This is the mail archive of the gdb-patches@sources.redhat.com mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[intercu] Handle simple DIEs pre-emptively


This change cuts down on the number of DIEs we have to handle in
scan_partial_symbols in two ways:

 - I was saving DIEs for a number of tags that don't need to be handled.
   The largest example was DW_TAG_pointer_type, which generally will not
   have a name.

 - Some DIEs we can determine that we will not need to refer back to, and
   process immediately.  This wipes out the data cache effects that were
   causing the largest slowdown.

This is the last of my queued patches, returning the branch to within 4-8 %
performance of mainline.  If you want to have a net performance increase the
way to do it is to specialize bcache.c for particular types that have fixed
lengths and more efficient hash functions; becase of symbol_set_names the
name pointer can act as a hash key for the entire partial symbol, and that
hash function amounts for at least 4% of total startup time on my C tests.
C++ tests are still dominated by compare_psymbols and strcmp_iw.

Checked in to the intercu branch.  Real inter-CU support for partial symbols
will be next, I think.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

2004-02-21  Daniel Jacobowitz  <drow@mvista.com>

	* dwarf2read.c (maybe_specification_partial_die): Removed.
	(is_type_tag_for_partial): Renamed from is_type_tag.  Remove
	some tags.
	(load_partial_dies): Call is_type_tag_for_partial.  Handle simple
	DIEs early.

Index: dwarf2read.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2read.c,v
retrieving revision 1.135.2.13
diff -u -p -r1.135.2.13 dwarf2read.c
--- dwarf2read.c	21 Feb 2004 21:32:25 -0000	1.135.2.13
+++ dwarf2read.c	22 Feb 2004 00:54:09 -0000
@@ -4462,63 +4462,37 @@ dwarf2_lookup_abbrev (unsigned int numbe
   return NULL;
 }
 
-/* Returns nonzero if TAG represents a type.  */
+/* Returns nonzero if TAG represents a type that we might generate a partial
+   symbol for.  */
 
 static int
-is_type_tag (int tag)
+is_type_tag_for_partial (int tag)
 { 
   switch (tag)
     {
+#if 0
+    /* Some types that would be reasonable to generate partial symbols for,
+       that we don't at present.  */
     case DW_TAG_array_type:
+    case DW_TAG_file_type:
+    case DW_TAG_ptr_to_member_type:
+    case DW_TAG_set_type:
+    case DW_TAG_string_type:
+    case DW_TAG_subroutine_type:
+#endif
+    case DW_TAG_base_type:
     case DW_TAG_class_type:
     case DW_TAG_enumeration_type:
-    case DW_TAG_pointer_type:
-    case DW_TAG_reference_type:
-    case DW_TAG_string_type:
     case DW_TAG_structure_type:
-    case DW_TAG_subroutine_type:
-    case DW_TAG_union_type:
-    case DW_TAG_ptr_to_member_type:
-    case DW_TAG_set_type:
     case DW_TAG_subrange_type:
-    case DW_TAG_base_type:
-    case DW_TAG_const_type:
-    case DW_TAG_file_type:
-    case DW_TAG_packed_type:
-    case DW_TAG_volatile_type:
     case DW_TAG_typedef:
+    case DW_TAG_union_type:
       return 1;
     default:
       return 0;
     }
 }
 
-#if 0
-/* Returns non-zero if PART_DIE might be referenced via DW_AT_specification
-   or DW_AT_abstract_origin.  */
-
-static int
-maybe_specification_partial_die (struct partial_die_info *part_die)
-{
-  if (is_type_tag (part_die->tag))
-    return 1;
-
-  /* FIXME: How does DW_AT_abstract_origin play into this?  Is it
-     possible to reference abstract origin in another CU (probably);
-     will they all be declarations (probably not?).
-
-     Understand this issue before posting patch.  */
-  if (part_die->is_declaration && !part_die->has_specification)
-    return 1;
-
-  /* FIXME: That's all GCC checks (is_symbol_die); but GCC can emit
-     DW_AT_specification for non-declarations.  I think they ought to
-     be tagged as declarations... */
-
-  return 0;
-}
-#endif
-
 /* Load all DIEs that are interesting for partial symbols into memory.  */
 
 static struct partial_die_info *
@@ -4558,7 +4532,7 @@ load_partial_dies (bfd *abfd, char *info
 	}
 
       /* Check whether this DIE is interesting enough to save.  */
-      if (!is_type_tag (abbrev->tag)
+      if (!is_type_tag_for_partial (abbrev->tag)
 	  && abbrev->tag != DW_TAG_enumerator
 	  && abbrev->tag != DW_TAG_subprogram
 	  && abbrev->tag != DW_TAG_variable
@@ -4573,6 +4547,62 @@ load_partial_dies (bfd *abfd, char *info
       //      fprintf_unfiltered (gdb_stderr, "Loading DIE %x\n", info_ptr - dwarf_info_buffer);
       info_ptr = load_partial_die (part_die, abbrev, bytes_read,
 				   abfd, info_ptr, cu);
+
+      /* This two-pass algorithm for processing partial symbols has a high
+	 cost in cache pressure.  Thus, handle some trivial cases here
+	 which cover the majority of C partial symbols.  DIEs which
+         neither have specification tags in them, nor could have specification
+         tags elsewhere pointing at them, can simply be processed and
+	 discarded.
+
+	 This segment is also optional; scan_partial_symbols and
+	 add_partial_symbol will handle these DIEs if we simply chain
+	 them in normally.  When compilers which do not emit large
+         quantities of duplicate debug information are more common,
+         this code can probably be removed.  */
+
+      /* Any complete simple types at the top level (pretty much all
+	 of them, for a language without namespaces), can be processed
+	 directly.  */
+      if (parent_die == NULL
+	  && part_die->has_specification == 0
+	  && part_die->is_declaration == 0
+	  && (part_die->tag == DW_TAG_typedef
+	      || part_die->tag == DW_TAG_base_type
+	      || part_die->tag == DW_TAG_subrange_type
+	      || part_die->tag == DW_TAG_enumeration_type))
+	{
+	  add_psymbol_to_list (part_die->name, strlen (part_die->name),
+			       VAR_DOMAIN, LOC_TYPEDEF,
+			       &cu->objfile->static_psymbols,
+			       0, (CORE_ADDR) 0, cu->language, cu->objfile);
+	  info_ptr = locate_pdi_sibling (part_die, info_ptr, abfd, cu);
+	  continue;
+	}
+
+      /* If we're at the second level, and we're an enumerator, and
+	 our parent has no specification (meaning possibly lives in a
+	 namespace elsewhere), then we can add the partial symbol now
+	 instead of queueing it.  */
+      if (part_die->tag == DW_TAG_enumerator
+	  && parent_die != NULL
+	  && parent_die->die_parent == NULL
+	  && parent_die->tag == DW_TAG_enumeration_type
+	  && parent_die->has_specification == 0)
+	{
+	  if (part_die->name == NULL)
+	    complaint (&symfile_complaints, "malformed enumerator DIE ignored");
+	  else
+	    add_psymbol_to_list (part_die->name, strlen (part_die->name),
+				 VAR_DOMAIN, LOC_CONST,
+				 cu->language == language_cplus
+				 ? &cu->objfile->global_psymbols
+				 : &cu->objfile->static_psymbols,
+				 0, (CORE_ADDR) 0, cu->language, cu->objfile);
+
+	  info_ptr = locate_pdi_sibling (part_die, info_ptr, abfd, cu);
+	  continue;
+	}
 
       /* We'll save this DIE so link it in.  */
       part_die->die_parent = parent_die;


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]