[2/2] line-log: avoid unnecessary full tree diffs
diff mbox series

Message ID 20190821110424.18184-3-szeder.dev@gmail.com
State New
Headers show
Series
  • line-log: avoid unnecessary full tree diffs
Related show

Commit Message

SZEDER Gábor Aug. 21, 2019, 11:04 a.m. UTC
With rename detection enabled the line-level log is able to trace the
evolution of line ranges across whole-file renames [1].  Alas, to
achieve that it uses the diff machinery very inefficiently, making the
operation very slow [2].  And since rename detection is enabled by
default, the line-level log is very slow by default.

When the line-level log processes a commit with rename detection
enabled, it currently does the following (see queue_diffs()):

  1. Computes a full tree diff between the commit and (one of) its
     parent(s), i.e. invokes diff_tree_oid() with an empty
     'diffopt->pathspec'.
  2. Checks whether any paths in the line ranges were modified.
  3. Checks whether any modified paths in the line ranges are missing
     in the parent commit's tree.
  4. If there is such a missing path, then calls diffcore_std() to
     figure out whether the path was indeed renamed based on the
     previously computed full tree diff.
  5. Continues doing stuff that are unrelated to the slowness.

So basically the line-level log computes a full tree diff for each
commit-parent pair in step (1) to be used for rename detection in step
(4) in the off chance that an interesting path is missing from the
parent.

Avoid these expensive and mostly unnecessary full tree diffs by
limiting the diffs to paths in the line ranges.  This is much cheaper,
and makes step (2) unnecessary.  If it turns out that an interesting
path is missing from the parent, then fall back and compute a full
tree diff, so the rename detection will still work.

Care must be taken when to update the pathspec used to limit the diff
in case of renames.  A path might be renamed on one branch and
modified on several parallel running branches, and while processing
commits on these branches the line-level log might have to alternate
between looking at a path's new and old name.  However, at any one
time there is only a single 'diffopt->pathspec'.

So add a step (0) to the above to ensure that the paths in the
pathspec match the paths in the line ranges associated with the
currently processed commit, and re-parse the pathspec from the paths
in the line ranges if they differ.

The new test cases include a specially crafted piece of history with
two merged branches and two files, where each branch modifies both
files, renames on of them, and then modifies both again.  Then two
separate 'git log -L' invocations check the line-level log of each of
those two files, which ensures that at least one of those invocations
have to do that back-and-forth between the file's old and new name (no
matter which branch is traversed first).  't/t4211-line-log.sh'
already contains two tests involving renames, they don't don't trigger
this back-and-forth.

Avoiding these unnecessary full tree diffs can have huge impact on
performance, especially in big repositories with big trees and mergy
history.  Tracing the evolution of a function through the whole
history:

  # git.git
  $ time git --no-pager log -L:read_alternate_refs:sha1-file.c v2.23.0

  Before:

    real    0m8.874s
    user    0m8.816s
    sys     0m0.057s

  After:

    real    0m2.516s
    user    0m2.456s
    sys     0m0.060s

  # linux.git
  $ time ~/src/git/git --no-pager log \
    -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2

  Before:

    real    3m50.033s
    user    3m48.041s
    sys     0m0.300s

  After:

    real    0m2.599s
    user    0m2.466s
    sys     0m0.157s

That's just over 88x speedup.

[1] Line-level log's rename following is quite similar to 'git log
    --follow path', with the notable differences that it does handle
    multiple paths at once as well, and that it doesn't show the
    commit performing the rename if it's an exact rename.

[2] This slowness might not have been apparent initially, because back
    when the line-level log feature was introduced rename detection
    was not yet enabled by default; 12da1d1f6f (Implement line-history
    search (git log -L), 2013-03-28) and 5404c116aa (diff: activate
    diff.renames by default, 2016-02-25).

Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
---
 line-log.c          | 43 ++++++++++++++++++++----
 t/t4211-line-log.sh | 82 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 118 insertions(+), 7 deletions(-)

Comments

Derrick Stolee Aug. 21, 2019, 3:53 p.m. UTC | #1
On 8/21/2019 7:04 AM, SZEDER Gábor wrote:
> With rename detection enabled the line-level log is able to trace the
> evolution of line ranges across whole-file renames [1].  Alas, to
> achieve that it uses the diff machinery very inefficiently, making the
> operation very slow [2].  And since rename detection is enabled by
> default, the line-level log is very slow by default.
> 
> When the line-level log processes a commit with rename detection
> enabled, it currently does the following (see queue_diffs()):
> 
>   1. Computes a full tree diff between the commit and (one of) its
>      parent(s), i.e. invokes diff_tree_oid() with an empty
>      'diffopt->pathspec'.
>   2. Checks whether any paths in the line ranges were modified.
>   3. Checks whether any modified paths in the line ranges are missing
>      in the parent commit's tree.
>   4. If there is such a missing path, then calls diffcore_std() to
>      figure out whether the path was indeed renamed based on the
>      previously computed full tree diff.
>   5. Continues doing stuff that are unrelated to the slowness.
> 
> So basically the line-level log computes a full tree diff for each
> commit-parent pair in step (1) to be used for rename detection in step
> (4) in the off chance that an interesting path is missing from the
> parent.
> 
> Avoid these expensive and mostly unnecessary full tree diffs by
> limiting the diffs to paths in the line ranges.  This is much cheaper,
> and makes step (2) unnecessary.  If it turns out that an interesting
> path is missing from the parent, then fall back and compute a full
> tree diff, so the rename detection will still work.

I applied your patches and tried them on our VFS-enabled version of Git
(see [1]). Unfortunately, the new logic is still triggering rename
detection, as measured by the number of objects being downloaded.

[1] https://github.com/microsoft/git/pull/182

My *guess* is that the repo has a lot of merge commits, and for many
of those, the file does not exist in the first parent. Since we are
essentially doing a --full-history, this means that edge tries a
rename detection. If we used the file-history simplification route of
traveling along a treesame edge instead of caring about both parents,
then maybe this would be avoided.

