Re: "git log --first-parent" shows parents that are not first

しらいしななこ  [off-list ref] writes:

quoted

The result given by "git log --first-parent" ('next' version) is
unexpected to me.

  % git rev-parse origin/next
  4eddac518225621c3e4f7285beb879d2b4bad38a
  % git log --abbrev-commit --pretty=oneline --first-parent origin/next^..origin/next
  4eddac5... Merge branch 'master' into next
  1f8115b... Merge branch 'maint'
...
  921177f... Documentation: improve "add", "pull" and "format-patch" examples
  c904bf3... Be more careful with objects directory permissions on clone

I asked for the log between one commit before the tip of "origin/next"
and the tip of the branch, following only the first-parent links.
v1.5.5 is not broken and shows the expected result:

  % ~/git-v1.5.5/bin/git log --abbrev-commit --pretty=oneline --first-parent origin/next^..origin/next
  4eddac5... Merge branch 'master' into next

Could you please revert d9c292e8bbd51c84cb9ecd86cb89b8a1b35a2a82?  With
that patch reverted from 'next', the problem disappears.

That's d9c292e (Simplify and fix --first-parent implementation,
2008-04-27) by Stephen.

I know that the alleged "fix" works around a corner-case, a fast-forward
situation that was artificually recorded as a merge, but if the "cure"
breaks a normal case like this, it is worse than the disease.

Stephen, do you have a fix?

In add_parents_to_list, if any parent of a revision had already been
SEEN, the current code would continue with the next parent. But if the
first parent has been SEEN and --first-parent has been specified we need
to break, not continue.

Signed-off-by: Lars Hjemli <redacted>
---
 revision.c |    5 ++++-
 1 files changed, 4 insertions(+), 1 deletions(-)

diff --git a/revision.c b/revision.c
index 44d780b..974ad10 100644
--- a/revision.c
+++ b/revision.c

@@ -468,8 +468,11 @@ static int add_parents_to_list(struct rev_info *revs, struct commit *commit, str
 		if (parse_commit(p) < 0)
 			return -1;
 		p->object.flags |= left_flag;
-		if (p->object.flags & SEEN)
+		if (p->object.flags & SEEN) {
+			if (revs->first_parent_only)
+				break;
 			continue;
+		}
 		p->object.flags |= SEEN;
 		insert_by_date(p, list);
 		if(revs->first_parent_only)

-- 
1.5.5.1.148.g8ee22.dirty

In add_parents_to_list, if any parent of a revision had already been
SEEN, the current code would continue with the next parent, skipping
the test for --first-parent. This patch inverts the test for SEEN so
that the test for --first-parent is always performed.

Signed-off-by: Lars Hjemli <redacted>
---
This is a slightly different approach which I think is less ugly.

 revision.c |    8 ++++----
 1 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/revision.c b/revision.c
index 44d780b..2fc26b8 100644
--- a/revision.c
+++ b/revision.c

@@ -468,10 +468,10 @@ static int add_parents_to_list(struct rev_info *revs, struct commit *commit, str
 		if (parse_commit(p) < 0)
 			return -1;
 		p->object.flags |= left_flag;
-		if (p->object.flags & SEEN)
-			continue;
-		p->object.flags |= SEEN;
-		insert_by_date(p, list);
+		if (!(p->object.flags & SEEN)) {
+			p->object.flags |= SEEN;
+			insert_by_date(p, list);
+		}
 		if(revs->first_parent_only)
 			break;
 	}

-- 
1.5.5.1.148.g8ee22.dirty

Lars Hjemli wrote:

In add_parents_to_list, if any parent of a revision had already been
SEEN, the current code would continue with the next parent, skipping
the test for --first-parent. This patch inverts the test for SEEN so
that the test for --first-parent is always performed.

