On Wednesday, 30 August 2017 at 21:33:30 UTC, Jonathan M Davis
wrote:
On Wednesday, August 30, 2017 20:47:12 EntangledQuanta via
Digitalmars-d- learn wrote:
This is quite surprising!
public struct S(T)
{
T s;
}
interface I
{
void Go(T)(S!T s);
static final I New()
{
return new C();
}
}
abstract class A : I
{
}
class C : A
{
void Go(T)(S!T s)
{
}
}
void main()
{
S!int s;
auto c = I.New();
c.Go(s); // fails!
//(cast(C)c).Go(s); // Works, only difference is we have
made c
an explicit C.
}
https://dpaste.dzfl.pl/dbc5a0663802
Everything works when Go is not templatized(we explicitly make
T
an int)
This is a blocker for me! Can someone open a ticket?
It is not possible to have a function be both virtual and
templated. A function template generates a new function
definition every time that it's a called with a new set of
template arguments. So, the actual functions are not known up
front, and that fundamentally does not work with virtual
functions, where the functions need to be known up front, and
you get a different function by a look-up for occurring in the
virtual function call table for the class. Templates and
virtual functions simply don't mix. You're going to have to
come up with a solution that does not try and mix templates and
virtual functions.
- Jonathan M Davis
I have a finite number of possible values of T, lets say 3. They
are known at compile time, just because you are or D thinks they
are not simply means you or D is not trying hard enough. So,
saying that virtual methods and templates are not compatible is
wrong. Just because you think they are or D thinks they are means
you haven't thought about it hard enough.
If I can overload a virtual function to get all my use cases and
that is all I need then I **should** be able to do it with
templates. Simple as that, if D can't do that then D needs to be
enhanced to do so.
e.g.,
class C
{
Go(Primitive!T)(T t);
}
The compiler can realize that T can only be a primitive, and
generates all possible combinations of primitives, which is
finite. This is doable, it is not impossible, regardless of what
you think. It is equivalent to
class C
{
Go(Primitive1 t);
Go(Primitive2 t);
...
Go(PrimitiveN t);
}
In fact, we can use string mixins to generate such code, but it
doens't save us trouble, which is what templates are suppose to
do in the first place. Just become someone hasn't implemented
special cases does not mean it is theoretically impossible to do.
A different syntax would be better
interface I
{
Go(T in [float, double, int])(T t);
}
class C : I
{
Go(T in [float, double, int])(T t) { }
}
which the compiler "unrolls" to
interface I
{
Go(float t);
Go(double t);
Go(int t);
}
class C
{
Go(float t) { }
Go(double t) { }
Go(int t) { }
}
Which, is standard D code. There is nothing wrong with
specializing the most common cases.
The point you are trying to making, and not doing a great job, is
that the compiler cannot create an unknown set of virtual
functions from a single templated virtual function. BUT, when you
realize that is what the problem is, the unknown set is the issue
NOT templated virtual functions. Make the set known and finite
somehow then you have a solution, and it's not that difficult.
Just requires some elbow grease.
Primitives are obviously known at compile time so that is a
doable special case. Although there will probably be quite a bit
of wasted space since each primitive will have a function
generated for it for each templated function, that really isn't
an issue.
By adding a new syntax in D, we could allow for any arbitrary(but
known and finite) set to be used
Go(T in [A,B,C])(T t)
Where A,B,C are known types at compile time. This generates 3
functions and is doable. (should be simple for any D compiler
genius to add for testing)