member pointers

[Kirk McDonald]

Frits van Bommel wrote:
> Daniel Keep wrote:
>> bc wrote:
>>> am i being stupid or is there any chance of adding pointers to member
>>> functions and pointers to data members to D? i can't see any way of
>>> getting the same functionality through existing language constructs for
>>> member functions... alias seems like it might at first but i can't get
>>> it to work... and the offsetof syntax for data members seems buggy.
>>> otherwise, congrats on designing a really cool language
>>
> [snip code]
>>
>> Delegates are your pointer to member functions.  Pointers to data of an
>> object are just as you would expect.  Or did you have a different use in
>> mind?
> 
> In C++, a "pointer to member" is a pointer to a _class_ member, not an 
> _instance_ member. That pointer can then be combined with an instance 
> and dereference to the member of that class. If the pointer is to a 
> virtual function, calling the function pointed to will call the correct 
> override for that instance. (the pointer type is a combination of type 
> of the class and the type of the member)
> So there are basically three cases:
> * pointer to data member: equivalent to typed offset into object
> * virtual function: equivalent to typed offset into vtable
> * non-virtual function: equivalent to function pointer (but still 
> requires "this" pointer)
> 
> As far as I know, D has no direct equivalent to this, though pointers to 
> data members could probably be trivially written as a template struct.
> One complication with member function pointers is that the latter two 
> cases above must both fit into the same type (so a pointer to a member 
> function must be able to reference both virtual and non-virtual functions).
> The compiler could probably easily implement this though. One way to 
> make this easier to implement would be to give non-virtual functions 
> vtable offsets as well, so the pointer can be just an offset into the 
> vtable instead of a struct or "tagged union" type of thing. That way 
> they can also remain the same size as a size_t.
> Note: pointers to member functions can be up to 16 bytes in some popular 
> C++ compilers, though C++ does need to handle a few case that D doesn't 
> like multiple inheritance and classes that have only been 
> forward-declared, as well as classes that aren't allowed to contain a 
> vtable pointer)

It is interesting to note that Pyd exploits D's behavior w.r.t. pointers 
to virtual member functions to implement its class wrapping. That is, 
given a base class and a derived class:

class Base {
     void foo() {}
}

class Derived : Base {
     void foo() {}
}

You can call Base.foo on an instance of Derived.foo with some delegate 
trickery:

void delegate() dg;
dg.ptr = new Derived;
dg.funcptr = &Base.foo;
dg();

The reasons Pyd needs to do this are fairly complex, and I don't feel 
like explaining them in this space. Suffice to say, it's a good reason, 
and certain parts of Pyd's class wrapping implementation would be even 
more complex without it.

To perhaps put minds at ease, I will say that the context in which it 
does this is fairly well guaranteed to be safe. (The derived class is 
generated by Pyd, so I can be pretty sure it's not doing anything to 
blow this up.)

-- 
Kirk McDonald
http://kirkmcdonald.blogspot.com
Pyd: Connecting D and Python
http://pyd.dsource.org