Share DWARF partial symtabs

This changes the DWARF reader to share partial symtabs (or indices if
they are available) across objfiles.  This has a few parts.

* If multiple objfiles backed by the same BFD can share partial symtabs
  (see below), a single dwarf2_per_bfd is created.  It is stored in the
  per-bfd `dwarf2_per_bfd_bfd_data_key` registry.  Multiple
  dwarf2_per_objfile objects will point to the same instance.  The
  lifetime of these dwarf2_per_bfd objects is naturally handled.  When
  all the objfiles using the BFD are destroyed, the BFD's refount drops
  to 0, which triggers the removal of the corresponding dwarf2_per_bfd
  object from the registry and its destruction.

* If multiple objfiles backed by the same BFD can't share partial
  symtabs (see below), one dwarf2_per_bfd object is created for each
  objfile.  Each dwarf2_per_objfile will point to their own instance of
  dwarf2_per_bfd.  These instances of dwarf2_per_bfd are kept in a
  per-objfile registry, meaning that when the objfile gets destroyed,
  the dwarf2_per_bfd instance gets destroyed as well.

* objfile::partial_symtabs is changed to be a shared_ptr again.  This
  lets us stash a second reference in dwarf2_per_bfd; if the DWARF
  data is being shared, we can simply copy this value to the new
  objfile.

* Two dwarf2_per_objfile objects backed by the same BFD may share a
  dwarf2_per_bfd instance if:

  * No other symbol reader has found symbols, and
  * No BFD section rqeuires relocation

YYYY-MM-DD  Tom Tromey  <tom@tromey.com>
YYYY-MM-DD  Simon Marchi  <simon.marchi@efficios.com>

	* objfiles.h (struct objfile) <partial_symtabs>: Now a
	shared_ptr.
	* dwarf2/read.h (struct dwarf2_per_objfile) <partial_symtabs>: New
	member.
	* dwarf2/read.c (dwarf2_per_bfd_bfd_data_key,
	dwarf2_per_bfd_objfile_data_key>: New globals.
	(dwarf2_has_info): Use shared dwarf2_per_bfd if possible.
	(dwarf2_get_section_info): Use get_dwarf2_per_objfile.
	(dwarf2_initialize_objfile): Consider cases where per_bfd can be
	shared.
	(dwarf2_build_psymtabs): Set objfile::partial_symtabs and
	short-circuit when sharing.
	(dwarf2_build_psymtabs): Set dwarf2_per_objfile::partial_symtabs.
	(dwarf2_psymtab::expand_psymtab): Use free_cached_comp_units.
diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
index 3f93a51..2de4ead 100644
--- a/gdb/dwarf2/read.c
+++ b/gdb/dwarf2/read.c
@@ -105,7 +105,19 @@
 /* When true, do not reject deprecated .gdb_index sections.  */
 static bool use_deprecated_index_sections = false;
 
-static const struct objfile_key<dwarf2_per_objfile> dwarf2_objfile_data_key;
+/* This is used to store the data that is always per objfile.  */
+static const objfile_key<dwarf2_per_objfile> dwarf2_objfile_data_key;
+
+/* These are used to store the dwarf2_per_bfd objects.
+
+   objfiles having the same BFD, which doesn't require relocations, are going to
+   share a dwarf2_per_bfd object, which is held in the _bfd_data_key version.
+
+   Other objfiles are not going to share a dwarf2_per_bfd with any other
+   objfiles, so they'll have their own version kept in the _objfile_data_key
+   version.  */
+static const struct bfd_key<dwarf2_per_bfd> dwarf2_per_bfd_bfd_data_key;
+static const struct objfile_key<dwarf2_per_bfd> dwarf2_per_bfd_objfile_data_key;
 
