherwise we hit a BUG().

There are a few fixes possible here which at least fix the BUG(), but none of them seem optimal for other reasons; see the comments with the new testcase 13e in t6043 for details (which testcase triggered a BUG() prior to this patch). As such, just opt for a very conservative and simple choice that is still relatively reasonable: have the recursive case treat 'conflict' as 'false' for opt->detect_directory_renames.

---

I really should introduce constants like
DETECT_DIRECTORY_RENAMES_NEVER = 0
DETECT_DIRECTORY_RENAMES_CONFLICT = 1
DETECT_DIRECTORY_RENAMES_YES = 2
and then use them in the code to make it clearer, but I wanted to make the code change as simple and contained as possible given that this is built on maint, fixes a BUG() and we're already in -rc1.

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

(This fix does merge down cleanly to master, next, and pu.)

 merge-recursive.c                   |   3 +-
 t/t6043-merge-rename-directories.sh | 111 ++++++++++++++++++++++++++++
 2 files changed, 113 insertions(+), 1 deletion(-)
merge-recursive.c +2 -1
index d2e380b7ed..c7691d9b54 100644
@@ -2856,7 +2856,8 @@ static int detect_and_process_renames(struct merge_options *opt,
 	head_pairs = get_diffpairs(opt, common, head);
 	merge_pairs = get_diffpairs(opt, common, merge);
 
-	if (opt->detect_directory_renames) {
+	if ((opt->detect_directory_renames == 2) ||
+	    (opt->detect_directory_renames == 1 && !opt->call_depth)) {
 		dir_re_head = get_directory_renames(head_pairs);
 		dir_re_merge = get_directory_renames(merge_pairs);
 
t/t6043-merge-rename-directories.sh +111 -0

115 lines not on this page · raw message

Signature · 1 line
2.22.0.246.g5ddf3d502a
#2 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

Quoted 15 lines · Elijah Newren

Ever since commit 8c8e5bd6eb33 ("merge-recursive: switch directory rename detection default", 2019-04-05), the default handling with directory rename detection was to report a conflict and leave unstaged entries in the index. However, when creating a virtual merge base in the recursive case, we absolutely need a tree, and the only way a tree can be written is if we have no unstaged entries -- otherwise we hit a BUG().

There are a few fixes possible here which at least fix the BUG(), but none of them seem optimal for other reasons; see the comments with the new testcase 13e in t6043 for details (which testcase triggered a BUG() prior to this patch). As such, just opt for a very conservative and simple choice that is still relatively reasonable: have the recursive case treat 'conflict' as 'false' for opt->detect_directory_renames.

Quoted 8 lines

---

I really should introduce constants like
DETECT_DIRECTORY_RENAMES_NEVER = 0
DETECT_DIRECTORY_RENAMES_CONFLICT = 1
DETECT_DIRECTORY_RENAMES_YES = 2

How would they compare with MERGE_DIRECTORY_RENAMES_* macros I see at the tip of 'pu'? init_merge_options() seems to read one of those values from the "repo settings" and copies it to the detect_directory_renames field, so I am reading that they must be identical.

Quoted 8 lines

and then use them in the code to make it clearer, but I wanted to make the code change as simple and contained as possible given that this is built on maint, fixes a BUG() and we're already in -rc1.

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

I do agree that it is sensible to avoid doing any funky thing during the virtual base merges, whose result is much less observable (hence harder to form the right mental model in end user's head) than the outermost merge. Do we want to allow this for inner merges when the setting is 2? Wouldn't that hit the same BUG()?