I could also be completely wrong about how this line-log code works
with regards to --full-history.

> Care must be taken when to update the pathspec used to limit the diff
> in case of renames.  A path might be renamed on one branch and
> modified on several parallel running branches, and while processing
> commits on these branches the line-level log might have to alternate
> between looking at a path's new and old name.  However, at any one
> time there is only a single 'diffopt->pathspec'.
> 
> So add a step (0) to the above to ensure that the paths in the
> pathspec match the paths in the line ranges associated with the
> currently processed commit, and re-parse the pathspec from the paths
> in the line ranges if they differ.
> 
> The new test cases include a specially crafted piece of history with
> two merged branches and two files, where each branch modifies both
> files, renames on of them, and then modifies both again.  Then two
> separate 'git log -L' invocations check the line-level log of each of
> those two files, which ensures that at least one of those invocations
> have to do that back-and-forth between the file's old and new name (no
> matter which branch is traversed first).  't/t4211-line-log.sh'
> already contains two tests involving renames, they don't don't trigger
> this back-and-forth.
> 
> Avoiding these unnecessary full tree diffs can have huge impact on
> performance, especially in big repositories with big trees and mergy
> history.  Tracing the evolution of a function through the whole
> history:
> 
>   # git.git
>   $ time git --no-pager log -L:read_alternate_refs:sha1-file.c v2.23.0
> 
>   Before:
> 
>     real    0m8.874s
>     user    0m8.816s
>     sys     0m0.057s
> 
>   After:
> 
>     real    0m2.516s
>     user    0m2.456s
>     sys     0m0.060s
> 
>   # linux.git
>   $ time ~/src/git/git --no-pager log \
>     -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2
> 
>   Before:
> 
>     real    3m50.033s
>     user    3m48.041s
>     sys     0m0.300s
> 
>   After:
> 
>     real    0m2.599s
>     user    0m2.466s
>     sys     0m0.157s
> 
> That's just over 88x speedup.

These performance numbers are great! Please don't let my complaints of
"it doesn't work for my particularly bad example" be a deterrent to this
change. If I figure out what is going on in my case, then I can create
an update on top of your changes.

> diff --git a/line-log.c b/line-log.c
> index fddd91f060..9010e00950 100644
> --- a/line-log.c
> +++ b/line-log.c
> @@ -737,6 +737,22 @@ static struct line_log_data *lookup_line_range(struct rev_info *revs,
>  	return ret;
>  }
>  
> +static int same_paths_in_pathspec_and_range(struct pathspec *pathspec,
> +					    struct line_log_data *range)
> +{
> +	int i;
> +	struct line_log_data *r;
> +
> +	for (i = 0, r = range; i < pathspec->nr && r; i++, r = r->next)
> +		if (strcmp(pathspec->items[i].match, r->path))
> +			return 0;
> +	if (i < pathspec->nr || r)
> +		/* different number of pathspec items and ranges */
> +		return 0;
> +
> +	return 1;
> +}

This method is easy to digest. Looks correct.

> @@ -762,8 +778,7 @@ void line_log_init(struct rev_info *rev, const char *prefix, struct string_list
>  	range = parse_lines(rev->diffopt.repo, commit, prefix, args);
>  	add_line_range(rev, commit, range);
>  
> -	if (!rev->diffopt.detect_rename)
> -		parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
> +	parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
>  }

So we always parse the pathspec, even if we don't do detect renames.

> @@ -821,15 +836,29 @@ static void queue_diffs(struct line_log_data *range,
>  			struct diff_queue_struct *queue,
>  			struct commit *commit, struct commit *parent)
>  {
> +	struct object_id *tree_oid, *parent_tree_oid;
> +
>  	assert(commit);
>  
> +	tree_oid = get_commit_tree_oid(commit);
> +	parent_tree_oid = parent ? get_commit_tree_oid(parent) : NULL;
> +
> +	if (opt->detect_rename &&
> +	    !same_paths_in_pathspec_and_range(&opt->pathspec, range)) {
> +		clear_pathspec(&opt->pathspec);
> +		parse_pathspec_from_ranges(&opt->pathspec, range);
> +	}

If we are detecting renames and our pathspec is not up-to-date with the
range, then clear the pathspec and reparse. Makes sense.

>  	DIFF_QUEUE_CLEAR(&diff_queued_diff);
> -	diff_tree_oid(parent ? get_commit_tree_oid(parent) : NULL,
> -		      get_commit_tree_oid(commit), "", opt);
> +	diff_tree_oid(parent_tree_oid, tree_oid, "", opt);

(I rearranged a pair of -/+ lines in the diff to highlight this change.)

Makes sense, parent_tree_oid above was set using the same conditional.

> -	if (opt->detect_rename) {
> +	if (opt->detect_rename && diff_might_be_rename()) {

Here is the crux of the matter: diff_might_be_rename() can prevent
the full tree diff.

> +		/* must look at the full tree diff to detect renames */
> +		clear_pathspec(&opt->pathspec);
> +		DIFF_QUEUE_CLEAR(&diff_queued_diff);
> +
> +		diff_tree_oid(parent_tree_oid, tree_oid, "", opt);
> +
>  		filter_diffs_for_paths(range, 1);
> -		if (diff_might_be_rename())
> -			diffcore_std(opt);
> +		diffcore_std(opt);
>  		filter_diffs_for_paths(range, 0);
>  	}

So before, diff_might_be_rename() already prevented diffcore_std(), but
now it also prevents clearing the pathspec. diff_might_be_rename() has
a simple implementation:

static inline int diff_might_be_rename(void)
{
        int i;
        for (i = 0; i < diff_queued_diff.nr; i++)
                if (!DIFF_FILE_VALID(diff_queued_diff.queue[i]->one)) {
                        /* fprintf(stderr, "diff_might_be_rename found creation of: %s\n", */
                        /*      diff_queued_diff.queue[i]->two->path); */
                        return 1;
                }
        return 0;
}

So yes, it is triggered by any path appearing in the child but not
a parent.

