(ref.doc)maxtal 150395

Newsgroups: comp.std.c++     (Note:
 maxtal.)
Subject: Re: Partial specializations? (Was: Sealed Classes)
Date: Wed, 15 Mar 1995 21:33:35 GMT

>John, could you, or someone else in the know, let the rest of us know
>more or less what "partial specializations" are?

	Sure. Even better, I've been told that they _have_ been accepted
at least in some form. First, consider a class template:

	template<class T, class U> class X { .. };

As you know you can write a specialisation:

	class X<int, int> { ... };

But what if you only want to specialise the first argument?
Can you write:

	template<class U> class X<int, U> { .. };

?? Well, the answer is: two weeks ago you could not, but 
I believe now you can. Here's another example:

	template<class T> class X<T*, T***> { ... }

The resolution of which definition to use is done by
asking which of the "generic bindings" which might
apply to a particular instance is most specific (i.e. most
specialised). For example:

	<T*, T***> is more special than <A, B**>

Thats classes. Now consider the problem:

	template<class T> void f(T);
	template<class P> void f(P*);
	int i;
	f(&i); // Ambiguous??

Both templates can generate a definition of 

	f(int*)

Which one should be used? Two weeks ago, this was ambiguous,
now the answer is "the second template, because it is
more specialised".

You will note "specialisations" form a DAG. Do not be
surprised at the analogy with (multiple) inheritance: virtual
function overriding is exactly the same kind of specialisation
as template specialisations. (The principal difference is
the binding time)
-----------------------------------------------------------------

I'm not aware of the exact rules. I'm sure there will be some
interesting problems, however. But one thing is for sure:
templates without partial specialisations would fail completely
to be orthogonal enough to construct significant libraries.

In particular, partial specialisations are necessary to make STL
maximally useful.