Quoted 140 lines

(This fix does merge down cleanly to master, next, and pu.)

 merge-recursive.c                   |   3 +-
 t/t6043-merge-rename-directories.sh | 111 ++++++++++++++++++++++++++++
 2 files changed, 113 insertions(+), 1 deletion(-)
merge-recursive.c +2 -1
index d2e380b7ed..c7691d9b54 100644
@@ -2856,7 +2856,8 @@ static int detect_and_process_renames(struct merge_options *opt,
 	head_pairs = get_diffpairs(opt, common, head);
 	merge_pairs = get_diffpairs(opt, common, merge);
 
-	if (opt->detect_directory_renames) {
+	if ((opt->detect_directory_renames == 2) ||
+	    (opt->detect_directory_renames == 1 && !opt->call_depth)) {
 		dir_re_head = get_directory_renames(head_pairs);
 		dir_re_merge = get_directory_renames(merge_pairs);
 
t/t6043-merge-rename-directories.sh +111 -0
index 50b7543483..c966147d5d 100755
@@ -4403,4 +4403,115 @@ test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual
 	)
 '
 
+# Testcase 13e, directory rename in virtual merge base
+#
+# This testcase has a slightly different setup than all the above cases, in
+# order to include a recursive case:
+#
+#      A   C
+#      o - o
+#     / \ / \
+#  O o   X   ?
+#     \ / \ /
+#      o   o
+#      B   D
+#
+#   Commit O: a/{z,y}
+#   Commit A: b/{z,y}
+#   Commit B: a/{z,y,x}
+#   Commit C: b/{z,y,x}
+#   Commit D: b/{z,y}, a/x
+#   Expected: b/{z,y,x}  (sort of; see below for why this might not be expected)
+#
+#   NOTES: 'X' represents a virtual merge base.  With the default of
+#          directory rename detection yielding conflicts, merging A and B
+#          results in a conflict complaining about whether 'x' should be
+#          under 'a/' or 'b/'.  However, when creating the virtual merge
+#          base 'X', since virtual merge bases need to be written out as a
+#          tree, we cannot have a conflict, so some resolution has to be
+#          picked.
+#
+#          In choosing the right resolution, it's worth noting here that
+#          commits C & D are merges of A & B that choose different
+#          locations for 'x' (i.e. they resolve the conflict differently),
+#          and so it would be nice when merging C & D if git could detect
+#          this difference of opinion and report a conflict.  But the only
+#          way to do so that I can think of would be to have the virtual
+#          merge base place 'x' in some directory other than either 'a/' or
+#          'b/', which seems a little weird -- especially since it'd result
+#          in a rename/rename(1to2) conflict with a source path that never
+#          existed in any version.
+#
+#          So, for now, when directory rename detection is set to
+#          'conflict' just avoid doing directory rename detection at all in
+#          the recursive case.  This will not allow us to detect a conflict
+#          in the outer merge for this special kind of setup, but it at
+#          least avoids hitting a BUG().
+#
+test_expect_success '13e-setup: directory rename detection in recursive case' '
+	test_create_repo 13e &&
+	(
+		cd 13e &&
+
+		mkdir a &&
+		echo z >a/z &&
+		echo y >a/y &&
+		git add a &&
+		test_tick &&
+		git commit -m "O" &&
+
+		git branch O &&
+		git branch A &&
+		git branch B &&
+
+		git checkout A &&
+		git mv a/ b/ &&
+		test_tick &&
+		git commit -m "A" &&
+
+		git checkout B &&
+		echo x >a/x &&
+		git add a &&
+		test_tick &&
+		git commit -m "B" &&
+
+		git branch C A &&
+		git branch D B &&
+
+		git checkout C &&
+		test_must_fail git -c merge.directoryRenames=conflict merge B &&
+		git add b/x &&
+		test_tick &&
+		git commit -m "C" &&
+
+
+		git checkout D &&
+		test_must_fail git -c merge.directoryRenames=conflict merge A &&
+		git add b/x &&
+		mkdir a &&
+		git mv b/x a/x &&
+		test_tick &&
+		git commit -m "D"
+	)
+'
+
+test_expect_success '13e-check: directory rename detection in recursive case' '
+	(
+		cd 13e &&
+
+		git checkout --quiet D^0 &&
+
+		git -c merge.directoryRenames=conflict merge -s recursive C^0 >out 2>err &&
+
+		test_i18ngrep ! CONFLICT out &&
+		test_i18ngrep ! BUG: err &&
+		test_i18ngrep ! core.dumped err &&
+		test_must_be_empty err &&
+
+		git ls-files >paths &&
+		! grep a/x paths &&
+		grep b/x paths
+	)
+'
+
 test_done
#3 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