>  	move_diff_queue(queue, &diff_queued_diff);
> diff --git a/t/t4211-line-log.sh b/t/t4211-line-log.sh
> index 1db7bd0f59..8319163744 100755
> --- a/t/t4211-line-log.sh
> +++ b/t/t4211-line-log.sh
> @@ -132,4 +132,86 @@ test_expect_success '--raw is forbidden' '
>  	test_must_fail git log -L1,24:b.c --raw
>  '
>  
> +test_expect_success 'setup for checking fancy rename following' '
> +	git checkout --orphan moves-start &&
> +	git reset --hard &&
> +
> +	printf "%s\n"    12 13 14 15      b c d e   >file-1 &&
> +	printf "%s\n"    22 23 24 25      B C D E   >file-2 &&
> +	git add file-1 file-2 &&
> +	test_tick &&
> +	git commit -m "Add file-1 and file-2" &&
> +	oid_add_f1_f2=$(git rev-parse --short HEAD) &&
> +
> +	git checkout -b moves-main &&
> +	printf "%s\n" 11 12 13 14 15      b c d e   >file-1 &&
> +	git commit -a -m "Modify file-1 on main" &&
> +	oid_mod_f1_main=$(git rev-parse --short HEAD) &&
> +
> +	printf "%s\n" 21 22 23 24 25      B C D E   >file-2 &&
> +	git commit -a -m "Modify file-2 on main #1" &&
> +	oid_mod_f2_main_1=$(git rev-parse --short HEAD) &&
> +
> +	git mv file-1 renamed-1 &&
> +	git commit -m "Rename file-1 to renamed-1 on main" &&
> +
> +	printf "%s\n" 11 12 13 14 15      b c d e f >renamed-1 &&
> +	git commit -a -m "Modify renamed-1 on main" &&
> +	oid_mod_r1_main=$(git rev-parse --short HEAD) &&
> +
> +	printf "%s\n" 21 22 23 24 25      B C D E F >file-2 &&
> +	git commit -a -m "Modify file-2 on main #2" &&
> +	oid_mod_f2_main_2=$(git rev-parse --short HEAD) &&
> +
> +	git checkout -b moves-side moves-start &&
> +	printf "%s\n"    12 13 14 15 16   b c d e   >file-1 &&
> +	git commit -a -m "Modify file-1 on side #1" &&
> +	oid_mod_f1_side_1=$(git rev-parse --short HEAD) &&
> +
> +	printf "%s\n"    22 23 24 25 26   B C D E   >file-2 &&
> +	git commit -a -m "Modify file-2 on side" &&
> +	oid_mod_f2_side=$(git rev-parse --short HEAD) &&
> +
> +	git mv file-2 renamed-2 &&
> +	git commit -m "Rename file-2 to renamed-2 on side" &&
> +
> +	printf "%s\n"    12 13 14 15 16 a b c d e   >file-1 &&
> +	git commit -a -m "Modify file-1 on side #2" &&
> +	oid_mod_f1_side_2=$(git rev-parse --short HEAD) &&
> +
> +	printf "%s\n"    22 23 24 25 26 A B C D E   >renamed-2 &&
> +	git commit -a -m "Modify renamed-2 on side" &&
> +	oid_mod_r2_side=$(git rev-parse --short HEAD) &&
> +
> +	git checkout moves-main &&
> +	git merge moves-side &&
> +	oid_merge=$(git rev-parse --short HEAD)
> +'
> +
> +test_expect_success 'fancy rename following #1' '
> +	cat >expect <<-EOF &&
> +	$oid_merge Merge branch '\''moves-side'\'' into moves-main
> +	$oid_mod_f1_side_2 Modify file-1 on side #2
> +	$oid_mod_f1_side_1 Modify file-1 on side #1
> +	$oid_mod_r1_main Modify renamed-1 on main
> +	$oid_mod_f1_main Modify file-1 on main
> +	$oid_add_f1_f2 Add file-1 and file-2
> +	EOF
> +	git log -L1:renamed-1 --oneline --no-patch >actual &&
> +	test_cmp expect actual
> +'
> +
> +test_expect_success 'fancy rename following #2' '
> +	cat >expect <<-EOF &&
> +	$oid_merge Merge branch '\''moves-side'\'' into moves-main
> +	$oid_mod_r2_side Modify renamed-2 on side
> +	$oid_mod_f2_side Modify file-2 on side
> +	$oid_mod_f2_main_2 Modify file-2 on main #2
> +	$oid_mod_f2_main_1 Modify file-2 on main #1
> +	$oid_add_f1_f2 Add file-1 and file-2
> +	EOF
> +	git log -L1:renamed-2 --oneline --no-patch >actual &&
> +	test_cmp expect actual
> +'

These look to be suitably interesting test cases. Thanks!

Looking at your patch, I can mostly follow the logic, but my
unfamiliarity with the code is keeping me from being confident
in full understanding. I hope someone who is familiar can
chime in, because I really like the direction here.

Hopefully I will have time in the next few weeks to revisit
this and work to resolve my abnormal case.

-Stolee
Junio C Hamano Aug. 21, 2019, 5:29 p.m. UTC | #2
SZEDER Gábor <szeder.dev@gmail.com> writes:

> So basically the line-level log computes a full tree diff for each
> commit-parent pair in step (1) to be used for rename detection in step
> (4) in the off chance that an interesting path is missing from the
> parent.

Good explanation.  As we are not supporting swapping of two (or
more) files, we only need rename processing when a path we have been
inspecting disappears, at which point it is worth spending cycles to
see where the path used to be in the parent commit.

> [1] Line-level log's rename following is quite similar to 'git log
>     --follow path', with the notable differences that it does handle
>     multiple paths at once as well, and that it doesn't show the
>     commit performing the rename if it's an exact rename.

Yeah, it's one of the reasons why "log --follow" is not (yet) a
serious "feature" but merely a "checkbox item".

