Re: [RFA] dwarf2read.c: set TYPE_DOMAIN_TYPE correctly for methods

[David Carlton <carlton at math dot stanford dot edu>]

In article <20020822204211.GA31727@nevyn.them.org>, Daniel Jacobowitz <drow@mvista.com> writes:

>> * dwarf2read.c (dwarf2_add_member_fn): Add back in the type
>> argument that was deleted on 2002-06-14: it was needed after all,
>> as PR gdb/653 demonstrates.  Update call to smash_to_method_type.
>> (read_structure_scope): Update call to dwarf2_add_member_fn.

> Can you explain why this is necessary?  I could not find any path to
> that point where type != die->type.

dwarf2_add_member_fn is called by read_structure_scope.  The 'die'
argument to dwarf2_add_member_fn is the 'child_die' variable of
read_structure_scope.  So what dwarf2_add_member_fn thinks of as
'die->type' is what read_structure_scope thinks of as
'child_die->type', in other words the type associated to the method
that is being added to the structure in question.

But the second argument to smash_to_method_type is supposed to be the
domain of the given method, i.e. the type of the structure owning the
method.  This is dramatically different from read_structure_scope's
child_die->type; it equals read_structure_scope's die->type.

So that's why we want to pass in read_structure_scope's die->type as
an argument.  (I don't see any way to recover die->type from
child_die, though admittedly I didn't try very hard.)

For a reasonably dramatic illustration of the effects of this, try
compiling the following program

/* Create some objects, and try to print out their methods.  */

class A {
public:
  virtual void virt() {};
  void nonvirt() {};
};

int main()
{
  A *theA = new A;

  return 0;
}

and debug it with GDB from the current CVS.  Run it until theA has
been created, and then try 'p theA->virt'.

David Carlton
carlton@math.stanford.edu