Quoted 27 lines · Junio C Hamano
Quoted 15 lines · Elijah Newren

Ever since commit 8c8e5bd6eb33 ("merge-recursive: switch directory rename detection default", 2019-04-05), the default handling with directory rename detection was to report a conflict and leave unstaged entries in the index. However, when creating a virtual merge base in the recursive case, we absolutely need a tree, and the only way a tree can be written is if we have no unstaged entries -- otherwise we hit a BUG().

There are a few fixes possible here which at least fix the BUG(), but none of them seem optimal for other reasons; see the comments with the new testcase 13e in t6043 for details (which testcase triggered a BUG() prior to this patch). As such, just opt for a very conservative and simple choice that is still relatively reasonable: have the recursive case treat 'conflict' as 'false' for opt->detect_directory_renames.

Quoted 8 lines

---

I really should introduce constants like
DETECT_DIRECTORY_RENAMES_NEVER = 0
DETECT_DIRECTORY_RENAMES_CONFLICT = 1
DETECT_DIRECTORY_RENAMES_YES = 2

Quoted 5 lines

How would they compare with MERGE_DIRECTORY_RENAMES_* macros I see at the tip of 'pu'? init_merge_options() seems to read one of those values from the "repo settings" and copies it to the detect_directory_renames field, so I am reading that they must be identical.

Indeed, it looks like Stolee has done my work for me -- though I would suspect that his change won't be merged down to maint, and I am assuming that my patch might be. If you don't want to merge it down to maint, I can rebase on Stolee's patch; just let me know. (And if you do want to merge it down to maint, I can send a subsequent patch to modify and adopt Stolee's naming.)

Quoted 9 lines
Quoted 8 lines

and then use them in the code to make it clearer, but I wanted to make the code change as simple and contained as possible given that this is built on maint, fixes a BUG() and we're already in -rc1.

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

Quoted 5 lines

I do agree that it is sensible to avoid doing any funky thing during the virtual base merges, whose result is much less observable (hence harder to form the right mental model in end user's head) than the outermost merge. Do we want to allow this for inner merges when the setting is 2? Wouldn't that hit the same BUG()?

No, it doesn't hit the same BUG(). Hitting the BUG() only came from having unstaged entries at the end of an inner merge, which only came from the 'conflict' setting for directory rename detection. Having directory rename detection completely on (detect the rename and accept it as the resolution), or completely off (don't detect directory renames, i.e. leave old pathname as the resolution) both have well defined resolutions.

But that doesn't completely answer your question about whether we want to have directory rename detection for the inner merges, it just suggests we have to avoid 'conflict' as a setting in that case. Let me look at the same testcase:

Quoted 10 lines
Quoted 10 lines
t/t6043-merge-rename-directories.sh +10 -0
+# Testcase 13e, directory rename in virtual merge base
+#
+# This testcase has a slightly different setup than all the above cases, in
+# order to include a recursive case:
+#
+#      A   C
+#      o - o
+#     / \ / \
+#  O o   X   ?
+#     \ / \ /
Quoted 8 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4423

Quoted 8 lines
t/t6043-merge-rename-directories.sh +8 -0
+#      o   o
+#      B   D
+#
+#   Commit O: a/{z,y}
+#   Commit A: b/{z,y}
+#   Commit B: a/{z,y,x}
+#   Commit C: b/{z,y,x}
+#   Commit D: b/{z,y}, a/x

Here, if the user has directory rename detection fully on (opt->detect_directory_renames == 2), then the fact that commits C and D resolved the merge differently is perhaps more surprising, because the default would be to just accept the rename. That means one side (commit C) is likely to have just taken the default resolution, but someone (whoever created commit D) made a manual effort to undo the directory rename. Now, if the virtual merge base doesn't do directory rename detection, then it'll match commit D which undid the directory rename detection, and result in commit C winning and using the directory rename. That seems like the wrong choice to me.