> [2] This slowness might not have been apparent initially, because back
>     when the line-level log feature was introduced rename detection

s/introduced/&,/

>     was not yet enabled by default; 12da1d1f6f (Implement line-history
>     search (git log -L), 2013-03-28) and 5404c116aa (diff: activate
>     diff.renames by default, 2016-02-25).

s/\.$/ are about 3 years apart&/ or something like that to make what
follows the semicolon a full sentence?
SZEDER Gábor Aug. 21, 2019, 5:35 p.m. UTC | #3
On Wed, Aug 21, 2019 at 11:53:28AM -0400, Derrick Stolee wrote:
> On 8/21/2019 7:04 AM, SZEDER Gábor wrote:
> > With rename detection enabled the line-level log is able to trace the
> > evolution of line ranges across whole-file renames [1].  Alas, to
> > achieve that it uses the diff machinery very inefficiently, making the
> > operation very slow [2].  And since rename detection is enabled by
> > default, the line-level log is very slow by default.
> > 
> > When the line-level log processes a commit with rename detection
> > enabled, it currently does the following (see queue_diffs()):
> > 
> >   1. Computes a full tree diff between the commit and (one of) its
> >      parent(s), i.e. invokes diff_tree_oid() with an empty
> >      'diffopt->pathspec'.
> >   2. Checks whether any paths in the line ranges were modified.
> >   3. Checks whether any modified paths in the line ranges are missing
> >      in the parent commit's tree.
> >   4. If there is such a missing path, then calls diffcore_std() to
> >      figure out whether the path was indeed renamed based on the
> >      previously computed full tree diff.
> >   5. Continues doing stuff that are unrelated to the slowness.
> > 
> > So basically the line-level log computes a full tree diff for each
> > commit-parent pair in step (1) to be used for rename detection in step
> > (4) in the off chance that an interesting path is missing from the
> > parent.
> > 
> > Avoid these expensive and mostly unnecessary full tree diffs by
> > limiting the diffs to paths in the line ranges.  This is much cheaper,
> > and makes step (2) unnecessary.  If it turns out that an interesting
> > path is missing from the parent, then fall back and compute a full
> > tree diff, so the rename detection will still work.
> 
> I applied your patches and tried them on our VFS-enabled version of Git
> (see [1]). Unfortunately, the new logic is still triggering rename
> detection, as measured by the number of objects being downloaded.

Well, the goal of this patch was to avoid full tree diffs if possible,
not to avoid rename detection :)

Anyway, I wonder how does 'git log -L1:your-evil-path --no-renames'
fare as a baseline?

> My *guess* is that the repo has a lot of merge commits, and for many
> of those, the file does not exist in the first parent.

Yeah, actual renames are only one of the possible causes that trigger
rename detection.  Surprisingly (well, to me, at least), in git.git
rename detection in line-level log is most often triggered by the
subtree merges of gitk and git-gui, not by actual renames.

> Since we are
> essentially doing a --full-history, this means that edge tries a
> rename detection. If we used the file-history simplification route of
> traveling along a treesame edge instead of caring about both parents,
> then maybe this would be avoided.
> 
> I could also be completely wrong about how this line-log code works
> with regards to --full-history.

Line-level log doesn't do '--full-history', though it does seem to
compute more diffs than a simple 'git log -- path'.

I applied the following diff on current master to add a bit of ad-hoc
tracing to see how many and which commit-parent diffs are computed:

diff --git a/line-log.c b/line-log.c
index 9010e00950..64f2c4d216 100644
--- a/line-log.c
+++ b/line-log.c
@@ -839,6 +839,8 @@ static void queue_diffs(struct line_log_data *range,
 	struct object_id *tree_oid, *parent_tree_oid;
 
 	assert(commit);
+	fprintf(stderr, "%s %s\n", oid_to_hex(&commit->object.oid),
+		parent ? oid_to_hex(&parent->object.oid) : "-");
 
 	tree_oid = get_commit_tree_oid(commit);
 	parent_tree_oid = parent ? get_commit_tree_oid(parent) : NULL;
diff --git a/revision.c b/revision.c
index 07412297f0..3f2182e32a 100644
--- a/revision.c
+++ b/revision.c
@@ -627,6 +627,8 @@ static int rev_compare_tree(struct rev_info *revs,
 	struct tree *t1 = get_commit_tree(parent);
 	struct tree *t2 = get_commit_tree(commit);
 
+	fprintf(stderr, "%s %s\n", oid_to_hex(&commit->object.oid),
+		parent ? oid_to_hex(&parent->object.oid) : "-");
 	if (!t1)
 		return REV_TREE_NEW;
 	if (!t2)


  $ ./git log v2.23.0 -- builtin/rev-list.c >/dev/null 2>P
  $ ./git log --full-history v2.23.0 -- builtin/rev-list.c >/dev/null 2>FH
  $ ./git log -L1:builtin/rev-list.c v2.23.0 >/dev/null 2>LL
  $ wc -l P FH LL
    17230 P
    70842 FH
    25995 LL
   114067 total

So line-level log clearly computes a lot less diffs than
'--full-history', though still about 50% more than a regular
pathspec-limited history traversal.  Looking at the commit-parent
pairs in the output, it appears that the difference comes mostly from
merge commits, because line-level log compares a merge commit with all
of its parents.

The number of processed commits is close enough, though:

  $ cut -d' ' -f1 P |sort -u |wc -l
  17164
  $ cut -d' ' -f1 LL |sort -u |wc -l
  17894

It seems there is still more room for improvements by avoiding
commit-non_first_parent diffs when the first parent is TREESAME, and
doing so could hopefully avoid triggering rename detection in those
subtree merges or in case of your evil path.


> > @@ -762,8 +778,7 @@ void line_log_init(struct rev_info *rev, const char *prefix, struct string_list
> >  	range = parse_lines(rev->diffopt.repo, commit, prefix, args);
> >  	add_line_range(rev, commit, range);
> >  
> > -	if (!rev->diffopt.detect_rename)
> > -		parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
> > +	parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
> >  }
> 
> So we always parse the pathspec, even if we don't do detect renames.

Erm, no: we always parse the pathspec, even if we DO detect renames.