 /* The "aclass" indices for various kinds of computed DWARF symbols.  */
 
@@ -1853,9 +1865,30 @@
 
   if (dwarf2_per_objfile == NULL)
     {
-      /* For now, each dwarf2_per_objfile owns its own dwarf2_per_bfd (no
-         sharing yet).  */
-      dwarf2_per_bfd *per_bfd = new dwarf2_per_bfd (objfile->obfd, names, can_copy);
+      dwarf2_per_bfd *per_bfd;
+
+      /* We can share a "dwarf2_per_bfd" with other objfiles if the BFD
+         doesn't require relocations and if there aren't partial symbols
+	 from some other reader.  */
+      if (!objfile_has_partial_symbols (objfile)
+	  && !gdb_bfd_requires_relocations (objfile->obfd))
+	{
+	  /* See if one has been created for this BFD yet.  */
+	  per_bfd = dwarf2_per_bfd_bfd_data_key.get (objfile->obfd);
+
+	  if (per_bfd == nullptr)
+	    {
+	      /* No, create it now.  */
+	      per_bfd = new dwarf2_per_bfd (objfile->obfd, names, can_copy);
+	      dwarf2_per_bfd_bfd_data_key.set (objfile->obfd, per_bfd);
+	    }
+	}
+      else
+	{
+	  /* No sharing possible, create one specifically for this objfile.  */
+	  per_bfd = new dwarf2_per_bfd (objfile->obfd, names, can_copy);
+	  dwarf2_per_bfd_objfile_data_key.set (objfile, per_bfd);
+	}
 
       dwarf2_per_objfile = dwarf2_objfile_data_key.emplace (objfile, objfile, per_bfd);
     }
@@ -2017,7 +2050,7 @@
                          asection **sectp, const gdb_byte **bufp,
                          bfd_size_type *sizep)
 {
-  struct dwarf2_per_objfile *data = dwarf2_objfile_data_key.get (objfile);
+  struct dwarf2_per_objfile *data = get_dwarf2_per_objfile (objfile);
   struct dwarf2_section_info *info;
 
   /* We may see an objfile without any DWARF, in which case we just
@@ -5874,6 +5907,7 @@
 {
   struct dwarf2_per_objfile *dwarf2_per_objfile
     = get_dwarf2_per_objfile (objfile);
+  dwarf2_per_bfd *per_bfd = dwarf2_per_objfile->per_bfd;
 
   /* If we're about to read full symbols, don't bother with the
      indices.  In this case we also don't care if some other debug
@@ -5881,20 +5915,28 @@
      expanded anyway.  */
   if ((objfile->flags & OBJF_READNOW))
     {
-      dwarf2_per_objfile->per_bfd->using_index = 1;
+      /* When using READNOW, the using_index flag (set below) indicates that
+	 PER_BFD was already initialized, when we loaded some other objfile.  */
+      if (per_bfd->using_index)
+	{
+	  *index_kind = dw_index_kind::GDB_INDEX;
+	  dwarf2_per_objfile->resize_symtabs ();
+	  return true;
+	}
+
+      per_bfd->using_index = 1;
       create_all_comp_units (dwarf2_per_objfile);
       create_all_type_units (dwarf2_per_objfile);
-      dwarf2_per_objfile->per_bfd->quick_file_names_table
-	= create_quick_file_names_table
-	    (dwarf2_per_objfile->per_bfd->all_comp_units.size ());
+      per_bfd->quick_file_names_table
+	= create_quick_file_names_table (per_bfd->all_comp_units.size ());
       dwarf2_per_objfile->resize_symtabs ();
 
-      for (int i = 0; i < (dwarf2_per_objfile->per_bfd->all_comp_units.size ()
-			   + dwarf2_per_objfile->per_bfd->all_type_units.size ()); ++i)
+      for (int i = 0; i < (per_bfd->all_comp_units.size ()
+			   + per_bfd->all_type_units.size ()); ++i)
 	{
-	  dwarf2_per_cu_data *per_cu = dwarf2_per_objfile->per_bfd->get_cutu (i);
+	  dwarf2_per_cu_data *per_cu = per_bfd->get_cutu (i);
 
