Delegate contravariance

[Christopher Wright]

Jason House wrote:
> Silvio Ricardo Cordeiro wrote:
> 
>> Is there any good reason why the following code doesn't work?
>> The function "foo" requires as its argument a delegate that
>> receives a B. This means that, because of the type soundness
>> of the D language, the delegate will only be called with instances
>> of B. Now, why can't it be "abc", then? The "abc" delegate just
>> happens to handle more than it is required...
>>
>>
>> class A { }
>> class B : A { }
>> void foo(void delegate(B b) dg) {}
>>
>> auto abc = (A a) { }
>> auto qwe = (B b) { }
>>
>> void main() {
>> foo(abc); // this line won't compile
>> foo(qwe);
>> }
>>
>>
>> BTW: Same thing happens for const(B). B "sort of" a derivative
>> of const(B), so it would make sense. Every possible argument that can
>> be passed to a delegate(B) can also be passed to a delegate(A) or
>> delegate(const(B)). Why the restriction, then? Am I missing something?
> 
> Your code above is wrong, but you are right about a bug in dmd.  
> 
> Your code sample should not compile because of two issues:
> 1. You're missing semicolons on the declarations of abc and qwe.
> 2.  foo(abc) is implicitly requires casting A's to B's which is not guaranteed to be correct.  

No, it implicitly casts Bs (derived) to As (base). The original example 
was correct (modulo semicolon issues). Your example has implicit casts 
of a base class to a derived class.

On the other hand, out parameters and return values have to be handled 
in the reverse manner, and ref parameters have to use exact matching.