So, I think that if opt->detect_directory_renames == 2, then we really do want to use directory rename detection on the virtual merge base so that we don't silently miss someone manually undoing the directory rename detection on one side of history.

The trickier choice comes when opt->detect_directory_renames == 1 (or 'conflict') because then we don't have a sane avoid-accidentally-matching-one-side solution. I'm not sure there's a "best" choice for this case, but we should certainly at least avoid a BUG(). The simple ways to do that were to translate 'conflict' to either 'false' or 'true'. I can't give a good rationale for why I picked 'false' over 'true' in that case with this patch; either seem equally good, but 'false' has a slight performance advantage over 'true' so I went with it.

#4 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

In reply to Junio C Hamano
Quoted 5 lines · Junio C Hamano

I do agree that it is sensible to avoid doing any funky thing during the virtual base merges, whose result is much less observable (hence harder to form the right mental model in end user's head) than the outermost merge. Do we want to allow this for inner merges when the setting is 2? Wouldn't that hit the same BUG()?

Ah, actually no. 2 would make a choice of committing to one possible merge result and does not leave options to be resolved as conflicted, so we should be able to write out its decision as a tree. If we cannot form a tree object when operating with 2, we do have a bug and we do want to know about it.

#5 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

In reply to Elijah Newren
Quoted 4 lines · Elijah Newren

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

For maintenance and upcoming release, a safer "fix" to do might be to also (in addition to this patch) flip the default to no to revert to the stable state before "directory renames" was introduced, while still allowing those who want to help can explore the right fix to this codepath.

Quoted 17 lines

(This fix does merge down cleanly to master, next, and pu.)

 merge-recursive.c                   |   3 +-
 t/t6043-merge-rename-directories.sh | 111 ++++++++++++++++++++++++++++
 2 files changed, 113 insertions(+), 1 deletion(-)
merge-recursive.c +2 -1
index d2e380b7ed..c7691d9b54 100644
@@ -2856,7 +2856,8 @@ static int detect_and_process_renames(struct merge_options *opt,
 	head_pairs = get_diffpairs(opt, common, head);
 	merge_pairs = get_diffpairs(opt, common, merge);
 
-	if (opt->detect_directory_renames) {
+	if ((opt->detect_directory_renames == 2) ||
+	    (opt->detect_directory_renames == 1 && !opt->call_depth)) {
#6 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

Quoted 8 lines · Junio C Hamano
Quoted 4 lines · Elijah Newren

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

Quoted 5 lines

For maintenance and upcoming release, a safer "fix" to do might be to also (in addition to this patch) flip the default to no to revert to the stable state before "directory renames" was introduced, while still allowing those who want to help can explore the right fix to this codepath.

That might be a bit more jarring. Directory rename detection did not have this bug in git 2.18, 2.19, 2.20, or 2.21. It was new to 2.22 with the 'conflict' setting

#7 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

In reply to Elijah Newren
Quoted 27 lines · Elijah Newren

Ever since commit 8c8e5bd6eb33 ("merge-recursive: switch directory rename detection default", 2019-04-05), the default handling with directory rename detection was to report a conflict and leave unstaged entries in the index. However, when creating a virtual merge base in the recursive case, we absolutely need a tree, and the only way a tree can be written is if we have no unstaged entries -- otherwise we hit a BUG().