This condition made 'git log -L... --no-renames' fast.

> > @@ -821,15 +836,29 @@ static void queue_diffs(struct line_log_data *range,
> >  			struct diff_queue_struct *queue,
> >  			struct commit *commit, struct commit *parent)
> >  {
> > +	struct object_id *tree_oid, *parent_tree_oid;
> > +
> >  	assert(commit);
> >  
> > +	tree_oid = get_commit_tree_oid(commit);
> > +	parent_tree_oid = parent ? get_commit_tree_oid(parent) : NULL;
> > +
> > +	if (opt->detect_rename &&
> > +	    !same_paths_in_pathspec_and_range(&opt->pathspec, range)) {
> > +		clear_pathspec(&opt->pathspec);
> > +		parse_pathspec_from_ranges(&opt->pathspec, range);
> > +	}
> 
> If we are detecting renames and our pathspec is not up-to-date with the
> range, then clear the pathspec and reparse. Makes sense.

Renames (and subtree merges, etc.) are rare, so the paths rarely
change, and clearing and parsing pathspecs involves 2-3 memory
allocations and frees per path, so it's worth checking first.  It had
a slight but measurable performance impact, about 2% in the Linux
repository.
Derrick Stolee Aug. 21, 2019, 6:12 p.m. UTC | #4
On 8/21/2019 1:35 PM, SZEDER Gábor wrote:
> On Wed, Aug 21, 2019 at 11:53:28AM -0400, Derrick Stolee wrote:
>> On 8/21/2019 7:04 AM, SZEDER Gábor wrote:
>>> With rename detection enabled the line-level log is able to trace the
>>> evolution of line ranges across whole-file renames [1].  Alas, to
>>> achieve that it uses the diff machinery very inefficiently, making the
>>> operation very slow [2].  And since rename detection is enabled by
>>> default, the line-level log is very slow by default.
>>>
>>> When the line-level log processes a commit with rename detection
>>> enabled, it currently does the following (see queue_diffs()):
>>>
>>>   1. Computes a full tree diff between the commit and (one of) its
>>>      parent(s), i.e. invokes diff_tree_oid() with an empty
>>>      'diffopt->pathspec'.
>>>   2. Checks whether any paths in the line ranges were modified.
>>>   3. Checks whether any modified paths in the line ranges are missing
>>>      in the parent commit's tree.
>>>   4. If there is such a missing path, then calls diffcore_std() to
>>>      figure out whether the path was indeed renamed based on the
>>>      previously computed full tree diff.
>>>   5. Continues doing stuff that are unrelated to the slowness.
>>>
>>> So basically the line-level log computes a full tree diff for each
>>> commit-parent pair in step (1) to be used for rename detection in step
>>> (4) in the off chance that an interesting path is missing from the
>>> parent.
>>>
>>> Avoid these expensive and mostly unnecessary full tree diffs by
>>> limiting the diffs to paths in the line ranges.  This is much cheaper,
>>> and makes step (2) unnecessary.  If it turns out that an interesting
>>> path is missing from the parent, then fall back and compute a full
>>> tree diff, so the rename detection will still work.
>>
>> I applied your patches and tried them on our VFS-enabled version of Git
>> (see [1]). Unfortunately, the new logic is still triggering rename
>> detection, as measured by the number of objects being downloaded.
> 
> Well, the goal of this patch was to avoid full tree diffs if possible,
> not to avoid rename detection :)
> 
> Anyway, I wonder how does 'git log -L1:your-evil-path --no-renames'
> fare as a baseline?

Yeah, adding --no-renames does really well, comparatively. Perhaps I'll
just recommend to users to use that flag for now.

Thanks,
-Stolee
SZEDER Gábor Aug. 22, 2019, 8:41 a.m. UTC | #5
On Wed, Aug 21, 2019 at 07:35:15PM +0200, SZEDER Gábor wrote:
> So line-level log clearly computes a lot less diffs than
> '--full-history', though still about 50% more than a regular
> pathspec-limited history traversal.  Looking at the commit-parent
> pairs in the output, it appears that the difference comes mostly from
> merge commits, because line-level log compares a merge commit with all
> of its parents.

> It seems there is still more room for improvements by avoiding
> commit-non_first_parent diffs when the first parent is TREESAME, and
> doing so could hopefully avoid triggering rename detection in those
> subtree merges or in case of your evil path.

Well, that fruit hung much lower than I though, just look at the size
of the WIP patch below.  I just hope that there are no unexpected
surprises, but FWIW it produces the exact same output for all files up
to 't/t5515' in v2.23.0 as the previous patch.

Can't wait to see how it fares with that evil Windows path :)

  --- >8 ---

Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent

  # git.git, ~25% of all commits are merges
  $ time git --no-pager log -L:read_alternate_refs:sha1-file.c v2.23.0

  Before:

    real    0m2.516s
    user    0m2.456s
    sys     0m0.060s

  After:

    real    0m1.132s
    user    0m1.096s
    sys     0m0.036s

  # linux.git, ~7% of all commits are merges
  $ time ~/src/git/git --no-pager log \
    -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2

  Before:

    real    0m2.599s
    user    0m2.466s
    sys     0m0.157s

  After:

    real    0m1.976s
    user    0m1.856s
    sys     0m0.121s

[TODO: get rid of unnecessary arrays, tests?, write commit message...]
---
 line-log.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/line-log.c b/line-log.c
index 9010e00950..a4b032f83a 100644
--- a/line-log.c
+++ b/line-log.c
@@ -1184,13 +1184,11 @@ static int process_ranges_merge_commit(struct rev_info *rev, struct commit *comm
 
 	p = commit->parents;
 	for (i = 0; i < nparents; i++) {
+		int changed;
 		parents[i] = p->item;
 		p = p->next;
 		queue_diffs(range, &rev->diffopt, &diffqueues[i], commit, parents[i]);
-	}
 
