Is this a bug?

[John Colvin via Digitalmars-d]

On Thursday, 26 June 2014 at 10:09:53 UTC, Rene Zwanenburg wrote:
> I /think/ this is a bug, but I'm not 100% sure. The following 
> compiles without any problems, as it should:
>
> import std.typecons;
>
> alias Handle(T) = RefCounted!(T, RefCountedAutoInitialize.no);
>
> auto initialized(T)() if(is(T == RefCounted!S, S...))
> {
> 	T refCounted;
> 	refCounted.refCountedStore.ensureInitialized();
> 	return refCounted;
> }
>
> alias S = Handle!S_Impl;
> struct S_Impl
> {
> }
>
> void main()
> {
> 	auto s = initialized!S;
> }
>
>
> Change 'initialized' to:
> auto initialized(T)() if(is(T == Handle!S, S))
> and the compiler will complain:
>
> Error: template instance test.initialized!(RefCounted!(S_Impl, 
> cast(RefCountedAutoInitialize)0)) does not match template 
> declaration initialized(T)() if (is(T == Handle!S, S))

I don't think that's a bug, but I think the template constraint 
itself shouldn't be legal to write. It's not possible to support 
the general case of a template A in is(T == A!X, X) as it would 
require following all possible instantiation paths of A.

The compiler could keep track of how a type was created on a 
per-identifier basis and then use that to check, but that would 
mean rejecting types that had a different instantiation path 
(e.g. is(RefCounted!(int, RefCountedAutoInitialize.no) == 
Handle!S, S) == false).

What might be feasible would be to extend is(T == A!X, X) to work 
with simple tmeplates and paramterised aliases like Handle, while 
statically disallowing more complicated / undecidable ones..