There are a few fixes possible here which at least fix the BUG(), but none of them seem optimal for other reasons; see the comments with the new testcase 13e in t6043 for details (which testcase triggered a BUG() prior to this patch). As such, just opt for a very conservative and simple choice that is still relatively reasonable: have the recursive case treat 'conflict' as 'false' for opt->detect_directory_renames.

---

I really should introduce constants like
DETECT_DIRECTORY_RENAMES_NEVER = 0
DETECT_DIRECTORY_RENAMES_CONFLICT = 1
DETECT_DIRECTORY_RENAMES_YES = 2
and then use them in the code to make it clearer, but I wanted to make the code change as simple and contained as possible given that this is built on maint, fixes a BUG() and we're already in -rc1.

Quoted 6 lines

I know this bug doesn't satisfy the normal criteria for making it into 2.23 (it's a bug that was present in 2.22 rather than a regression in 2.23), but given that it's a BUG() condition, I was hoping it is important and safe enough to include anyway.

(This fix does merge down cleanly to master, next, and pu.)

Thanks for picking this up and sorry I didn't end up sending anything - priority shifts on this end. :)

Quoted 10 lines
 merge-recursive.c                   |   3 +-
 t/t6043-merge-rename-directories.sh | 111 ++++++++++++++++++++++++++++
 2 files changed, 113 insertions(+), 1 deletion(-)
merge-recursive.c +0 -0
index d2e380b7ed..c7691d9b54 100644

merge-recursive.c static int detect_and_process_renames(struct merge_options *opt, · line 2859

Quoted 4 lines
merge-recursive.c +0 -1
@@ -2856,7 +2856,8 @@ static int detect_and_process_renames(struct merge_options *opt,
 	head_pairs = get_diffpairs(opt, common, head);
 	merge_pairs = get_diffpairs(opt, common, merge);
 
-	if (opt->detect_directory_renames) {

So we used to say, "If you want to apply directory renames, or mark directory renames as merge conflicts, we will go look for renames no matter what."

merge-recursive.c static int detect_and_process_renames(struct merge_options *opt, · line 2860

Quoted 2 lines
merge-recursive.c +2 -0
+	if ((opt->detect_directory_renames == 2) ||
+	    (opt->detect_directory_renames == 1 && !opt->call_depth)) {

But now we say, "If you want to apply directory renames always, or if you want to mark them as conflicts AND you aren't below the first layer of a recursive merge, then we will go look for renames."

merge-recursive.c static int detect_and_process_renames(struct merge_options *opt, · line 2862

Quoted 2 lines
merge-recursive.c +0 -0
 		dir_re_head = get_directory_renames(head_pairs);
 		dir_re_merge = get_directory_renames(merge_pairs);

That means that when we usually prefer to mark directory renames as conflicts and we are putting together a virtual ancestor, we don't try to detect renames at all, which I imagine leaves it to the top level merge to mark as conflicts for the user to resolve much later.

Quoted 64 lines
t/t6043-merge-rename-directories.sh +55 -0
index 50b7543483..c966147d5d 100755
@@ -4403,4 +4403,115 @@ test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual
 	)
 '
 