-	for (i = 0; i < nparents; i++) {
-		int changed;
 		cand[i] = NULL;
 		changed = process_all_files(&cand[i], rev, &diffqueues[i], range);
 		if (!changed) {
@@ -1203,7 +1201,7 @@ static int process_ranges_merge_commit(struct rev_info *rev, struct commit *comm
 			commit_list_append(parents[i], &commit->parents);
 			free(parents);
 			free(cand);
-			free_diffqueues(nparents, diffqueues);
+			free_diffqueues(i, diffqueues);
 			/* NEEDSWORK leaking like a sieve */
 			return 0;
 		}
Derrick Stolee Aug. 22, 2019, 2:53 p.m. UTC | #6
On 8/22/2019 4:41 AM, SZEDER Gábor wrote:
> On Wed, Aug 21, 2019 at 07:35:15PM +0200, SZEDER Gábor wrote:
>> So line-level log clearly computes a lot less diffs than
>> '--full-history', though still about 50% more than a regular
>> pathspec-limited history traversal.  Looking at the commit-parent
>> pairs in the output, it appears that the difference comes mostly from
>> merge commits, because line-level log compares a merge commit with all
>> of its parents.
> 
>> It seems there is still more room for improvements by avoiding
>> commit-non_first_parent diffs when the first parent is TREESAME, and
>> doing so could hopefully avoid triggering rename detection in those
>> subtree merges or in case of your evil path.
> 
> Well, that fruit hung much lower than I though, just look at the size
> of the WIP patch below.  I just hope that there are no unexpected
> surprises, but FWIW it produces the exact same output for all files up
> to 't/t5515' in v2.23.0 as the previous patch.
> 
> Can't wait to see how it fares with that evil Windows path :)

Thanks for this! With this patch, we finally have the time down to ~20s.

This is a HUGE improvement, especially considering there is only one result
for the particular section, so the entire history is explored in that time.
 
>   --- >8 ---
> 
> Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent
> 
>   # git.git, ~25% of all commits are merges
>   $ time git --no-pager log -L:read_alternate_refs:sha1-file.c v2.23.0
> 
>   Before:
> 
>     real    0m2.516s
>     user    0m2.456s
>     sys     0m0.060s
> 
>   After:
> 
>     real    0m1.132s
>     user    0m1.096s
>     sys     0m0.036s
> 
>   # linux.git, ~7% of all commits are merges
>   $ time ~/src/git/git --no-pager log \
>     -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2
> 
>   Before:
> 
>     real    0m2.599s
>     user    0m2.466s
>     sys     0m0.157s
> 
>   After:
> 
>     real    0m1.976s
>     user    0m1.856s
>     sys     0m0.121s
> 
> [TODO: get rid of unnecessary arrays, tests?, write commit message...]
> ---
>  line-log.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
> 
> diff --git a/line-log.c b/line-log.c
> index 9010e00950..a4b032f83a 100644
> --- a/line-log.c
> +++ b/line-log.c
> @@ -1184,13 +1184,11 @@ static int process_ranges_merge_commit(struct rev_info *rev, struct commit *comm
>  
>  	p = commit->parents;
>  	for (i = 0; i < nparents; i++) {
> +		int changed;
>  		parents[i] = p->item;
>  		p = p->next;
>  		queue_diffs(range, &rev->diffopt, &diffqueues[i], commit, parents[i]);
> -	}
>  
> -	for (i = 0; i < nparents; i++) {
> -		int changed;
>  		cand[i] = NULL;
>  		changed = process_all_files(&cand[i], rev, &diffqueues[i], range);
>  		if (!changed) {

Interesting. The old logic computed ALL the diffs, then started navigating.

By navigating before computing all the diffs, we are now avoiding the rename logic
on the SECOND parent, and there will be a lot of second parents that do not include
the file (depending on the number of parallel topics being merged independently).
That's why git.git has a better performance difference than linux.git.

> @@ -1203,7 +1201,7 @@ static int process_ranges_merge_commit(struct rev_info *rev, struct commit *comm
>  			commit_list_append(parents[i], &commit->parents);
>  			free(parents);
>  			free(cand);
> -			free_diffqueues(nparents, diffqueues);
> +			free_diffqueues(i, diffqueues);

Good point here, as we haven't initialized all of the queues.

Thanks,
-Stolee
Junio C Hamano Aug. 22, 2019, 4:01 p.m. UTC | #7
SZEDER Gábor <szeder.dev@gmail.com> writes:

> Well, that fruit hung much lower than I though, just look at the size
> of the WIP patch below.  I just hope that there are no unexpected
> surprises, but FWIW it produces the exact same output for all files up
> to 't/t5515' in v2.23.0 as the previous patch.
>
> Can't wait to see how it fares with that evil Windows path :)
>
>   --- >8 ---
>
> Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent

A quick question.  That we need "stop diffing after first treesame"
patch in the first place means we have always been attempting to
follow all the parents of a merge?  I'd expect that to happen when
"--full-history" was given to "git log -L..." invocation.  When we
are simplifying side branches without "--full-history", I agree that
we should see if any parent is treesame with respect to the paths we
are interested in, and if so ignore all other parents.

Or am I misunderstanding the issue here?
SZEDER Gábor Aug. 22, 2019, 4:26 p.m. UTC | #8
On Thu, Aug 22, 2019 at 09:01:44AM -0700, Junio C Hamano wrote:
> SZEDER Gábor <szeder.dev@gmail.com> writes:
> 
> > Well, that fruit hung much lower than I though, just look at the size
> > of the WIP patch below.  I just hope that there are no unexpected
> > surprises, but FWIW it produces the exact same output for all files up
> > to 't/t5515' in v2.23.0 as the previous patch.
> >
> > Can't wait to see how it fares with that evil Windows path :)
> >
> >   --- >8 ---
> >
> > Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent
> 
> A quick question.  That we need "stop diffing after first treesame"
> patch in the first place means we have always been attempting to
> follow all the parents of a merge?

To follow, no.

But there are two subsequent loops: the first loop computed the diffs
between the merge and each of its parents, while the second processed
those diffs, and returned as soon as it found a treesame parent,
without following the others.

