Re: [RFA] Patch PR debug/41065

[Dodji Seketeli <dodji at redhat dot com>]

Le 21/09/2009 17:08, Jason Merrill a écrit :
> On 09/01/2009 05:22 PM, Dodji Seketeli wrote:
> > +hashval_t
> > +types_used_by_vars_do_hash (const void *x)
> > +{
> > + return htab_hash_pointer
> > + (((const struct types_used_by_vars_entry*)x)->type);
> > +}
> > +
> > +/*Equality function of the types_used_by_vars_entry hash table. */
> > +
> > +int
> > +types_used_by_vars_eq (const void *x1, const void *x2)
> > +{
> > + return htab_eq_pointer
> > + (((const struct types_used_by_vars_entry *)x1)->type,
> > + (const union tree_node*) x2);
>
> Why don't these hash/compare the var_decl?

My understanding is that we want to keep the type used by the var_decl 
because the var_decl itself might loose the information. For instance, in 
the example of the patch, the information about v being initialized by the 
CONST_DECL 'a' is lost as we only keep the integral value of 'a' in 
finish_id_expression:

 /* Similarly, we resolve enumeration constants to their 

     underlying values.  */ 

  else if (TREE_CODE (decl) == CONST_DECL)
    {
      *idk = CP_ID_KIND_NONE;
      if (!processing_template_decl) 

        {
          used_types_insert (TREE_TYPE (decl)); 

          return DECL_INITIAL (decl); 

        }
      return decl;
    }

Hence the type hashing, rather than the var_decl hashing.

--
Dodji Seketeli
Red Hat, Inc.