[05/14] Introduce dwarf2_unshareable and move die_type_hash

The goal of this series is to share dwarf2_per_objfile across
objfiles.  However, this is difficult to do, because it has some
members that are inherently per-objfile.

This patch introduces a new "unshareable" object.  The idea here is
that we'll have one such object per objfile, and we'll temporarily
update the pointer when needed.  This is a bit ugly, but it's
convenient, because it doesn't require a huge rewriting of the DWARF
code.  (I would be in favor of such a rewriting, but ... it's huge.)

This patch further moves one such member, the DIE type hash, to the
new structure.

In this patch, the new object is managed by the dwarf2_per_objfile.
However, by the end of the series, this will change.

gdb/ChangeLog
2020-02-15  Tom Tromey  <tom@tromey.com>

	* dwarf2/read.h (struct dwarf2_unshareable): New type.
	(struct dwarf2_per_objfile) <die_type_hash>: Move to
	dwarf2_unshareable.
	<unshareable>: New member.
	* dwarf2/read.c (dwarf2_per_objfile::dwarf2_per_objfile):
	Initialize "unshareable".
	(dwarf2_per_objfile::~dwarf2_per_objfile): Delete unshareable.
	(set_die_type, get_die_type_at_offset): Update.
---
 gdb/ChangeLog     | 11 +++++++++++
 gdb/dwarf2/read.c | 19 +++++++++++++------
 gdb/dwarf2/read.h | 28 +++++++++++++++++++++++-----
 3 files changed, 47 insertions(+), 11 deletions(-)

Just one nit...

On 2/15/20 1:54 PM, Tom Tromey wrote:
> The goal of this series is to share dwarf2_per_objfile across
> objfiles.  However, this is difficult to do, because it has some
> members that are inherently per-objfile.
> 
> This patch introduces a new "unshareable" object.  The idea here is
> that we'll have one such object per objfile, and we'll temporarily
> update the pointer when needed.  This is a bit ugly, but it's
> convenient, because it doesn't require a huge rewriting of the DWARF
> code.  (I would be in favor of such a rewriting, but ... it's huge.)
> 
> This patch further moves one such member, the DIE type hash, to the
> new structure.
> 
> In this patch, the new object is managed by the dwarf2_per_objfile.
> However, by the end of the series, this will change.
> 
> gdb/ChangeLog
> 2020-02-15  Tom Tromey  <tom@tromey.com>
> 
> 	* dwarf2/read.h (struct dwarf2_unshareable): New type.
> 	(struct dwarf2_per_objfile) <die_type_hash>: Move to
> 	dwarf2_unshareable.
> 	<unshareable>: New member.
> 	* dwarf2/read.c (dwarf2_per_objfile::dwarf2_per_objfile):
> 	Initialize "unshareable".
> 	(dwarf2_per_objfile::~dwarf2_per_objfile): Delete unshareable.
> 	(set_die_type, get_die_type_at_offset): Update.
> ---
>   gdb/ChangeLog     | 11 +++++++++++
>   gdb/dwarf2/read.c | 19 +++++++++++++------
>   gdb/dwarf2/read.h | 28 +++++++++++++++++++++++-----
>   3 files changed, 47 insertions(+), 11 deletions(-)
> 
> diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
> index a897b11693a..cb04eb19beb 100644
> --- a/gdb/dwarf2/read.c
> +++ b/gdb/dwarf2/read.c
> @@ -1774,7 +1774,10 @@ dwarf2_per_objfile::dwarf2_per_objfile (struct objfile *objfile_,
>   					const dwarf2_debug_sections *names,
>   					bool can_copy_)
>     : objfile (objfile_),
> -    can_copy (can_copy_)
> +    can_copy (can_copy_),
> +    /* Temporarily just allocate one per objfile -- we don't have
> +       sharing yet.  */
> +    unshareable (new dwarf2_unshareable)
>   {
>     if (names == NULL)
>       names = &dwarf2_elf_names;
> @@ -1796,6 +1799,8 @@ 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.  */
>   }
>   
> @@ -24434,8 +24439,8 @@ set_die_type (struct die_info *die, struct type *type, struct dwarf2_cu *cu)
>   			    cu->per_cu->addr_type ()))
>       add_dyn_prop (DYN_PROP_DATA_LOCATION, prop, type);
>   
> -  if (dwarf2_per_objfile->die_type_hash == NULL)
> -    dwarf2_per_objfile->die_type_hash
> +  if (dwarf2_per_objfile->unshareable->die_type_hash == NULL)
> +    dwarf2_per_objfile->unshareable->die_type_hash
>         = htab_up (htab_create_alloc (127,
>   				    per_cu_offset_and_type_hash,
>   				    per_cu_offset_and_type_eq,
> @@ -24445,7 +24450,8 @@ set_die_type (struct die_info *die, struct type *type, struct dwarf2_cu *cu)
>     ofs.sect_off = die->sect_off;
>     ofs.type = type;
>     slot = (struct dwarf2_per_cu_offset_and_type **)
> -    htab_find_slot (dwarf2_per_objfile->die_type_hash.get (), &ofs, INSERT);
> +    htab_find_slot (dwarf2_per_objfile->unshareable->die_type_hash.get (),
> +		    &ofs, INSERT);
>     if (*slot)
>       complaint (_("A problem internal to GDB: DIE %s has type already set"),
>   	       sect_offset_str (die->sect_off));
> @@ -24465,13 +24471,14 @@ get_die_type_at_offset (sect_offset sect_off,
>     struct dwarf2_per_cu_offset_and_type *slot, ofs;
>     struct dwarf2_per_objfile *dwarf2_per_objfile = per_cu->dwarf2_per_objfile;
>   
> -  if (dwarf2_per_objfile->die_type_hash == NULL)
> +  if (dwarf2_per_objfile->unshareable->die_type_hash == NULL)
>       return NULL;
>   
>     ofs.per_cu = per_cu;
>     ofs.sect_off = sect_off;
>     slot = ((struct dwarf2_per_cu_offset_and_type *)
> -	  htab_find (dwarf2_per_objfile->die_type_hash.get (), &ofs));
> +	  htab_find (dwarf2_per_objfile->unshareable->die_type_hash.get (),
> +		     &ofs));
>     if (slot)
>       return slot->type;
>     else
> diff --git a/gdb/dwarf2/read.h b/gdb/dwarf2/read.h
> index 5fa8dec9bfb..06a0fa3a19a 100644
> --- a/gdb/dwarf2/read.h
> +++ b/gdb/dwarf2/read.h
> @@ -66,6 +66,24 @@ struct dwarf2_queue_item
>     enum language pretend_language;
>   };
>   
> +/* Some DWARF data cannot (currently) be shared across objfiles.  Such
> +   data is stored in this object.
> +
> +   dwarf2_per_objfile holds a pointer to an instance of this object.
> +   This pointer is temporarily set when expanding CUs.  This hackish
> +   approach is due to the history of the DWARF reader.  In the past,
> +   all objects were stored per-objfile, and this object was introduced
> +   in the process of separating out the shareable and per-objfile
> +   state.  */
> +
> +struct dwarf2_unshareable
> +{
> +  /* Table mapping type DIEs to their struct type *.
> +     This is NULL if not allocated yet.