This patch unified those two loops so it computes the diff with the
first parent, and then processes that diff right away, and returns if
treesame, thereby avoding diffing the remaining parents.

>  I'd expect that to happen when
> "--full-history" was given to "git log -L..." invocation.

Oh, right, I didn't consider '--full-history'.  In that case it should
not stop at the first parent.

Hmm, looking into this, it seems that line-level log doesn't work with
'--full-history' to begin with:

  # Each commit does what the subject says.
  $ git log --oneline --graph
  *   f9bf557 (HEAD -> master) Merge branch 'branch'
  |\  
  | * 1b573fb (branch) Revert "Modify file"
  | * 3634cf3 Modify file
  |/  
  * 8842c18 Add file
  $ git log --oneline file
  8842c18 Add file
  $ git log --oneline --full-history file
  1b573fb (branch) Revert "Modify file"
  3634cf3 Modify file
  8842c18 Add file
  $ ~/src/git/BUILDS/v2.23.0/bin/git log -L1:file --oneline --no-patch
  8842c18 Add file
  $ ~/src/git/BUILDS/v2.23.0/bin/git log -L1:file --oneline --no-patch --full-history
  8842c18 Add file
Derrick Stolee Aug. 22, 2019, 4:51 p.m. UTC | #9
On 8/22/2019 12:26 PM, SZEDER Gábor wrote:
> On Thu, Aug 22, 2019 at 09:01:44AM -0700, Junio C Hamano wrote:
>> SZEDER Gábor <szeder.dev@gmail.com> writes:
>>
>>> Well, that fruit hung much lower than I though, just look at the size
>>> of the WIP patch below.  I just hope that there are no unexpected
>>> surprises, but FWIW it produces the exact same output for all files up
>>> to 't/t5515' in v2.23.0 as the previous patch.
>>>
>>> Can't wait to see how it fares with that evil Windows path :)
>>>
>>>   --- >8 ---
>>>
>>> Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent
>>
>> A quick question.  That we need "stop diffing after first treesame"
>> patch in the first place means we have always been attempting to
>> follow all the parents of a merge?
> 
> To follow, no.
> 
> But there are two subsequent loops: the first loop computed the diffs
> between the merge and each of its parents, while the second processed
> those diffs, and returned as soon as it found a treesame parent,
> without following the others.
> 
> This patch unified those two loops so it computes the diff with the
> first parent, and then processes that diff right away, and returns if
> treesame, thereby avoding diffing the remaining parents.

The change you've proposed could be made a bit better in the following
way: first look for a treesame parent by computing the diffs without
rename detection. Re-enable rename detection only if no treesame parent
is found.

That would be sure to avoid the rename detection even if the merge has
a treesame second parent.

Thanks,
-Stolee
SZEDER Gábor Aug. 23, 2019, 10:04 a.m. UTC | #10
On Thu, Aug 22, 2019 at 10:41:58AM +0200, SZEDER Gábor wrote:
> On Wed, Aug 21, 2019 at 07:35:15PM +0200, SZEDER Gábor wrote:
> Subject: [PATCH 3/2] WIP line-log: stop diff-ing after first TREESAME merge parent

>   # linux.git, ~7% of all commits are merges
>   $ time ~/src/git/git --no-pager log \
>     -L:build_restore_work_registers:arch/mips/mm/tlbex.c v5.2
> 
>   Before:
> 
>     real    0m2.599s
>     user    0m2.466s
>     sys     0m0.157s
> 
>   After:
> 
>     real    0m1.976s
>     user    0m1.856s
>     sys     0m0.121s

So, to recap, the above command follows the given line range through
the whole history, and the timings were done with these three patches
on top of v2.23.0 and with a commit-graph file present and used.

When merged with my more responsive line-level log series, the same
command takes about 12% longer:

  real    0m2.216s
  user    0m2.108s
  sys     0m0.109s

Looking at the flame graphs generated from the perf profiles I notice
the following:

  - (line_log_)process_ranges_arbitrary_commit(), i.e. the function
    responsible for processing all commits for the line-level log,
    shows up in 2408 samples before the merge and in 2387 samples
    after the merge.  I'm inclined to write it off as noise.

  - Before the merge limit_list() and sort_in_topological_order(), the
    two functions responsible for topo-ordering without generation
    numbers, show up in 2484 and 707 samples, respectively, which is
    3191 samples combined.

  - After the merge expand_topo_walk() alone shows up in 4341 samples,
    with next_topo_commit() in 72 samples.
    
So it appears that the generation numbers-based topo-ordering is
almost 40% slower than "traditional" topo-ordering.  It came as a
surprise to me, but perhaps to you it's expected?

Anyway, the time to show the first commit is still much better with
generarion numbers, so overall I think it more than offsets this ~10%
difference.

Patch
diff mbox series

diff --git a/line-log.c b/line-log.c
index fddd91f060..9010e00950 100644
--- a/line-log.c
+++ b/line-log.c
@@ -737,6 +737,22 @@  static struct line_log_data *lookup_line_range(struct rev_info *revs,
 	return ret;
 }
 
+static int same_paths_in_pathspec_and_range(struct pathspec *pathspec,
+					    struct line_log_data *range)
+{
+	int i;
+	struct line_log_data *r;
+
+	for (i = 0, r = range; i < pathspec->nr && r; i++, r = r->next)
+		if (strcmp(pathspec->items[i].match, r->path))
+			return 0;
+	if (i < pathspec->nr || r)
+		/* different number of pathspec items and ranges */
+		return 0;
+
+	return 1;
+}
+
 static void parse_pathspec_from_ranges(struct pathspec *pathspec,
 				       struct line_log_data *range)
 {
@@ -762,8 +778,7 @@  void line_log_init(struct rev_info *rev, const char *prefix, struct string_list
 	range = parse_lines(rev->diffopt.repo, commit, prefix, args);
 	add_line_range(rev, commit, range);
 
