[PATCH 5/6] .gdb_index prod perf regression: Estimate size of psyms_seen

Pedro Alves palves@redhat.com
Mon Jun 12 16:19:00 GMT 2017


Using the same test as the previous patch, perf shows GDB spending
over 7% in "free".  A substantial number of those calls comes from
insertions in the psyms_seen unordered_set causing lots of rehashing
and recreating buckets.  Fix this by computing an estimate of the size
of the set upfront.

Using the same test as in the previous patch, against the same gdb
inferior, timing improves ~8% further:

  ~6.5s => ~6.0s (average of 5 runs).

gdb/ChangeLog:
2017-06-12  Pedro Alves  <palves@redhat.com>

	* dwarf2read.c (recursively_count_psymbols): New function.
	(write_psymtabs_to_index): Call it to compute number of psyms and
	pass estimate size of psyms_seen to unordered_set's ctor.
---
 gdb/ChangeLog    |  6 ++++++
 gdb/dwarf2read.c | 36 +++++++++++++++++++++++++++++++++++-
 2 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 01b66a1..9dbc059 100644
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,5 +1,11 @@
 2017-06-12  Pedro Alves  <palves@redhat.com>
 
+	* dwarf2read.c (recursively_count_psymbols): New function.
+	(write_psymtabs_to_index): Call it to compute number of psyms and
+	pass estimate size of psyms_seen to unordered_set's ctor.
+
+2017-06-12  Pedro Alves  <palves@redhat.com>
+
 	* dwarf2read.c (write_hash_table): Check if key already exists
 	before emplacing.
 
diff --git a/gdb/dwarf2read.c b/gdb/dwarf2read.c
index 93fd275..bff2fcb 100644
--- a/gdb/dwarf2read.c
+++ b/gdb/dwarf2read.c
@@ -23691,6 +23691,22 @@ write_one_signatured_type (void **slot, void *d)
   return 1;
 }
 
+/* Recurse into all "included" dependencies and count their symbols as
+   if they appeared in this psymtab.  */
+
+static void
+recursively_count_psymbols (struct partial_symtab *psymtab,
+			    size_t &psyms_seen)
+{
+  for (int i = 0; i < psymtab->number_of_dependencies; ++i)
+    if (psymtab->dependencies[i]->user != NULL)
+      recursively_count_psymbols (psymtab->dependencies[i],
+				  psyms_seen);
+
+  psyms_seen += psymtab->n_global_syms;
+  psyms_seen += psymtab->n_static_syms;
+}
+
 /* Recurse into all "included" dependencies and write their symbols as
    if they appeared in this psymtab.  */
 
@@ -23764,7 +23780,6 @@ write_psymtabs_to_index (struct objfile *objfile, const char *dir)
 
   mapped_symtab symtab;
   data_buf cu_list;
-  std::unordered_set<partial_symbol *> psyms_seen;
 
   /* While we're scanning CU's create a table that maps a psymtab pointer
      (which is what addrmap records) to its index (which is what is recorded
@@ -23776,6 +23791,25 @@ write_psymtabs_to_index (struct objfile *objfile, const char *dir)
   /* The CU list is already sorted, so we don't need to do additional
      work here.  Also, the debug_types entries do not appear in
      all_comp_units, but only in their own hash table.  */
+
+  /* The psyms_seen set is potentially going to be largish (~40k
+     elements when indexing a -g3 build of GDB itself).  Estimate the
+     number of elements in order to avoid too many rehashes, which
+     require rebuilding buckets and thus many trips to
+     malloc/free.  */
+  size_t psyms_count = 0;
+  for (int i = 0; i < dwarf2_per_objfile->n_comp_units; ++i)
+    {
+      struct dwarf2_per_cu_data *per_cu
+	= dwarf2_per_objfile->all_comp_units[i];
+      struct partial_symtab *psymtab = per_cu->v.psymtab;
+
+      if (psymtab != NULL && psymtab->user == NULL)
+	recursively_count_psymbols (psymtab, psyms_count);
+    }
+  /* Generating an index for gdb itself shows a ratio of
+     TOTAL_SEEN_SYMS/UNIQUE_SYMS or ~5.  4 seems like a good bet.  */
+  std::unordered_set<partial_symbol *> psyms_seen (psyms_count / 4);
   for (int i = 0; i < dwarf2_per_objfile->n_comp_units; ++i)
     {
       struct dwarf2_per_cu_data *per_cu
-- 
2.5.5



More information about the Gdb-patches mailing list