tests/org.eclipse.ocl.examples.impactanalyzer.testmodel.ngpm/rose-model/Polymorphism.txt

Polymorphic operation calls:
----------------------------

An OperationExpression references an Operation in the
metamodel. However, this is only the static view. During runtime, a
different operation that conforms to that referenced by the
OperationExpression may be invoked, depending on the actual type of
the object on which the operation is called.

The type system guarantees that when the code calling an operation on
an object has passed the compiler checks, it will find an operation on
the actual type of the object on which the operation has to be invoked
at runtime that at least conforms to the operation referenced at
compile time. The operation is found either on the type of the object
itself, on a TypeAdapter of that type or on a type to which the
object's type delegates.

Polymorphic association access:
-------------------------------

Requirement: A structure that is a superstructure of another and that
uses equal names to identify corresponding connecting elements
(probably the association ends) should be considered "conforming" to
the substructure. Rationale: All navigations possible in the
substructure are also possible in the superstructure and yield
conforming elements.

Caveat: Taken together with the fact that conforming instances can be
inserted into an end, navigating back may use a different
association.

Question: Can users easily determine whether an exposed association end
conforms to another one? Tool support could help.

Discussion:

When an AssociationNavigationExpression references an AssociationEnd,
at runtime the navigation may use a conforming end attached to the
actual type of the object on which the navigation is started.
Otherwise, it would not be possible to pass a conforming structure and
navigate it using (statically) the association ends of the structure
to which the passed one conforms.

When using an AddLink or RemoveLink statement, this may also have to
work polymorphically w.r.t the association being used. Example:

class A {
  a <-AtoB-> B b; // bidirectional assoc, end names "a" and "b"
}

class B {
  // no exposed ends
}

// C conforms to A because it exposes a conforming association end
class C {
  c <-CtoB-> B b; // exposes an end conforming to AtoB.b
}

B b1 = ...;
C c1 = ...;
A a1 = c1;

AddLink(AtoB, a1, b1); // dynamically, this would probably have to
                       // fill CtoB because a1 is bound to an object
		       // of type C.

If there was no CtoB association, C would have to expose the B end of
the AtoB association in order to conform to A. If A did not expose the
b end, C would conform to A without exposing the b end. In both cases
it would be possible to call AddLink(AtoB, a1, b1) such that the C
instance bound to a1 and the B instance bound to b1 are then linked.

A problem with polymorphic associations could be "catcalls" like with
arrays in Java. Therefore we have forbidden covariant types of
association ends that could polymorphically be used instead of
another.

The easiest option seems to be to require that "conforming" for
exposed ends only means that the *same* end can be exposed in the
conforming class. Otherwise, multiple associations would point to the
same type with one end but only to conforming types with the other
end. Then, what are the semantics for navigating from the common type
to the other end? If the one other end does not conform to the other
other end, they could probably not be used polymorphically, and thus
computing the union would not make sense. Instead, the program would
have to disambiguate the reference to the association end, or
otherwise it would always mean the end that leads to the class to
which the other one conforms (the "superclass").

The issue occurs w.r.t the conformance that depends on exposed ends
which in turn is important for nested structures and their
conformance. It seems intuitive that two structurally equivalent types
A and B should conform to each other. However, again intuitively, this
wouldn't require the associations that they use to define their
internal structure to be the same but only conform to each other. This
probably assumes that such a hierarchical object is not operated
through the associations themselves but rather through the (exposed
only?) ends ("properties") which then need to conform to each
other. In this sense, using an exposed end to access an association
must work polymorphically.

Solution proposal:

There is no notion of conformance of associations with other
associations. However, association ends may conform to other
association ends:

  context AssociationEnd::conformsTo(ae:AssociationEnd):
    post:
      result = (self.multiplicity=ae.multiplicity and
                self.navigable=ae.navigable and
		self.composite=ae.composite and
		self.type.conformsTo(ae.type) and
		(not ae.readonly implies ae.type.conformsTo(self.type)))

In other words, an association end conforms to another end "ae" if it
has equal navigability, composition property and multiplicity, and its
type conforms to ae's type, and if ae can be used to manipulate the
association (ae is not readonly, currently meaning that ae's
association is not readonly) then ae's type also conforms to this
end's type (bidirectional conformance).

An AssociationNavigationExpression works polymorphically, if and only
if the end it navigates to is exposed by the static (compile-time)
type of the object on which the navigation is performed
(ane.object.type) and the actual type of this object exposes a
conforming association end. In all other cases, the association
navigated by the AssociationNavigationExpression is the one that is
determined at compile time.