perf hists: Changes on hists__{link,match}

PATCH Superseded Namhyung Kim 3 participants 9 messages Dec 4, 2012 perf events

lore

Share & tools

lore

Namhyung Kim

author maintainer

2012-12-04 04:44 UTC #0 lore raw

cover perf hists: Changes on hists__{link,match}

Hi Arnaldo,

This is what I talked to Jiri yesterday, and it can be a common basis of both event group and multiple diff patchset. The point is using internal input or collapsed rb tree to sort hist entries rather than output tree with unnessary resort.

Please take a look.

Thanks,
Namhyung

Cc Jiri Olsa <[email protected]>

Cc Stephane Eranian <[email protected]>

Cc Namhyung Kim <[email protected]>

Namhyung Kim (3):
perf hists: Exchange order of comparing items when collapsing hists perf hists: Link hist entries before inserting to an output tree perf diff: Use internal rb tree for compute resort

 tools/perf/builtin-diff.c | 95 +++++++++++++++++++++--------------------------
 tools/perf/util/hist.c    | 51 ++++++++++++++++++-------
 2 files changed, 80 insertions(+), 66 deletions(-)

Signature · 1 line

1.7.11.7

Namhyung Kim

author maintainer

2012-12-04 04:44 UTC #1 lore raw

PATCH 1/3 perf hists: Exchange order of comparing items when collapsing hists

From: Namhyung Kim <[email protected]>

When comparing entries for collapsing put the given entry first, and then the iterated entry. This is not the case of hist_entry__cmp() when called if given sort keys don't require collapsing. So change the order for the sake of consistency. It will be required for matching and/or linking multiple hist entries.

Cc Jiri Olsa <[email protected]>

Cc Stephane Eranian <[email protected]>

Signed-off-by Namhyung Kim <[email protected]>

---
 tools/perf/util/hist.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

tools/perf/util/hist.c +1 -1

index 82df1b26f0d4..161c35e7ed0e 100644
@@ -433,7 +433,7 @@ static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused,
 		parent = *p;
 		iter = rb_entry(parent, struct hist_entry, rb_node_in);
 
-		cmp = hist_entry__collapse(iter, he);
+		cmp = hist_entry__collapse(he, iter);
 
 		if (!cmp) {
 			he_stat__add_stat(&iter->stat, &he->stat);

Signature · 1 line

1.7.11.7

Arnaldo Carvalho de Melo

2012-12-04 13:44 UTC #2 lore raw

Re: [PATCH 1/3] perf hists: Exchange order of comparing items when collapsing hists

Em Tue, Dec 04, 2012 at 01:44:23PM +0900, Namhyung Kim escreveu:

Quoted 20 lines

From: Namhyung Kim <[email protected]>
When comparing entries for collapsing put the given entry first, and then the iterated entry. This is not the case of hist_entry__cmp() when called if given sort keys don't require collapsing. So change the order for the sake of consistency. It will be required for matching and/or linking multiple hist entries.
Cc Jiri Olsa <[email protected]>
Cc Stephane Eranian <[email protected]>
Signed-off-by Namhyung Kim <[email protected]>
---
 tools/perf/util/hist.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/hist.c +0 -0
index 82df1b26f0d4..161c35e7ed0e 100644

tools/perf/util/hist.c static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused, · line 439

Quoted 8 lines

tools/perf/util/hist.c +1 -1

@@ -433,7 +433,7 @@ static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused,
 		parent = *p;
 		iter = rb_entry(parent, struct hist_entry, rb_node_in);
 
-		cmp = hist_entry__collapse(iter, he);
+		cmp = hist_entry__collapse(he, iter);
 
 		if (!cmp) {
 			he_stat__add_stat(&iter->stat, &he->stat);

What about this he_stat__add_stat call? Now the hist_entry__collapse receives (he, iter) while this right next function call receives (iter, he).

- Arnaldo

Quoted 2 lines

Signature · 1 line
1.7.11.7

Namhyung Kim

author maintainer

2012-12-04 14:57 UTC #3 lore raw

Re: [PATCH 1/3] perf hists: Exchange order of comparing items when collapsing hists

Quoted 27 lines · Arnaldo Carvalho de Melo

Em Tue, Dec 04, 2012 at 01:44:23PM +0900, Namhyung Kim escreveu:
Quoted 26 lines
From: Namhyung Kim <[email protected]>
When comparing entries for collapsing put the given entry first, and then the iterated entry. This is not the case of hist_entry__cmp() when called if given sort keys don't require collapsing. So change the order for the sake of consistency. It will be required for matching and/or linking multiple hist entries.
Cc Jiri Olsa <[email protected]>
Cc Stephane Eranian <[email protected]>
Signed-off-by Namhyung Kim <[email protected]>
---
 tools/perf/util/hist.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/hist.c +1 -1
index 82df1b26f0d4..161c35e7ed0e 100644
@@ -433,7 +433,7 @@ static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused,
 		parent = *p;
 		iter = rb_entry(parent, struct hist_entry, rb_node_in);
 