-	  per_cu->v.quick = OBSTACK_ZALLOC (&dwarf2_per_objfile->per_bfd->obstack,
+	  per_cu->v.quick = OBSTACK_ZALLOC (&per_bfd->obstack,
 					    struct dwarf2_per_cu_quick_data);
 	}
 
@@ -5905,6 +5947,24 @@
       return true;
     }
 
+  /* Was a debug names index already read when we processed an objfile sharing
+     PER_BFD?  */
+  if (per_bfd->debug_names_table != nullptr)
+    {
+      *index_kind = dw_index_kind::DEBUG_NAMES;
+      dwarf2_per_objfile->resize_symtabs ();
+      return true;
+    }
+
+  /* Was a GDB index already read when we processed an objfile sharing
+     PER_BFD?  */
+  if (per_bfd->index_table != nullptr)
+    {
+      *index_kind = dw_index_kind::GDB_INDEX;
+      dwarf2_per_objfile->resize_symtabs ();
+      return true;
+    }
+
   if (dwarf2_read_debug_names (dwarf2_per_objfile))
     {
       *index_kind = dw_index_kind::DEBUG_NAMES;
@@ -5945,6 +6005,16 @@
 {
   struct dwarf2_per_objfile *dwarf2_per_objfile
     = get_dwarf2_per_objfile (objfile);
+  dwarf2_per_bfd *per_bfd = dwarf2_per_objfile->per_bfd;
+
+  if (per_bfd->partial_symtabs != nullptr)
+    {
+      /* Partial symbols were already read, so now we can simply
+	 attach them.  */
+      objfile->partial_symtabs = per_bfd->partial_symtabs;
+      dwarf2_per_objfile->resize_symtabs ();
+      return;
+    }
 
   init_psymbol_list (objfile, 1024);
 
@@ -5966,6 +6036,12 @@
     {
       exception_print (gdb_stderr, except);
     }
+
+  /* Finish by setting the local reference to partial symtabs, so that
+     we don't try to read them again if reading another objfile with the same
+     BFD.  If we can't in fact share, this won't make a difference anyway as
+     the dwarf2_per_bfd object won't be shared.  */
+  per_bfd->partial_symtabs = objfile->partial_symtabs;
 }
 
 /* Find the base address of the compilation unit for range lists and
@@ -9014,9 +9090,10 @@
 {
   gdb_assert (!readin_p (objfile));
 
+  dwarf2_per_objfile *per_objfile = get_dwarf2_per_objfile (objfile);
+  free_cached_comp_units freer (per_objfile);
   expand_dependencies (objfile);
 
-  dwarf2_per_objfile *per_objfile = get_dwarf2_per_objfile (objfile);
   dw2_do_instantiate_symtab (per_cu_data, per_objfile, false);
   gdb_assert (get_compunit_symtab (objfile) != nullptr);
 }
diff --git a/gdb/dwarf2/read.h b/gdb/dwarf2/read.h
index 996cf55..53f0c73 100644
--- a/gdb/dwarf2/read.h
+++ b/gdb/dwarf2/read.h
@@ -254,6 +254,11 @@
   /* CUs that are queued to be read.  */
   std::queue<dwarf2_queue_item> queue;
 
+  /* We keep a separate reference to the partial symtabs, in case we
+     are sharing them between objfiles.  This is only set after
+     partial symbols have been read the first time.  */
+  std::shared_ptr<psymtab_storage> partial_symtabs;
+
 private:
 
   /* The total number of per_cu and signatured_type objects that have
diff --git a/gdb/objfiles.h b/gdb/objfiles.h
index 0b47bd0..56ff521 100644
--- a/gdb/objfiles.h
+++ b/gdb/objfiles.h
@@ -576,7 +576,7 @@
 
   /* The partial symbol tables.  */
 
-  std::unique_ptr<psymtab_storage> partial_symtabs;
+  std::shared_ptr<psymtab_storage> partial_symtabs;
 
   /* The object file's BFD.  Can be null if the objfile contains only
      minimal symbols, e.g. the run time common symbols for SunOS4.  */