Let's put it this way:
- If there would have been only one path to any particular point in the
  tree, then the --first-parent flag makes no differences, because the
  tree wouldn't contain any merges to begin with.
- If a tree contains *any* merges (i.e. a commit with multiple parents),
  then there are always multiple paths to some common ancestor, and
  therefore depending on which path you travel up first, you sometimes get
  clashes with the SEEN flag (unpredictable by definition).
- It would seem logical and sufficient to avoid this unpredictability by
  utilising the --first-parent flag to present and walk a tree of commits
  AS IF there were no merges.
- My original patch did just that, it simplified the code to make sure
  that all other parents beside the first parent are ignored when
  walking the tree.
- Your code now doesn't simplify the (IMO) convoluted walk, and still
  marks things as seen, even though in the first-parent case, these
  commits are not really seen at all.  It implies that your code
  generates differing output, depending on the merges present.
- The question now is, do we want the output of --first-parent to be
  immutable with respect to merges being present (but hidden from sight
  during a --first-parent run), or do we want the output of
  --first-parent to actually change depending on variations in parents
  other than the first parent?

I'd say it's better to keep the code simpler, and to make sure the
output does *not* depend on any parents other than the first (as
implemented in my original patch).

This is a slightly different approach which I think is less ugly.

Your patch is smaller, and therefore (perhaps) less ugly; the resulting
code and logic of my original patch is simpler (IMHO), and therefore
cleaner (but it all depends on (the lack of) consensus over the points above).
-- 
Sincerely,                                                          srb@cuci.nl
           Stephen R. van den Berg.

"If I had to live my life again, I'd make the same mistakes, only sooner."

On Tue, May 13, 2008 at 10:15 PM, Stephen R. van den Berg [off-list ref] wrote:

Lars Hjemli wrote:
 >In add_parents_to_list, if any parent of a revision had already been
 >SEEN, the current code would continue with the next parent, skipping
 >the test for --first-parent. This patch inverts the test for SEEN so
 >that the test for --first-parent is always performed.

 Let's put it this way:
 - If there would have been only one path to any particular point in the
  tree, then the --first-parent flag makes no differences, because the
  tree wouldn't contain any merges to begin with.

True

 - If a tree contains *any* merges (i.e. a commit with multiple parents),
  then there are always multiple paths to some common ancestor, and
  therefore depending on which path you travel up first, you sometimes get
  clashes with the SEEN flag (unpredictable by definition).

True

 - It would seem logical and sufficient to avoid this unpredictability by
  utilising the --first-parent flag to present and walk a tree of commits
  AS IF there were no merges.

True

 - My original patch did just that, it simplified the code to make sure
  that all other parents beside the first parent are ignored when
  walking the tree.

Except for the case where the first parent had been already SEEN; then
it would continue to test the next parents until one was found which
was not already SEEN and _that_ parent would be treated as if it was
first. And as Nanako showed, a simple `git rev-list HEAD^..HEAD` marks
both HEAD and HEAD^ as seen. When combined with --first-parent, the
result (with your patch) is that HEAD^2 is treated as the first
parent. With my patch on top of yours, the walk stops as HEAD^, which
is what we probably both want.

 - Your code now doesn't simplify the (IMO) convoluted walk, and still
  marks things as seen, even though in the first-parent case, these
  commits are not really seen at all.  It implies that your code
  generates differing output, depending on the merges present.

I don't think so. My code should neither follow nor mark as SEEN any
parent but the first (but I could obviously be wrong).

 - The question now is, do we want the output of --first-parent to be
  immutable with respect to merges being present (but hidden from sight
  during a --first-parent run), or do we want the output of
  --first-parent to actually change depending on variations in parents
  other than the first parent?

 I'd say it's better to keep the code simpler, and to make sure the
 output does *not* depend on any parents other than the first (as
 implemented in my original patch).

I agree with your reasoning, and your patch with mine on top seems to
achieve that goal.

--
larsh

"Lars Hjemli" [off-list ref] writes:

quoted

 - My original patch did just that, it simplified the code to make sure
  that all other parents beside the first parent are ignored when
  walking the tree.

Except for the case where the first parent had been already SEEN; then
it would continue to test the next parents until one was found which
was not already SEEN and _that_ parent would be treated as if it was
first. And as Nanako showed, a simple `git rev-list HEAD^..HEAD` marks
both HEAD and HEAD^ as seen. When combined with --first-parent, the
result (with your patch) is that HEAD^2 is treated as the first
parent. With my patch on top of yours, the walk stops as HEAD^, which
is what we probably both want.

quoted

 - Your code now doesn't simplify the (IMO) convoluted walk, and still
  marks things as seen, even though in the first-parent case, these
  commits are not really seen at all.  It implies that your code
  generates differing output, depending on the merges present.

