Re: [PATCH 02/23] dwarf: add DW_AT_data_location support

[Keven Boell <keven dot boell at linux dot intel dot com>]

Hi Joel,

I've addressed your comments and the changes will be part of patch series v2.

Keven


> Hello,
> 
>> An object might have a descriptor proceeding the actual value.
>> To point the debugger to the actually value of an object
>> DW_AT_data_location is used for. For example the compile may
>> emit for this entity:
>>
>>   1| int foo[N];
>>
>> the following descriptor:
>>
>> struct array {
>>   size_t size;
>>   void*  data; // DW_AT_data_location describes this location
>> }
>>
>> This allows GDB to print the actual data of an type.
>>
>> 2014-05-28  Sanimir Agovic  <sanimir.agovic@intel.com>
>>             Keven Boell  <keven.boell@intel.com>
>>
>> 	* dwarf2read.c (set_die_type): Parse and save DW_AT_data_location
>> 	attribute.
>> 	* gdbtypes.c (is_dynamic_type): Consider a type being dynamic if
>> 	the data location has not yet been resolved.
>> 	(resolve_dynamic_type): Evaluate data location baton
>> 	if present and save its value.
>> 	* gdbtypes.h <main_type>: Add data_location.
>> 	(TYPE_DATA_LOCATION): New macro.
>> 	(TYPE_DATA_LOCATION_ADDR): New macro.
>> 	(TYPE_DATA_LOCATION_IS_ADDRESS): New macro.
>> 	* value.c: Include dwarf2loc.h.
>> 	(value_fetch_lazy): Use data location addres to read value from
>> 	memory.
>> 	(coerce_ref): Construct new value from data location.
> 
> Here are some comments and questions, but it would be nice,  IMO, if
> the patch was also reviewed by Tom, particularly since it introduces
> a new field in struct main_type, which is a size-sensitive, I think.
> 
> Also, it would be nice if you could include a copy of the actual
> DWARF output in the revision log of your patch, for easy reference
> of what you're trying to support.
> 
>> diff --git a/gdb/dwarf2read.c b/gdb/dwarf2read.c
>> index 6ebfffc..7a0f7f4 100644
>> --- a/gdb/dwarf2read.c
>> +++ b/gdb/dwarf2read.c
>> @@ -21499,6 +21499,7 @@ set_die_type (struct die_info *die, struct type *type, struct dwarf2_cu *cu)
>>  {
>>    struct dwarf2_per_cu_offset_and_type **slot, ofs;
>>    struct objfile *objfile = cu->objfile;
>> +  struct attribute *attr;
>>  
>>    /* For Ada types, make sure that the gnat-specific data is always
>>       initialized (if not already set).  There are a few types where
>> @@ -21513,6 +21514,20 @@ set_die_type (struct die_info *die, struct type *type, struct dwarf2_cu *cu)
>>        && !HAVE_GNAT_AUX_INFO (type))
>>      INIT_GNAT_SPECIFIC (type);
>>  
>> +  /* Read DW_AT_data_location and set in type.  */
>> +  attr = dwarf2_attr (die, DW_AT_data_location, cu);
>> +  if (attr_form_is_block (attr))
> 
> Here (in set_die_type), why do you limit the processing the block-form
> data_location attributes? Why not just call attr_to_dynamic_prop
> regardless of the form, and let that function deal with whatever
> the form is? In other words, trop the form check and just keep
> the following bit:
> 
>> +      struct dynamic_prop prop;
>> +
>> +      if (attr_to_dynamic_prop (attr, die, cu, &prop))
>> +	{
>> +	  TYPE_DATA_LOCATION (type)
>> +	    = obstack_alloc (&objfile->objfile_obstack, sizeof (prop));
>> +	  *TYPE_DATA_LOCATION (type) = prop;
>> +	}

Good catch, there is no reason to limit the processing here.

