[PATCH 14/14] Share DWARF partial symtabs

Luis Machado luis.machado@linaro.org
Tue Feb 18 12:26:00 GMT 2020


On 2/15/20 1:54 PM, Tom Tromey wrote:
> This changes the DWARF reader to share partial symtabs (or indices if
> they are available) across objfiles.  This has a few parts.
> 
> * The dwarf2_unshareable object is created per-objfile and is stored
>    on the objfile using a registry key.  dwarf2_enter_objfile is
>    modified to save and restore this pointer in the dwarf2_per_objfile.
> 
> * objfile::partial_symtabs is changed to be a shared_ptr again.  This
>    lets us stash a second reference in dwarf2_per_objfile; if the DWARF
>    data is being shared, we can simply copy this value to the new
>    objfile.
> 
> * The dwarf2_per_objfile itself may be shared via the BFD -- using a
>    new per-BFD registry key -- or not.  This depends both on whether
>    some other symbol reader has found symbols, and whether any BFD
>    sections require relocation.
> 
> 2020-02-15  Tom Tromey  <tom@tromey.com>
> 
> 	* objfiles.h (struct objfile) <partial_symtabs>: Now a
> 	shared_ptr.
> 	* dwarf2/read.h (struct dwarf2_per_objfile) <partial_symtabs>: New
> 	member.
> 	(class dwarf2_enter_objfile): Constructor no longer inline.
> 	<m_restore_unshared>: New member.
> 	* dwarf2/read.c (dwarf2_unshared_data_key, dwarf2_bfd_data_key):
> 	New globals.
> 	(dwarf2_objfile_data_key): Add comment.
> 	(get_dwarf2_per_objfile): Rewrite.
> 	(dwarf2_enter_objfile::dwarf2_enter_objfile): Move from read.h.
> 	Initialize m_restore_unshared.
> 	(dwarf2_per_objfile::dwarf2_per_objfile): Add "unshared"
> 	parameter.  Don't initialize objfile member.
> 	(dwarf2_per_objfile::~dwarf2_per_objfile): Don't deleted
> 	"unshareable".

"Don't delete..."