-		cmp = hist_entry__collapse(iter, he);
+		cmp = hist_entry__collapse(he, iter);
 

Quoted 6 lines

tools/perf/util/hist.c static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused, · line 439
Quoted 2 lines
tools/perf/util/hist.c +0 -0
 		if (!cmp) {
 			he_stat__add_stat(&iter->stat, &he->stat);
What about this he_stat__add_stat call? Now the hist_entry__collapse receives (he, iter) while this right next function call receives (iter, he).

Hmm.. I thought they're diffent kind of operation. hist_entry__collapse is in a process of iteration and he_stat__add_stat is not. It's just adding or copying entry's stat, so I thought it's something like memcpy
- hence the order.

If you really concern about ordering between them, maybe I can change hist_entry__cmp() to receive (iter, he).

Thanks,
Namhyung

Namhyung Kim

author maintainer

2012-12-04 04:44 UTC #4 lore raw

PATCH 2/3 perf hists: Link hist entries before inserting to an output tree

In reply to Namhyung Kim

From: Namhyung Kim <[email protected]>

For matching and/or linking hist entries, they need to be sorted by given sort keys. However current hists__match/link did this on the output trees, so that the entries in the output tree need to be resort before doing it.

This looks not so good since we have trees for collecting or collapsing entries before passing them to an output tree and they're already sorted by the given sort keys. Since we don't need to print anything at the time of matching/linking, we can use these internal trees directly instead of bothering with double resort on the output tree.

Its only user - at the time of this writing - perf diff can be easily converted to use the internal tree and can save some lines too by getting rid of unnecessary resorting codes.

It seems we need to do output resort on baseline hists for displacement prior to the match/link. But AFAIK it's okay since the resort doesn't make any change to the internal tree - i.e. every entries in the internal tree remain as is and just linked to both of internal and output tree using different keys. So linking entries in the internal tree after output resorting will not cause any harm IMHO.

Cc Jiri Olsa <[email protected]>

Cc Stephane Eranian <[email protected]>

Signed-off-by Namhyung Kim <[email protected]>

---
 tools/perf/builtin-diff.c | 73 ++++++++++++++++++-----------------------------
 tools/perf/util/hist.c    | 49 +++++++++++++++++++++++--------
 2 files changed, 65 insertions(+), 57 deletions(-)

tools/perf/builtin-diff.c +28 ~~-45~~

index d869029fb75e..b52f5d8c4a6b 100644
@@ -276,46 +276,19 @@ static struct perf_tool tool = {
 	.ordering_requires_timestamps = true,
 };
 
-static void insert_hist_entry_by_name(struct rb_root *root,
-				      struct hist_entry *he)
-{
-	struct rb_node **p = &root->rb_node;
-	struct rb_node *parent = NULL;
-	struct hist_entry *iter;
-
-	while (*p != NULL) {
-		parent = *p;
-		iter = rb_entry(parent, struct hist_entry, rb_node);
-		if (hist_entry__cmp(he, iter) < 0)
-			p = &(*p)->rb_left;
-		else
-			p = &(*p)->rb_right;
-	}
-
-	rb_link_node(&he->rb_node, parent, p);
-	rb_insert_color(&he->rb_node, root);
-}
-
-static void hists__name_resort(struct hists *self, bool sort)
+static void hists__compute_position(struct hists *hists)
 {
 	unsigned long position = 1;
-	struct rb_root tmp = RB_ROOT;
-	struct rb_node *next = rb_first(&self->entries);
+	struct rb_node *next = rb_first(&hists->entries);
 
 	while (next != NULL) {
-		struct hist_entry *n = rb_entry(next, struct hist_entry, rb_node);
+		struct hist_entry *he;
 
-		next = rb_next(&n->rb_node);
-		n->position = position++;
+		he = rb_entry(next, struct hist_entry, rb_node);
+		he->position = position++;
 
-		if (sort) {
-			rb_erase(&n->rb_node, &self->entries);
-			insert_hist_entry_by_name(&tmp, n);
-		}
+		next = rb_next(next);
 	}
-
-	if (sort)
-		self->entries = tmp;
 }
 
 static struct perf_evsel *evsel_match(struct perf_evsel *evsel,

@@ -330,34 +303,39 @@ static struct perf_evsel *evsel_match(struct perf_evsel *evsel,
 	return NULL;
 }
 
