Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!usc!apple!landon
From: landon@Apple.COM (Landon Dyer)
Newsgroups: comp.lang.c++
Subject: Pure virtual functions considered broken
Message-ID: <36904@apple.Apple.COM>
Date: 30 Nov 89 23:55:24 GMT
Organization: Apple Computer Inc., Cupertino, CA
Lines: 77


Nano-summary:	Pure virtual functions suffer from an apparently arbitrary
		language design decision.


Simplified, I have a three-layer class hierarchy, something like this:


        TBasicBehaviour     -- abstract base, common behaviour for all objects
         /
      TMumble               -- a more specific abstract base
       /
    TRealMumble             -- a class that actually gets instantiated


TBasicBehaviour has a number of pure virtual functions.  Derived from this
class are additional abstract bases that add more behaviour, and derived from
THOSE are the classes of which instances are actually made.

In CFront it's necessary to (re-)declare a pure virtual function of a parent
class, even if the subclass itself is abstract.  [C++ Reference Manual,
section 10.3, p68 in my copy.]


The problem is that if you have MANY derived abstract bases (and why not?),
this requirement can become painful to maintain.  Furthermore, the requirement
appears to be completely arbitrary -- a warning, rather than a hard error
would be sufficient for "mothering" purposes.


So there are a number of possibilities:

    o  It's impossible to lift the restriction for obscure reasons;
    o  It's another CFront 2.0 bug;
    o  This is the way "it turned out";
    o  It really is a feature;
    o  I'm completely mistaken and deserve enlightenment ["Oww!"].

I'm willing to listen to arguments about how CFront is doing the right thing,
or about how the current behaviour is merely part of CFront's implementation,
and will be fixed someday.

It WOULD be nice to see a rationale ...


----------------

Here's an example, which you can skip if you like.  (You'll have to flatten
this 2-D code to compile it...)


	struct A {
	  virtual f() = 0;
	};

	struct B1 : A {		|	struct B2 : A {		// etc ==>
	  virtual f() = 0;	|	  virtual f() = 0;
	  virtual g() = 0;	|	  virtual h() = 0;
	};			|	};

	struct C1 : B1 {	|	struct C2 : B2 {	// etc ==>
	  virtual f();		|	  virtual f();
	  virtual g();		|	  virtual h();
	};			|	};


IMHO, The redeclarations B1::f() and B2::f() should not be required.


Peace,

-- 

-----------------------------------------  "Mmmph!  Urghurmph!  Grugmph!"
Landon Dyer, Apple Computer, Inc.          "What's he trying to say?"
Development Systems Group (MPW / DSG)      "I dunno -- there's a lawyer
NOT THE VIEWS OF APPLE COMPUTER              crammed in his mouth."