+# Testcase 13e, directory rename in virtual merge base
+#
+# This testcase has a slightly different setup than all the above cases, in
+# order to include a recursive case:
+#
+#      A   C
+#      o - o
+#     / \ / \
+#  O o   X   ?
+#     \ / \ /
+#      o   o
+#      B   D
+#
+#   Commit O: a/{z,y}
+#   Commit A: b/{z,y}
+#   Commit B: a/{z,y,x}
+#   Commit C: b/{z,y,x}
+#   Commit D: b/{z,y}, a/x
+#   Expected: b/{z,y,x}  (sort of; see below for why this might not be expected)
+#
+#   NOTES: 'X' represents a virtual merge base.  With the default of
+#          directory rename detection yielding conflicts, merging A and B
+#          results in a conflict complaining about whether 'x' should be
+#          under 'a/' or 'b/'.  However, when creating the virtual merge
+#          base 'X', since virtual merge bases need to be written out as a
+#          tree, we cannot have a conflict, so some resolution has to be
+#          picked.
+#
+#          In choosing the right resolution, it's worth noting here that
+#          commits C & D are merges of A & B that choose different
+#          locations for 'x' (i.e. they resolve the conflict differently),
+#          and so it would be nice when merging C & D if git could detect
+#          this difference of opinion and report a conflict.  But the only
+#          way to do so that I can think of would be to have the virtual
+#          merge base place 'x' in some directory other than either 'a/' or
+#          'b/', which seems a little weird -- especially since it'd result
+#          in a rename/rename(1to2) conflict with a source path that never
+#          existed in any version.
+#
+#          So, for now, when directory rename detection is set to
+#          'conflict' just avoid doing directory rename detection at all in
+#          the recursive case.  This will not allow us to detect a conflict
+#          in the outer merge for this special kind of setup, but it at
+#          least avoids hitting a BUG().
+#
+test_expect_success '13e-setup: directory rename detection in recursive case' '
+	test_create_repo 13e &&
+	(
+		cd 13e &&
+
+		mkdir a &&
+		echo z >a/z &&
+		echo y >a/y &&
+		git add a &&
+		test_tick &&

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4470

Quoted 10 lines
t/t6043-merge-rename-directories.sh +10 -0
+		git commit -m "O" &&
+
+		git branch O &&
+		git branch A &&
+		git branch B &&
+
+		git checkout A &&
+		git mv a/ b/ &&
+		test_tick &&
+		git commit -m "A" &&

So we do a directory rename in branch A...

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4476

Quoted 6 lines
t/t6043-merge-rename-directories.sh +6 -0
+
+		git checkout B &&
+		echo x >a/x &&
+		git add a &&
+		test_tick &&
+		git commit -m "B" &&

... And we add a new file to the directory in question in branch B...

So A and B will be the ones combined to make a virtual ancestor.

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4485

Quoted 9 lines
t/t6043-merge-rename-directories.sh +9 -0
+
+		git branch C A &&
+		git branch D B &&
+
+		git checkout C &&
+		test_must_fail git -c merge.directoryRenames=conflict merge B &&
+		git add b/x &&
+		test_tick &&
+		git commit -m "C" &&

Then we merge C with B which places B as a mutual ancestor of D as well as C.

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4489

Quoted 4 lines
t/t6043-merge-rename-directories.sh +4 -0
+
+
+		git checkout D &&
+		test_must_fail git -c merge.directoryRenames=conflict merge A &&

Now we do the same thing merging A with D, which means that D has ancestors B from branching and A from merge, and C has ancestors A from branching and B from merge. So D and C have two closest ancestors (criss-cross merge).

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4492

Quoted 3 lines
t/t6043-merge-rename-directories.sh +3 -0
+		git add b/x &&
+		mkdir a &&
+		git mv b/x a/x &&

Now D adds contention over a/x and b/x (which were both mentioned in the ancestry too) to induce a conflict... or is this adding a resolution which can be decided on automatically? I guess later you are looking to make sure no CONFLICT still exists in the output, so you must be resolving the conflict here?

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4504

Quoted 12 lines
t/t6043-merge-rename-directories.sh +12 -0
+		test_tick &&
+		git commit -m "D"
+	)
+'
+
+test_expect_success '13e-check: directory rename detection in recursive case' '
+	(
+		cd 13e &&
+
+		git checkout --quiet D^0 &&
+
+		git -c merge.directoryRenames=conflict merge -s recursive C^0 >out 2>err &&

Now we finally do the recursive merge - C and D have equally likely ancestors A and B, and A and B have a rename conflict.

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4507

Quoted 3 lines
t/t6043-merge-rename-directories.sh +3 -0
+
+		test_i18ngrep ! CONFLICT out &&
+		test_i18ngrep ! BUG: err &&

The BUG is gone. But should it not use i18ngrep? BUG() isn't localized.

