Re: C++/Java regressions

[Daniel Jacobowitz <drow at mvista dot com>]

On Tue, Nov 25, 2003 at 11:04:41PM -0500, Ian Lance Taylor wrote:
> Then when gdb goes to look up T5<int>, it finds the DWARF psymbol for
> the class itself, but it also finds that the symbol might be a
> namespace.  It then decides that it is a namespace.

That sounds like the bit that's wrong.  If it found a psymbol for the
class why does it decide to call it a namespace?

> The problem with t5i, besides some demangler changes which have
> already been addressed, is that gdb doesn't recognize that a
> constructor function is, in fact, a constructor.  That's because the
> code in c-typeprint.c does this:
> 
> 	      char *method_name = TYPE_FN_FIELDLIST_NAME (type, i);
> 	      char *name = type_name_no_tag (type);
> 	      int is_constructor = name && strcmp (method_name, name) == 0;
> 
> and this:
> 
> 		  char *physname = TYPE_FN_FIELD_PHYSNAME (f, j);
> 		  int is_full_physname_constructor =
> 		   is_constructor_name (physname) 
> 		   || is_destructor_name (physname)
> 		   || method_name[0] == '~';
> 
> The first fails because name is "T5<int>" and method_name is "T5".
> The second fails because this is dealing with debugging information,
> and gcc does not emit any physical symbol name information for class
> constructors.
> 
> This could be addressed by changing the simple test for is_constructor
> to a more complex test which ignore the template arguments when
> determining whether method_name was the same as name.

Yes, I have a patch sitting here (for some time) which does this.  It
got mixed in with some other patches that aren't ready for prime-time
yet.

I'm considering a routine for canonicalization of C++ demangled names. 
That's the only way we're going to get templates.exp to start working
robustly.  For cv-qualifiers it's fairly easy, and most of what gets
mangled is to accomodate either space efficiency or linker character
sets so that's not all so important.  "Unsigned" <-> "unsigned int" is
also easy.  Template arguments can be a little harder so we may punt.

Perhaps next month.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer