public inbox for dwz@sourceware.org
 help / color / mirror / Atom feed
* [PATCH] Precompute partitions
@ 2021-02-25 14:45 Tom de Vries
  0 siblings, 0 replies; only message in thread
From: Tom de Vries @ 2021-02-25 14:45 UTC (permalink / raw)
  To: dwz, jakub, mark

Hi,

Currently, we calculate the duplicate chain partitions in partition_dups_1.
This is done twice, once for phase 1 and once for phase 2.

Instead, calculate the partitions once in partition_dups, and use those for
both calls to partition_dups_1.

Performance tested on cc1.  This causes a small performance regresssion:
...
user:
series:  5140 5100 5120 5130 5150 5160 5160 5180 5170 5130 \
         5180 5160 5090 5230 5140 5140 5210 5100 5170 5130
mean:  5149.50 (100%)
stddev:  35.46
series:  5120 5190 5230 5190 5200 5160 5170 5210 5270 5180 \
         5270 5240 5200 5200 5200 5170 5150 5220 5180 5140
mean:  5194.50 (100.87%)
stddev:  39.13
...

There's no significant increase of memory usage:
...
mem:
series:  1260512 1260456 1260492 1260368 1260608 1260268 1260656 1260488 \
         1260420 1260332 1260464 1260488 1260536 1260340 1260352 1260492 \
         1260268 1260276 1260316 1260316
mean:  1260422.40 (100%)
stddev:  113.73
series:  1260456 1260296 1260244 1260360 1260584 1260344 1260548 1260388 \
         1260424 1260304 1260252 1260560 1260664 1260476 1260480 1260416 \
         1260580 1260504 1260604 1260324
mean:  1260440.40 (100.00%)
...

We accept the small performance penalty because this patch is a prerequisite
for the PR25424 bug fix.

Any comments?

Thanks,
- Tom

Precompute partitions

2021-02-25  Tom de Vries  <tdevries@suse.de>

	* dwz.c	(calculate_partitions): New function, factored out of ...
	(partition_dups_1): ... here.  Drop vec_size parameter.  Add
	nr_partitions and partitions parameter.  Iterate over partitions array.
	(partition_dups): Call calculate_partitions.  Update calls to
	partition_dups_1.

---
 dwz.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++--------------------
 1 file changed, 53 insertions(+), 22 deletions(-)

diff --git a/dwz.c b/dwz.c
index e71b3fa..f435428 100644
--- a/dwz.c
+++ b/dwz.c
@@ -7989,34 +7989,26 @@ cnt_ref_cus (dw_die_ref ref)
    multiple CUs might be worthwhile to be moved into partial units,
    construct those partial units.  */
 static bool
-partition_dups_1 (dw_die_ref *arr, size_t vec_size,
+partition_dups_1 (dw_die_ref *arr, size_t nr_partitions, size_t *partitions,
 		  dw_cu_ref *first_partial_cu,
 		  dw_cu_ref *last_partial_cu,
 		  bool second_phase)
 {
-  size_t i, j;
+  size_t i, j, cnt;
   bool ret = false;
-  for (i = 0; i < vec_size; i = j)
+  size_t idx = 0;
+  for (idx = 0; idx < nr_partitions * 2; idx += 2)
     {
+      i = partitions[idx];
+      cnt = partitions[idx + 1];
+      j = partitions[idx + 2];
+
+      if (arr[i]->die_dup != NULL)
+	continue;
+
       dw_die_ref ref;
-      size_t cnt = 0, size = 0, k, orig_size, new_size, namespaces = 0;
+      size_t size = 0, k, orig_size, new_size, namespaces = 0;
       unsigned int force = 0;
-      if (arr[i]->die_dup != NULL)
-	{
-	  j = i + 1;
-	  continue;
-	}
-      for (j = i + 1; j < vec_size; j++)
-	{
-	  size_t this_cnt;
-	  if (!same_ref_cus_p (arr[i], arr[j], &this_cnt))
-	    break;
-	  cnt = this_cnt;
-	}
-      if (stats_p && !second_phase)
-	stats->part_cnt++;
-      if (cnt == 0)
-	cnt = cnt_ref_cus (arr[i]);
       enum dwarf_source_language part_lang
 	= gen_cu_p ? partition_lang (arr[i]) : 0;
       for (k = i; k < j; k++)
@@ -8289,6 +8281,36 @@ partition_dups_1 (dw_die_ref *arr, size_t vec_size,
   return ret;
 }
 
+/* Partition the duplicate chains in array ARR with size VEC_SIZE, and store
+   the partitions on obstack ob2, with for each partition two entries:
+   the start and the number of unique reffer CUs.  */
+static void
+calculate_partitions (dw_die_ref *arr, size_t vec_size)
+{
+  size_t i, j;
+  for (i = 0; i < vec_size; i = j)
+    {
+      size_t cnt = 0;
+      for (j = i + 1; j < vec_size; j++)
+	{
+	  size_t this_cnt;
+	  if (!same_ref_cus_p (arr[i], arr[j], &this_cnt))
+	    break;
+	  cnt = this_cnt;
+	}
+      if (cnt == 0)
+	cnt = cnt_ref_cus (arr[i]);
+      obstack_grow (&ob2, &i, sizeof (size_t));
+      obstack_grow (&ob2, &cnt, sizeof (size_t));
+    }
+
+  /* Add element to mark end of partition list.  This allows us to do
+     'j = partitions[idx + 2]' for all partitions.  */
+  obstack_grow (&ob2, &j, sizeof (size_t));
+  size_t zero = 0;
+  obstack_grow (&ob2, &zero, sizeof (size_t));
+}
+
 static inline void FORCE_INLINE
 reset_die_ref_seen (void)
 {
@@ -8443,7 +8465,16 @@ partition_dups (void)
 	  report_progress ();
 	  fprintf (stderr, "partition_dups after qsort\n");
 	}
-      if (partition_dups_1 (arr, vec_size, &first_partial_cu,
+
+      size_t *partitions = (size_t *) obstack_base (&ob2);
+      calculate_partitions (arr, vec_size);
+      size_t nr_partitions
+	= (obstack_object_size (&ob2) / sizeof (size_t)) / 2 - 1;
+      partitions = (size_t *) obstack_finish (&ob2);
+      if (stats_p)
+	stats->part_cnt += nr_partitions;
+
+      if (partition_dups_1 (arr, nr_partitions, partitions, &first_partial_cu,
 			    &last_partial_cu, false))
 	{
 	  for (i = 0; i < vec_size; i++)
@@ -8452,7 +8483,7 @@ partition_dups (void)
 	    if (arr[i]->die_dup != NULL)
 	      mark_refs (die_cu (arr[i]), arr[i], arr[i],
 			 MARK_REFS_FOLLOW_DUPS);
-	  partition_dups_1 (arr, vec_size, &first_partial_cu,
+	  partition_dups_1 (arr, nr_partitions, partitions, &first_partial_cu,
 			    &last_partial_cu, true);
 	  for (i = 0; i < vec_size; i++)
 	    arr[i]->die_ref_seen = 0;

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2021-02-25 14:45 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-02-25 14:45 [PATCH] Precompute partitions Tom de Vries

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).