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4512

Quoted 5 lines
t/t6043-merge-rename-directories.sh +5 -0
+		test_i18ngrep ! core.dumped err &&
+		test_must_be_empty err &&
+
+		git ls-files >paths &&
+		! grep a/x paths &&

Finally, make sure that a/x has been truly disappeared...

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4513

Quoted 1 line
t/t6043-merge-rename-directories.sh +1 -0
+		grep b/x paths

...and b/x is the only x left standing.

Quoted 7 lines
t/t6043-merge-rename-directories.sh +3 -0
+	)
+'
+
 test_done
Signature · 1 line
2.22.0.246.g5ddf3d502a
#8 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

Quoted 9 lines · Emily Shaffer
Quoted 7 lines · Elijah Newren

Ever since commit 8c8e5bd6eb33 ("merge-recursive: switch directory rename detection default", 2019-04-05), the default handling with directory rename detection was to report a conflict and leave unstaged entries in the index. However, when creating a virtual merge base in the recursive case, we absolutely need a tree, and the only way a tree can be written is if we have no unstaged entries -- otherwise we hit a BUG().

...

Quoted 2 lines

Thanks for picking this up and sorry I didn't end up sending anything - priority shifts on this end. :)

I totally understand.

...

Quoted 23 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4485

Quoted 5 lines
t/t6043-merge-rename-directories.sh +5 -0
+             git checkout C &&
+             test_must_fail git -c merge.directoryRenames=conflict merge B &&
+             git add b/x &&
+             test_tick &&
+             git commit -m "C" &&

Then we merge C with B which places B as a mutual ancestor of D as well as C.

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4489

Quoted 4 lines
t/t6043-merge-rename-directories.sh +4 -0
+
+
+             git checkout D &&
+             test_must_fail git -c merge.directoryRenames=conflict merge A &&

Now we do the same thing merging A with D, which means that D has ancestors B from branching and A from merge, and C has ancestors A from branching and B from merge. So D and C have two closest ancestors (criss-cross merge).

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4492

Quoted 3 lines
t/t6043-merge-rename-directories.sh +3 -0
+             git add b/x &&
+             mkdir a &&
+             git mv b/x a/x &&
Quoted 5 lines

Now D adds contention over a/x and b/x (which were both mentioned in the ancestry too) to induce a conflict... or is this adding a resolution which can be decided on automatically? I guess later you are looking to make sure no CONFLICT still exists in the output, so you must be resolving the conflict here?

Yes, we are resolving the conflict for D by choosing to reject the directory rename, placing 'x' in a/x instead of b/x.

When merge.directoryRenames=conflict, it'll make only b/x be present in the working directory and mark it as conflicted (i.e. assume the directory rename is probably the right resolution, but print a warning and mark the index as needing to be updated to verify -- this allows e.g. "git add -u" to do the "right thing"). For commit C, we just did a "git add b/x" to accept the directory rename. For commit D, we wanted to say we didn't want the directory rename which you'd first guess would be "git mv b/x a/x" BUT: (1) a/x has unstaged entries which will cause git-mv to fail, and (2) directory a/ didn't exist -- both of these issues had to be corrected before running git-mv.

Quoted 16 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4504

Quoted 12 lines
t/t6043-merge-rename-directories.sh +12 -0
+             test_tick &&
+             git commit -m "D"
+     )
+'
+
+test_expect_success '13e-check: directory rename detection in recursive case' '
+     (
+             cd 13e &&
+
+             git checkout --quiet D^0 &&
+
+             git -c merge.directoryRenames=conflict merge -s recursive C^0 >out 2>err &&

Now we finally do the recursive merge - C and D have equally likely ancestors A and B, and A and B have a rename conflict.