> 	(dwarf2_has_info): Check dwarf2_bfd_data_key and
> 	dwarf2_unshared_data_key.
> 	(dwarf2_get_section_info): Use get_dwarf2_per_objfile.
> 	(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.
> ---
>   gdb/ChangeLog     | 26 ++++++++++++++++++
>   gdb/dwarf2/read.c | 70 +++++++++++++++++++++++++++++++++++++----------
>   gdb/dwarf2/read.h | 12 ++++----
>   gdb/objfiles.h    |  2 +-
>   4 files changed, 89 insertions(+), 21 deletions(-)
> 
> diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
> index 15813b72005..7be57661296 100644
> --- a/gdb/dwarf2/read.c
> +++ b/gdb/dwarf2/read.c
> @@ -100,7 +100,15 @@ static bool check_physname = false;
>   /* 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_unshareable> dwarf2_unshared_data_key;
> +
> +/* These are used to store the dwarf2_per_objfile; either on the
> +   objfile or on the BFD.  */
> +static const struct objfile_key<dwarf2_per_objfile>
> +  dwarf2_objfile_data_key;
> +static const struct bfd_key<dwarf2_per_objfile>
> +  dwarf2_bfd_data_key;
>   
>   /* The "aclass" indices for various kinds of computed DWARF symbols.  */
>   
> @@ -274,7 +282,18 @@ struct mapped_debug_names final : public mapped_index_base
>   dwarf2_per_objfile *
>   get_dwarf2_per_objfile (struct objfile *objfile)
>   {
> -  return dwarf2_objfile_data_key.get (objfile);
> +  dwarf2_per_objfile *result = dwarf2_bfd_data_key.get (objfile->obfd);
> +  if (result == nullptr)
> +    result = dwarf2_objfile_data_key.get (objfile);
> +  return result;
> +}
> +
> +dwarf2_enter_objfile::dwarf2_enter_objfile (struct objfile *objfile)
> +  : m_per_objfile (get_dwarf2_per_objfile (objfile)),
> +    m_restore_objfile (&m_per_objfile->objfile, objfile),
> +    m_restore_unshared (&m_per_objfile->unshareable,
> +			dwarf2_unshared_data_key.get (objfile))
> +{
>   }
>   
>   /* Default names of the debugging sections.  */
> @@ -1754,16 +1773,12 @@ line_header_eq_voidp (const void *item_lhs, const void *item_rhs)
>   dwarf2_per_objfile::dwarf2_per_objfile (struct objfile *objfile_,
>   					const dwarf2_debug_sections *names,
>   					bool can_copy_)
> -  : objfile (objfile_),
> -    can_copy (can_copy_),
> -    /* Temporarily just allocate one per objfile -- we don't have
> -       sharing yet.  */
> -    unshareable (new dwarf2_unshareable)
> +  : can_copy (can_copy_)
>   {
>     if (names == NULL)
>       names = &dwarf2_elf_names;
>   
> -  bfd *obfd = objfile->obfd;
> +  bfd *obfd = objfile_->obfd;
>   
>     for (asection *sec = obfd->sections; sec != NULL; sec = sec->next)
>       locate_sections (obfd, sec, *names);
> @@ -1780,8 +1795,6 @@ dwarf2_per_objfile::~dwarf2_per_objfile ()
>     for (signatured_type *sig_type : all_type_units)
>       sig_type->per_cu.imported_symtabs_free ();
>   
> -  delete unshareable;
> -
>     /* Everything else should be on the objfile obstack.  */
>   }
>   
> @@ -1841,13 +1854,26 @@ dwarf2_has_info (struct objfile *objfile,
>     if (objfile->flags & OBJF_READNEVER)
>       return 0;
>   
> +  struct dwarf2_unshareable *unshared = dwarf2_unshared_data_key.get (objfile);
> +  if (unshared == nullptr)
> +    unshared = dwarf2_unshared_data_key.emplace (objfile);
> +
>     struct dwarf2_per_objfile *dwarf2_per_objfile
>       = get_dwarf2_per_objfile (objfile);
>   
>     if (dwarf2_per_objfile == NULL)
> -    dwarf2_per_objfile = dwarf2_objfile_data_key.emplace (objfile, objfile,
> -							  names,
> -							  can_copy);
> +    {
> +      dwarf2_per_objfile = new ::dwarf2_per_objfile (objfile, names,
> +						     can_copy);
> +      /* We can share this if the objfile 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))
> +	dwarf2_bfd_data_key.set (objfile->obfd, dwarf2_per_objfile);
> +      else
> +	dwarf2_objfile_data_key.set (objfile, dwarf2_per_objfile);
> +    }
>   
>     return (!dwarf2_per_objfile->info.is_virtual
>   	  && dwarf2_per_objfile->info.s.section != NULL
> @@ -2006,7 +2032,7 @@ dwarf2_get_section_info (struct objfile *objfile,
>                            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
> @@ -5862,6 +5888,15 @@ dwarf2_build_psymtabs (struct objfile *objfile)
>     struct dwarf2_per_objfile *dwarf2_per_objfile
>       = get_dwarf2_per_objfile (objfile);
>   
> +  if (dwarf2_per_objfile->partial_symtabs != nullptr)
> +    {
> +      /* Partial symbols were already read, so now we can simply
> +	 attach them.  */
> +      objfile->partial_symtabs = dwarf2_per_objfile->partial_symtabs;
> +      dwarf2_resize_unshareable (dwarf2_per_objfile);
> +      return;
> +    }
> +
>     init_psymbol_list (objfile, 1024);
>   
>     try
> @@ -5882,6 +5917,11 @@ dwarf2_build_psymtabs (struct objfile *objfile)
>       {
>         exception_print (gdb_stderr, except);
>       }
> +
> +  /* Finish by setting the local reference to partial symtabs, so that
> +     we don't try to read again.  If we can't in fact share, this
> +     won't make a difference anyway.  */
> +  dwarf2_per_objfile->partial_symtabs = objfile->partial_symtabs;
>   }
>   
>   /* Find the base address of the compilation unit for range lists and
> @@ -8915,8 +8955,8 @@ dwarf2_psymtab::expand_psymtab (struct objfile *objfile)
>     if (symtab.has_value ())
>       return;
>   
> +  free_cached_comp_units freer (dwarf2_per_objfile);
>     read_dependencies (objfile);
> -
>     dw2_do_instantiate_symtab (per_cu_data, false);
>   }
>   

Spurious line removal, but it doesn't matter much.

> diff --git a/gdb/dwarf2/read.h b/gdb/dwarf2/read.h
> index 241f3c8e71e..891a844a673 100644
> --- a/gdb/dwarf2/read.h
> +++ b/gdb/dwarf2/read.h
> @@ -299,6 +299,11 @@ public:
>     /* 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;
> +
>     /* The total number of per_cu and signatured_type objects that have
>        been created for this reader.  */
>     size_t num_psymtabs = 0;
> @@ -321,11 +326,7 @@ class dwarf2_enter_objfile
>   {
>   public:
>   
> -  dwarf2_enter_objfile (struct objfile *objfile)
> -    : m_per_objfile (get_dwarf2_per_objfile (objfile)),
> -      m_restore_objfile (&m_per_objfile->objfile, objfile)
> -  {
> -  }
> +  dwarf2_enter_objfile (struct objfile *objfile);
>   
>     ~dwarf2_enter_objfile () = default;
>   
> @@ -335,6 +336,7 @@ private:
>   
>     dwarf2_per_objfile *m_per_objfile;
>     scoped_restore_tmpl<struct objfile *> m_restore_objfile;
> +  scoped_restore_tmpl<dwarf2_unshareable *> m_restore_unshared;
>   };
>   
>   /* A partial symtab specialized for DWARF.  */
> diff --git a/gdb/objfiles.h b/gdb/objfiles.h
> index b71a8a9edb8..12892f80f71 100644
> --- a/gdb/objfiles.h
> +++ b/gdb/objfiles.h
> @@ -558,7 +558,7 @@ public:
>   
>     /* 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.  */
> 

I've gone through the series and didn't see anything obviously wrong. I 
suppose this will get more thoroughly exercised when running in 
multi-process/multi-target mode with quite a few objfiles and libraries.

Is it currently possible to exercise such a scenario automatically? I 
suppose manually loading something like a web browser will be a good 
test as well.



More information about the Gdb-patches mailing list