"This is nullptr"

> +     The mapping is done via (CU/TU + DIE offset) -> type.  */
> +  htab_up die_type_hash;
> +};
> +
>   /* Collection of data recorded per objfile.
>      This hangs off of dwarf2_objfile_data_key.  */
>   
> @@ -214,11 +232,6 @@ public:
>        symbols.  */
>     bool reading_partial_symbols = false;
>   
> -  /* Table mapping type DIEs to their struct type *.
> -     This is NULL if not allocated yet.
> -     The mapping is done via (CU/TU + DIE offset) -> type.  */
> -  htab_up die_type_hash;
> -
>     /* The CUs we recently read.  */
>     std::vector<dwarf2_per_cu_data *> just_read_cus;
>   
> @@ -241,6 +254,11 @@ public:
>   
>     /* CUs that are queued to be read.  */
>     std::queue<dwarf2_queue_item> queue;
> +
> +  /* State that cannot be shared across objfiles.  This is normally
> +     nullptr and is temporarily set to the correct value at the entry
> +     points of the reader.  */
> +  dwarf2_unshareable *unshareable = nullptr;
>   };
>   
>   /* Get the dwarf2_per_objfile associated to OBJFILE.  */
>

On 2020-02-15 11:54 a.m., Tom Tromey wrote:
> The goal of this series is to share dwarf2_per_objfile across
> objfiles.  However, this is difficult to do, because it has some
> members that are inherently per-objfile.
> 
> This patch introduces a new "unshareable" object.  The idea here is
> that we'll have one such object per objfile, and we'll temporarily
> update the pointer when needed.  This is a bit ugly, but it's
> convenient, because it doesn't require a huge rewriting of the DWARF
> code.  (I would be in favor of such a rewriting, but ... it's huge.)
> 
> This patch further moves one such member, the DIE type hash, to the
> new structure.
> 
> In this patch, the new object is managed by the dwarf2_per_objfile.
> However, by the end of the series, this will change.

So, in the end we'll have dwarf2_per_objfile, of which there will be only
one and will be shared between the objfiles, and dwarf2_unshareable of which
there will be N (for N objfiles), which won't be shared between the objfiles?

I find it a bit confusing, because now the object called "per_objfile" is not
per objfile.

Simon

>>>>> "Simon" == Simon Marchi <simark@simark.ca> writes:

We discussed this online but I thought I'd reply for the record.

>> In this patch, the new object is managed by the dwarf2_per_objfile.
>> However, by the end of the series, this will change.

Simon> So, in the end we'll have dwarf2_per_objfile, of which there will be only
Simon> one and will be shared between the objfiles, and dwarf2_unshareable of which
Simon> there will be N (for N objfiles), which won't be shared between the objfiles?

Yes.

Simon> I find it a bit confusing, because now the object called "per_objfile" is not
Simon> per objfile.

Yeah.  In the cover letter I mentioned perhaps renaming this object.
That could be done with sed I suppose.

Tom