Quoted 5 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4507

Quoted 3 lines
t/t6043-merge-rename-directories.sh +3 -0
+
+             test_i18ngrep ! CONFLICT out &&
+             test_i18ngrep ! BUG: err &&

The BUG is gone. But should it not use i18ngrep? BUG() isn't localized.

Technically, yes, you're right. However, this line's purpose isn't correctness of the test but documentation for the person reading the testcase about what it's real original purpose was; my real test was the "test_must_be_empty err" check I have below it, but I added this line just to document the intent better. I kind of like the "CONFLICT" and "BUG" lines looking similar just so the reader of the testcase doesn't have to try to reason through why they are different, although I guess it does present the problem that more careful readers like yourself might do a double take.

If folks find it more readable to use regular grep instead of test_i18ngrep, I can change this line as well as the "core dumped" check immediately below over to regular grep.

Quoted 9 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4512

Quoted 5 lines
t/t6043-merge-rename-directories.sh +5 -0
+             test_i18ngrep ! core.dumped err &&
+             test_must_be_empty err &&
+
+             git ls-files >paths &&
+             ! grep a/x paths &&

Finally, make sure that a/x has been truly disappeared...

Quoted 1 line

+ grep b/x paths

...and b/x is the only x left standing.

Thanks for taking a look. :-)

author Reviewed-by
#9 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

Quoted 16 lines
Quoted 5 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4507

Quoted 3 lines
t/t6043-merge-rename-directories.sh +3 -0
+
+             test_i18ngrep ! CONFLICT out &&
+             test_i18ngrep ! BUG: err &&

The BUG is gone. But should it not use i18ngrep? BUG() isn't localized.

Technically, yes, you're right. However, this line's purpose isn't correctness of the test but documentation for the person reading the testcase about what it's real original purpose was; my real test was the "test_must_be_empty err" check I have below it, but I added this line just to document the intent better. I kind of like the "CONFLICT" and "BUG" lines looking similar just so the reader of the testcase doesn't have to try to reason through why they are different, although I guess it does present the problem that more careful readers like yourself might do a double take.

Quoted 4 lines

If folks find it more readable to use regular grep instead of test_i18ngrep, I can change this line as well as the "core dumped" check immediately below over to regular grep.

It's fine by me, I see your point.

Quoted 1 line

Thanks for taking a look. :-)

#10 lore raw

Re: [PATCH 1/1] merge-recursive: avoid directory rename detection in recursive case

In reply to Elijah Newren
Quoted 5 lines · Elijah Newren
Quoted 4 lines

t/t6043-merge-rename-directories.sh test_expect_success '13d-check(info): messages for rename/rename(1to1) via dual · line 4507

Quoted 2 lines
t/t6043-merge-rename-directories.sh +2 -0
+             test_i18ngrep ! CONFLICT out &&
+             test_i18ngrep ! BUG: err &&

The BUG is gone. But should it not use i18ngrep? BUG() isn't localized.

Quoted 9 lines

Technically, yes, you're right. However, this line's purpose isn't correctness of the test but documentation for the person reading the testcase about what it's real original purpose was; my real test was the "test_must_be_empty err" check I have below it, but I added this line just to document the intent better. I kind of like the "CONFLICT" and "BUG" lines looking similar just so the reader of the testcase doesn't have to try to reason through why they are different, although I guess it does present the problem that more careful readers like yourself might do a double take.

I think it would be better to drop the grep for BUG entirely.

Not BUG()-ing should be something we implicitly assume for all commands, and checking the exit code already covers that[1]. I don't think we should be cluttering up every test, even ones that are in response to a BUG(), with redundant checks. If you really want to document it further, a comment can do that without incurring extra run-time overhead. But I think in this case that your existing comments and commit message cover it quite well.

-Peff

[1] There are cases where there's a crash in a sub-process, but in that case the failure should be surfaced in the way the test is written. It is here, and I'd argue that any case where it isn't probably ought to be rewritten (because you're missing not just BUG()s, but probably die()).