-static void perf_evlist__resort_hists(struct perf_evlist *evlist, bool name)
+static void perf_evlist__resort_hists(struct perf_evlist *evlist, bool base)
 {
 	struct perf_evsel *evsel;
 
 	list_for_each_entry(evsel, &evlist->entries, node) {
 		struct hists *hists = &evsel->hists;
 
-		hists__output_resort(hists);
+		hists__collapse_resort(hists);
 
-		/*
-		 * The hists__name_resort only sets possition
-		 * if name is false.
-		 */
-		if (name || ((!name) && show_displacement))
-			hists__name_resort(hists, name);
+		if (base && show_displacement) {
+			hists__output_resort(hists);
+			hists__compute_position(hists);
+		}
 	}
 }
 
 static void hists__baseline_only(struct hists *hists)
 {
-	struct rb_node *next = rb_first(&hists->entries);
+	struct rb_root *root;
+	struct rb_node *next;
+
+	if (sort__need_collapse)
+		root = &hists->entries_collapsed;
+	else
+		root = hists->entries_in;
 
+	next = rb_first(root);
 	while (next != NULL) {
-		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node);
+		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node_in);
 
-		next = rb_next(&he->rb_node);
+		next = rb_next(&he->rb_node_in);
 		if (!hist_entry__next_pair(he)) {
-			rb_erase(&he->rb_node, &hists->entries);
+			rb_erase(&he->rb_node_in, root);
 			hist_entry__free(he);
 		}
 	}

@@ -481,6 +459,11 @@ static void hists__process(struct hists *old, struct hists *new)
 	else
 		hists__link(new, old);
 
+	hists__output_resort(new);
+
+	if (show_displacement)
+		hists__compute_position(new);
+
 	if (sort_compute) {
 		hists__precompute(new);
 		hists__compute_resort(new);

tools/perf/util/hist.c +37 ~~-12~~

94 lines not on this page · raw message

Signature · 1 line

1.7.11.7

Namhyung Kim

author maintainer

2012-12-04 04:45 UTC #5 lore raw

PATCH 3/3 perf diff: Use internal rb tree for compute resort

In reply to Namhyung Kim

From: Namhyung Kim <[email protected]>

There's no reason to run hists_compute_resort() using output tree. Convert it to use internal tree so that it can remove unnecessary _output_resort. Also move position computation below the resort since it changes the output ordering.

Cc Jiri Olsa <[email protected]>

Cc Stephane Eranian <[email protected]>

Signed-off-by Namhyung Kim <[email protected]>

---
 tools/perf/builtin-diff.c | 32 +++++++++++++++++++-------------
 1 file changed, 19 insertions(+), 13 deletions(-)

tools/perf/builtin-diff.c +19 ~~-13~~

index b52f5d8c4a6b..4e18cea7c845 100644
@@ -435,19 +435,25 @@ static void insert_hist_entry_by_compute(struct rb_root *root,
 
 static void hists__compute_resort(struct hists *hists)
 {
-	struct rb_root tmp = RB_ROOT;
-	struct rb_node *next = rb_first(&hists->entries);
+	struct rb_root *root;
+	struct rb_node *next;
+
+	if (sort__need_collapse)
+		root = &hists->entries_collapsed;
+	else
+		root = hists->entries_in;
+
+	hists->entries = RB_ROOT;
+	next = rb_first(root);
 
 	while (next != NULL) {
-		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node);
+		struct hist_entry *he;
 
-		next = rb_next(&he->rb_node);
+		he = rb_entry(next, struct hist_entry, rb_node_in);
+		next = rb_next(&he->rb_node_in);
 
-		rb_erase(&he->rb_node, &hists->entries);
-		insert_hist_entry_by_compute(&tmp, he, compute);
+		insert_hist_entry_by_compute(&hists->entries, he, compute);
 	}
-
-	hists->entries = tmp;
 }
 
 static void hists__process(struct hists *old, struct hists *new)

@@ -459,16 +465,16 @@ static void hists__process(struct hists *old, struct hists *new)
 	else
 		hists__link(new, old);
 
-	hists__output_resort(new);
-
-	if (show_displacement)
-		hists__compute_position(new);
-
 	if (sort_compute) {
 		hists__precompute(new);
 		hists__compute_resort(new);
+	} else {
+		hists__output_resort(new);
 	}
 
+	if (show_displacement)
+		hists__compute_position(new);
+
 	hists__fprintf(new, true, 0, 0, stdout);
 }

Signature · 1 line

1.7.11.7

Arnaldo Carvalho de Melo

2012-12-04 13:45 UTC #6 lore raw

Re: [PATCH 3/3] perf diff: Use internal rb tree for compute resort

Em Tue, Dec 04, 2012 at 01:44:25PM +0900, Namhyung Kim escreveu:

Quoted 6 lines

From: Namhyung Kim <[email protected]>
There's no reason to run hists_compute_resort() using output tree. Convert it to use internal tree so that it can remove unnecessary _output_resort. Also move position computation below the resort since it changes the output ordering.