> 
>> --- a/gdb/gdbtypes.c
>> +++ b/gdb/gdbtypes.c
>> @@ -1635,8 +1635,12 @@ is_dynamic_type (struct type *type)
>>  	   or the elements it contains have a dynamic contents.  */
>>  	if (is_dynamic_type (TYPE_INDEX_TYPE (type)))
>>  	  return 1;
>> -	else
>> -	  return is_dynamic_type (TYPE_TARGET_TYPE (type));
>> +	else if (TYPE_DATA_LOCATION (type) != NULL
>> +          && (TYPE_DATA_LOCATION_KIND (type) == PROP_LOCEXPR
>> +              || TYPE_DATA_LOCATION_KIND (type) == PROP_LOCLIST))
>> +    return 1;
>> +  else
>> +    return is_dynamic_type (TYPE_TARGET_TYPE (type));
> 
> The comment needs to be updated. Your change is also splitting
> the "if bounds or contents is dynamic" logic. Perhaps you could
> simply add the data-location check at the start of the function
> with a command that says: A type which has a data_location which
> [bla bla] is a dynamic type

I've updated the comment.

> 
> 
>> +  type = resolved_type;
>> +
>> +  /* Resolve data_location attribute.  */
>> +  prop = TYPE_DATA_LOCATION (type);
>> +  if (dwarf2_evaluate_property (prop, addr, &value))
>> +    {
>> +      TYPE_DATA_LOCATION_ADDR (type) = value;
>> +      TYPE_DATA_LOCATION_KIND (type) = PROP_CONST;
>> +    }
>> +  else
>> +    TYPE_DATA_LOCATION (type) = NULL;
> 
> Here, I do not understand why you override TYPE, instead of just
> using RESOLVED_TYPE directly.

I did this for improving the readability, however maybe this is just confusing. I changed it.

> 
>> @@ -722,6 +722,10 @@ struct main_type
>>  
>>      struct func_type *func_stuff;
>>    } type_specific;
>> +
>> +  /* * Indirection to actual data.  */
>> +
>> +  struct dynamic_prop *data_location;
> 
> I'd like to have a comment which is a little more descriptive of
> what this field contains. Someone who reads this comment should be
> able to understand it without having to grep for this field to
> get an idea of how this field is used.

Done.

> 
>> +/* Attribute accessors for VLA support.  */
> 
> I think this comment is too specific. Although this field is
> instroduced to support VLAs, descriptors can probably be used
> in other contexts. I don't think you really need a comment,
> in this case, though.

Done.

> 
>> +#define TYPE_DATA_LOCATION(thistype) \
>> +  TYPE_MAIN_TYPE(thistype)->data_location
>> +#define TYPE_DATA_LOCATION_BATON(thistype) \
>> +  TYPE_DATA_LOCATION (thistype)->data.baton
>> +#define TYPE_DATA_LOCATION_ADDR(thistype) \
>> +  TYPE_DATA_LOCATION (thistype)->data.const_val
>> +#define TYPE_DATA_LOCATION_KIND(thistype) \
>> +  TYPE_DATA_LOCATION (thistype)->kind
>> +
>>  /* Moto-specific stuff for FORTRAN arrays.  */
>>  
>>  #define TYPE_ARRAY_UPPER_BOUND_IS_UNDEFINED(arraytype) \
>> diff --git a/gdb/value.c b/gdb/value.c
>> index d125a09..1c88dfd 100644
>> --- a/gdb/value.c
>> +++ b/gdb/value.c
>> @@ -3635,8 +3635,14 @@ value_fetch_lazy (struct value *val)
>>      }
>>    else if (VALUE_LVAL (val) == lval_memory)
>>      {
>> -      CORE_ADDR addr = value_address (val);
>>        struct type *type = check_typedef (value_enclosing_type (val));
>> +      CORE_ADDR addr;
>> +
>> +      if (TYPE_DATA_LOCATION (type) != NULL
>> +	  && TYPE_DATA_LOCATION_KIND (type) == PROP_CONST)
>> +	addr = TYPE_DATA_LOCATION_ADDR (type);
>> +      else
>> +	addr = value_address (val);
> 
> I am wondering if this part shouldn't be in value_address itself.
> WDYT? Tom?
> 
>>  
>>        if (TYPE_LENGTH (type))
>>  	read_value_memory (val, 0, value_stack (val),
>> -- 
>> 1.7.9.5
>