-	if (!rev->diffopt.detect_rename)
-		parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
+	parse_pathspec_from_ranges(&rev->diffopt.pathspec, range);
 }
 
 static void move_diff_queue(struct diff_queue_struct *dst,
@@ -821,15 +836,29 @@  static void queue_diffs(struct line_log_data *range,
 			struct diff_queue_struct *queue,
 			struct commit *commit, struct commit *parent)
 {
+	struct object_id *tree_oid, *parent_tree_oid;
+
 	assert(commit);
 
+	tree_oid = get_commit_tree_oid(commit);
+	parent_tree_oid = parent ? get_commit_tree_oid(parent) : NULL;
+
+	if (opt->detect_rename &&
+	    !same_paths_in_pathspec_and_range(&opt->pathspec, range)) {
+		clear_pathspec(&opt->pathspec);
+		parse_pathspec_from_ranges(&opt->pathspec, range);
+	}
 	DIFF_QUEUE_CLEAR(&diff_queued_diff);
-	diff_tree_oid(parent ? get_commit_tree_oid(parent) : NULL,
-		      get_commit_tree_oid(commit), "", opt);
-	if (opt->detect_rename) {
+	diff_tree_oid(parent_tree_oid, tree_oid, "", opt);
+	if (opt->detect_rename && diff_might_be_rename()) {
+		/* must look at the full tree diff to detect renames */
+		clear_pathspec(&opt->pathspec);
+		DIFF_QUEUE_CLEAR(&diff_queued_diff);
+
+		diff_tree_oid(parent_tree_oid, tree_oid, "", opt);
+
 		filter_diffs_for_paths(range, 1);
-		if (diff_might_be_rename())
-			diffcore_std(opt);
+		diffcore_std(opt);
 		filter_diffs_for_paths(range, 0);
 	}
 	move_diff_queue(queue, &diff_queued_diff);
diff --git a/t/t4211-line-log.sh b/t/t4211-line-log.sh
index 1db7bd0f59..8319163744 100755
--- a/t/t4211-line-log.sh
+++ b/t/t4211-line-log.sh
@@ -132,4 +132,86 @@  test_expect_success '--raw is forbidden' '
 	test_must_fail git log -L1,24:b.c --raw
 '
 
+test_expect_success 'setup for checking fancy rename following' '
+	git checkout --orphan moves-start &&
+	git reset --hard &&
+
+	printf "%s\n"    12 13 14 15      b c d e   >file-1 &&
+	printf "%s\n"    22 23 24 25      B C D E   >file-2 &&
+	git add file-1 file-2 &&
+	test_tick &&
+	git commit -m "Add file-1 and file-2" &&
+	oid_add_f1_f2=$(git rev-parse --short HEAD) &&
+
+	git checkout -b moves-main &&
+	printf "%s\n" 11 12 13 14 15      b c d e   >file-1 &&
+	git commit -a -m "Modify file-1 on main" &&
+	oid_mod_f1_main=$(git rev-parse --short HEAD) &&
+
+	printf "%s\n" 21 22 23 24 25      B C D E   >file-2 &&
+	git commit -a -m "Modify file-2 on main #1" &&
+	oid_mod_f2_main_1=$(git rev-parse --short HEAD) &&
+
+	git mv file-1 renamed-1 &&
+	git commit -m "Rename file-1 to renamed-1 on main" &&
+
+	printf "%s\n" 11 12 13 14 15      b c d e f >renamed-1 &&
+	git commit -a -m "Modify renamed-1 on main" &&
+	oid_mod_r1_main=$(git rev-parse --short HEAD) &&
+
+	printf "%s\n" 21 22 23 24 25      B C D E F >file-2 &&
+	git commit -a -m "Modify file-2 on main #2" &&
+	oid_mod_f2_main_2=$(git rev-parse --short HEAD) &&
+
+	git checkout -b moves-side moves-start &&
+	printf "%s\n"    12 13 14 15 16   b c d e   >file-1 &&
+	git commit -a -m "Modify file-1 on side #1" &&
+	oid_mod_f1_side_1=$(git rev-parse --short HEAD) &&
+
+	printf "%s\n"    22 23 24 25 26   B C D E   >file-2 &&
+	git commit -a -m "Modify file-2 on side" &&
+	oid_mod_f2_side=$(git rev-parse --short HEAD) &&
+
+	git mv file-2 renamed-2 &&
+	git commit -m "Rename file-2 to renamed-2 on side" &&
+
+	printf "%s\n"    12 13 14 15 16 a b c d e   >file-1 &&
+	git commit -a -m "Modify file-1 on side #2" &&
+	oid_mod_f1_side_2=$(git rev-parse --short HEAD) &&
+
+	printf "%s\n"    22 23 24 25 26 A B C D E   >renamed-2 &&
+	git commit -a -m "Modify renamed-2 on side" &&
+	oid_mod_r2_side=$(git rev-parse --short HEAD) &&
+
+	git checkout moves-main &&
+	git merge moves-side &&
+	oid_merge=$(git rev-parse --short HEAD)
+'
+
+test_expect_success 'fancy rename following #1' '
+	cat >expect <<-EOF &&
+	$oid_merge Merge branch '\''moves-side'\'' into moves-main
+	$oid_mod_f1_side_2 Modify file-1 on side #2
+	$oid_mod_f1_side_1 Modify file-1 on side #1
+	$oid_mod_r1_main Modify renamed-1 on main
+	$oid_mod_f1_main Modify file-1 on main
+	$oid_add_f1_f2 Add file-1 and file-2
+	EOF
+	git log -L1:renamed-1 --oneline --no-patch >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success 'fancy rename following #2' '
+	cat >expect <<-EOF &&
+	$oid_merge Merge branch '\''moves-side'\'' into moves-main
+	$oid_mod_r2_side Modify renamed-2 on side
+	$oid_mod_f2_side Modify file-2 on side
+	$oid_mod_f2_main_2 Modify file-2 on main #2
+	$oid_mod_f2_main_1 Modify file-2 on main #1
+	$oid_add_f1_f2 Add file-1 and file-2
+	EOF
+	git log -L1:renamed-2 --oneline --no-patch >actual &&
+	test_cmp expect actual
+'
+
 test_done