Have you tested this with 'perf top'? With the highest frequency?

- Arnaldo

Quoted 69 lines

Cc Jiri Olsa <[email protected]>

Cc Stephane Eranian <[email protected]>

Signed-off-by Namhyung Kim <[email protected]>

---
 tools/perf/builtin-diff.c | 32 +++++++++++++++++++-------------
 1 file changed, 19 insertions(+), 13 deletions(-)

tools/perf/builtin-diff.c +19 ~~-13~~

index b52f5d8c4a6b..4e18cea7c845 100644
@@ -435,19 +435,25 @@ static void insert_hist_entry_by_compute(struct rb_root *root,
 
 static void hists__compute_resort(struct hists *hists)
 {
-	struct rb_root tmp = RB_ROOT;
-	struct rb_node *next = rb_first(&hists->entries);
+	struct rb_root *root;
+	struct rb_node *next;
+
+	if (sort__need_collapse)
+		root = &hists->entries_collapsed;
+	else
+		root = hists->entries_in;
+
+	hists->entries = RB_ROOT;
+	next = rb_first(root);
 
 	while (next != NULL) {
-		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node);
+		struct hist_entry *he;
 
-		next = rb_next(&he->rb_node);
+		he = rb_entry(next, struct hist_entry, rb_node_in);
+		next = rb_next(&he->rb_node_in);
 
-		rb_erase(&he->rb_node, &hists->entries);
-		insert_hist_entry_by_compute(&tmp, he, compute);
+		insert_hist_entry_by_compute(&hists->entries, he, compute);
 	}
-
-	hists->entries = tmp;
 }
 
 static void hists__process(struct hists *old, struct hists *new)

@@ -459,16 +465,16 @@ static void hists__process(struct hists *old, struct hists *new)
 	else
 		hists__link(new, old);
 
-	hists__output_resort(new);
-
-	if (show_displacement)
-		hists__compute_position(new);
-
 	if (sort_compute) {
 		hists__precompute(new);
 		hists__compute_resort(new);
+	} else {
+		hists__output_resort(new);
 	}
 
+	if (show_displacement)
+		hists__compute_position(new);
+
 	hists__fprintf(new, true, 0, 0, stdout);
 }

Signature · 1 line

1.7.11.7

Namhyung Kim

author maintainer

2012-12-04 15:03 UTC #7 lore raw

Re: [PATCH 3/3] perf diff: Use internal rb tree for compute resort

Quoted 9 lines · Arnaldo Carvalho de Melo

Em Tue, Dec 04, 2012 at 01:44:25PM +0900, Namhyung Kim escreveu:
Quoted 6 lines
From: Namhyung Kim <[email protected]>
There's no reason to run hists_compute_resort() using output tree. Convert it to use internal tree so that it can remove unnecessary _output_resort. Also move position computation below the resort since it changes the output ordering.
Have you tested this with 'perf top'? With the highest frequency?

After testing 'perf top -F 100000' couple of minutes, I couldn't find any visible problem. Basically this patchset changes hists__link/match path which only called from 'perf diff' - I'm working on making use of that for event group report though. So I didn't check perf top side seriously. Any reason do you mention it that I'm missing?

Thanks,
Namhyung

Jiri Olsa

2012-12-04 13:55 UTC #8 lore raw

In reply to Namhyung Kim

Quoted 13 lines · Namhyung Kim

Hi Arnaldo,
This is what I talked to Jiri yesterday, and it can be a common basis of both event group and multiple diff patchset. The point is using internal input or collapsed rb tree to sort hist entries rather than output tree with unnessary resort.
Please take a look.
Thanks,
Namhyung

Quoted 8 lines

Cc Jiri Olsa <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Namhyung Kim <[email protected]>
Namhyung Kim (3):
perf hists: Exchange order of comparing items when collapsing hists perf hists: Link hist entries before inserting to an output tree perf diff: Use internal rb tree for compute resort

hi,
could you please push branch with this?

thanks,
jirka