I don't think so. My code should neither follow nor mark as SEEN any
parent but the first (but I could obviously be wrong).

A major part of the "convoluted walk" is the (il-)logic that skipped
earlier SEEN parents and treated the first unseen one as if it was the
first parent, which is not exactly Stephen's fault.  It was placed by
yours truly in the very original code but it was done without much
thought.

I think your patch is the correct fix for that convolution, regardless of
the traversal order stability issue Stephen mentions.

Junio C Hamano wrote:

"Lars Hjemli" [off-list ref] writes:

quoted

quoted

 - Your code now doesn't simplify the (IMO) convoluted walk, and still
  marks things as seen, even though in the first-parent case, these
  commits are not really seen at all.  It implies that your code
  generates differing output, depending on the merges present.

quoted

I don't think so. My code should neither follow nor mark as SEEN any
parent but the first (but I could obviously be wrong).

A major part of the "convoluted walk" is the (il-)logic that skipped
earlier SEEN parents and treated the first unseen one as if it was the
first parent, which is not exactly Stephen's fault.  It was placed by
yours truly in the very original code but it was done without much
thought.

I think your patch is the correct fix for that convolution, regardless of
the traversal order stability issue Stephen mentions.

I agree that Lars' patch prevents parts of the tree to go "dark" (so did
my patch).
However, without either patch, that implies that the current --first-parent
code has a high probability of obscuring parts of the tree depending on
traversing order (in any tree which contains at least one merge).

So, I'd say, since the current code does not and cannot work reliably
for anyone specifically using --first-parent (with every merge
encountered, the probability of correctness is multiplied by 0.5 at
most/least), you are going to do them a favour anyway by fixing the code,
then why not simplify the convolution and make the code rock-steady (and
implement my patch)?

Anyone using --first-parent in production now has an embarrassingly high
probability of missing commits in his generated lists (I know that I
noticed the problem within 5 minutes from actually trying to use the
flag to get meaningful output).  So fixing and simplifying the code now
is rather unlikely to create any more surprises than the current code
already presents to existing users (if any).
-- 
Sincerely,                                                          srb@cuci.nl
           Stephen R. van den Berg.

What if there were no hypothetical questions?

On Wed, May 14, 2008 at 12:34 PM, Stephen R. van den Berg [off-list ref] wrote:

 So, I'd say, since the current code does not and cannot work reliably
 for anyone specifically using --first-parent (with every merge
 encountered, the probability of correctness is multiplied by 0.5 at
 most/least), you are going to do them a favour anyway by fixing the code,
 then why not simplify the convolution and make the code rock-steady (and
 implement my patch)?

The current 'next' branch in git.git contains your patch with my fixup
on top and I believe this fixes _both_ the original issue with
first-parent (thanks to your patch) and the issue Nanako discovered
(thanks to my patch). Am I missing something?

-- 
larsh

Lars Hjemli wrote:

On Wed, May 14, 2008 at 12:34 PM, Stephen R. van den Berg [off-list ref] wrote:

quoted

 So, I'd say, since the current code does not and cannot work reliably
 for anyone specifically using --first-parent (with every merge
 encountered, the probability of correctness is multiplied by 0.5 at
 most/least), you are going to do them a favour anyway by fixing the code,
 then why not simplify the convolution and make the code rock-steady (and
 implement my patch)?

The current 'next' branch in git.git contains your patch with my fixup
on top and I believe this fixes _both_ the original issue with
first-parent (thanks to your patch) and the issue Nanako discovered
(thanks to my patch). Am I missing something?

Probably not.  I didn't check 'next' yet, since neither mine nor your
patch had been Acked on the list (I guess it shows that I don't know the
procedures here all too well yet).
-- 
Sincerely,                                                          srb@cuci.nl
           Stephen R. van den Berg.

What if there were no hypothetical questions?