[Bug 65] New: Strange results overriding interface return with class return

[d-bugmail at puremagic.com]

http://d.puremagic.com/bugzilla/show_bug.cgi?id=65

           Summary: Strange results overriding interface return with class
                    return
           Product: D
           Version: 0.150
          Platform: PC
               URL: http://www.digitalmars.com/drn-
                    bin/wwwnews?digitalmars.D.bugs/3287
        OS/Version: Windows
            Status: NEW
          Keywords: wrong-code
          Severity: critical
          Priority: P2
         Component: DMD
        AssignedTo: bugzilla at digitalmars.com
        ReportedBy: smjg at iname.com


The compiler allows a method with an interface return type to be overridden
with a class return type.  However, when this is done, strange things happen,
from AVs to doing things that seem to have no relation to the method that was
called.

Two similar testcases:

----- covariant_int2.d -----
import std.stdio;

interface Father {}

class Mother {
     Father test() {
         writefln("Called Mother.test!");
         return new Child(42);
     }
}

class Child : Mother, Father {
     int data;

     this(int d) { data = d; }

     override Child test() {
         writefln("Called Child.test!");
         return new Child(69);
     }
}

void main() {
     Child aChild = new Child(105);
     Mother childsMum = aChild;
     Child childsChild = aChild.test();
     Child mumsChild = cast(Child) childsMum.test();
}

----- covariant_int4.d -----
import std.stdio;

interface Father {
     void showData();
}

class Mother {
     Father test() {
         writefln("Called Mother.test!");
         return new Child(42);
     }
}

class Child : Mother, Father {
     int data;

     this(int d) { data = d; }

     override Child test() {
         writefln("Called Child.test!");
         return new Child(69);
     }

     void showData() {
         writefln(data);
     }
}

void main() {
     Child aChild = new Child(105);
     Mother childsMum = aChild;

     aChild.test();
     Father mumTest = childsMum.test();
     aChild.showData();
     mumTest.showData();
}

----------
D:\My Documents\Programming\D\Tests\bugs>covariant_int2
Called Child.test!
Called Child.test!
Error: Access Violation

D:\My Documents\Programming\D\Tests\bugs>covariant_int4
Called Child.test!
Called Child.test!
105
Child
----------

I'm guessing that the underlying cause of both is the same - as speculated
before

http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D.bugs/2070

interface references aren't compatible with class references.  This means that
when a method is covariantly overridden from interface to class, and it is then
called through the base class, a class reference is returned, which is no good
as the base class method, and hence the caller through it, needs an interface
reference.

The spec doesn't explicitly forbid this, but if it isn't supposed to work then
the compiler should be giving an error (and the spec updated accordingly). 
Otherwise, it could be fixed to work like this:

- The compiler would detect that a method is being overridden from Father
(interface) to Child (class), and compile Child.test to return an interface
reference for compatibility.

- When the method is called through a Child reference, the caller would need to
implicitly convert the returned Father reference to a Child reference.  Of
course, this conversion can be optimised away if the context dictates that a
Father reference is required.

- It would be necessary to throw in a restriction or two.  A class cannot
derive from both a class and an interface, or multiple interfaces, if they
define methods with the same name and parameter types but one has a class
return and the other has an interface return.  Assuming that it would be
impossible to compile the method to be compatible with both simultaneously.

I haven't experimented with interface-to-interface covariant overrides, so
don't know if these work.  But I can imagine there being complications when
multiple interface inheritance is involved.

The question: Is it worth making this work?  Or do these complications mean
that we ought to disallow interface-to-class overrides altogether?


--