Bug 2524 – final override inconsistent when implementing interfaces

The following code builds: interface I { void foo(); } class C : I { override void foo(); } But the following does not work: class C : I { final override void foo(); } testfinaloverride.d(8): function testfinaloverride.C.foo does not override any function The two cases should be consistent. The spec is unclear to me as to whether implementing an interface function qualifies as overriding a function or not. I would vote for the case where override requires either to override a base class function or implement an interface function, since it is a pain if you change a base class to an interface, you'd have to remove all the override keywords. So I'm marking it as rejects-valid. The spec should also specifically lay out what override does in the case of interfaces.

It compiles in D2, though silent interface method implementation looks strange for me.

(In reply to comment #1) > It compiles in D2, though silent interface method implementation looks strange > for me. I've tested this with DMD 2.022 and 2.021 and I get this error in both cases.

*** Bug 2593 has been marked as a duplicate of this bug. ***

My report is also a duplicate of this bug. Note that Walter closed it with the following comment: «A "final private" method is not virtual, and hence won't work for an interface method. That's why the error message appears. You can make it an enhancement request if you like.»

*** Bug 2538 has been marked as a duplicate of this bug. ***

I don't think #2538 is a duplicate of this. It's the "private" part of "final private" that makes the method non-virtual. Implementing interface methods with final methods is absolutely legal.

In fact, private implies final (non-virtual), and does not put the function in the vtable (meaning it cannot override a base function). From the spec: "All non-static non-private non-template member functions are virtual." If someone casts your class to an interface, do you want them to now be able to call your private function?

(In reply to comment #7) > In fact, private implies final (non-virtual), and does not put the function in > the vtable (meaning it cannot override a base function). From the spec: "All > non-static non-private non-template member functions are virtual." The sentence you've quoted states nothing about functions that _are_ private. Moreover, a final method _can_ override a method in a base class. Even so, an interface has its own vtable, so the virtuality or not of a function with respect to its class (or a class from which it is derived) should have nothing to do with whether it can implement an interface method. > If someone casts your class to an interface, do you want them to > now be able to call your private function? Probably not. But the way it's attempted in issue 2538 brings us back to the problem of inheritance protection. But we can talk about this there.

(In reply to comment #7) > In fact, private implies final (non-virtual), and does not put the function in > the vtable (meaning it cannot override a base function). Protection should be orthogonal to "virtuality" attributes. Here is an example: class A { private void foo() { writefln("A.foo"); } } class B : A { private override void foo() { super.foo(); writefln("B.foo"); } } > From the spec: "All non-static non-private non-template member functions are virtual." I believe this is a design flow. Same as "package implies final". > If someone casts your class to an interface, do you want them to now > be able to call your private function? > Of course, why would you inherit or override it other than to allow virtual behavior? Here is an example: interface INetworkListener { void acceptNetworkPacket(); } class NetworkManager { static void registerNetworkListener(INetworkListener listener) { ... } } class SoundManager : private INetworkListener { this() { NetworkManager.registerNetworkListener(this); } private void acceptNetworkPacket() { // ... } } No-one should know that SoundManager implements INetworkListener interface. Since "acceptNetworkPacket" method is an implementation detail, I don't want it to be visible from the outside of this class (to prevent accidential invokation). Thus I mark it private. Note that private interface inheritance is harmless as it doesn't shadow Object's methods (opCmp, opEquals, toString etc).

But I can still cast to the interface. Protection attributes are compile-time entities, they are not flagged at runtime: SoundManager s; Object o = s; INetworkListener inl = cast(INetworkListener)o; The compiler just looks in the object's list of interfaces to see if it finds that interface. There is no protection associated with it at runtime. What you are asking for requires a major compiler redesign for limited value. You can easily implement what you want using a private inner class: interface INetworkListener { void acceptNetworkPacket(); } class NetworkManager { static void registerNetworkListener(INetworkListener listener) { ... } } class SoundManager { private class NetworkListener : INetworkListener { void acceptNetworkPacket() { // ... } } this() { NetworkManager.registerNetworkListener(new NetworkListener); } } This is similar to how Java works with anonymous classes. D can do the same thing I think, but I can't remember the syntax right now.

(In reply to comment #10) > But I can still cast to the interface. Protection attributes are compile-time > entities, they are not flagged at runtime: > > SoundManager s; > Object o = s; > > INetworkListener inl = cast(INetworkListener)o; > > The compiler just looks in the object's list of interfaces to see if it finds > that interface. There is no protection associated with it at runtime. > So what? How does this differ from the following in C++: class A { void foo(); } class B : protected A { } B* b = new B(); //b->foo(); // error //A* a = b; // error void* o = b; A* a = (A*)o; a->foo(); You just hijacked type system, that's it. One possible solution would be to add protection flag into class typeinfo which will be taken into account during dynamic cast. Or remove interface from classinfo.interfaces altogether (it should still be possible to cast known class to its known base class/interface statically, i.e. without dynamic cast). I know the issue is not of high value but it should either be fixed for consistency, or protection inheritance attributes should be removed from language.

(In reply to comment #10) > But I can still cast to the interface. Protection attributes are compile-time > entities, they are not flagged at runtime: > > SoundManager s; > Object o = s; > > INetworkListener inl = cast(INetworkListener)o; > > The compiler just looks in the object's list of interfaces to see if it finds > that interface. There is no protection associated with it at runtime. But the compiler is capable of checking to see whether the inheritance is public or private. It just doesn't at the moment. No runtime protection checking needed in this instance. > What you are asking for requires a major compiler redesign for limited value. I still think this "major compiler redesign" should be removing inheritance protection altogether. Which would have the value of simplifying the language by getting rid of these things that don't make sense. (In reply to comment #11) > So what? How does this differ from the following in C++: > > class A { > void foo(); > } > > class B : protected A { > } > > B* b = new B(); > //b->foo(); // error > //A* a = b; // error > void* o = b; > A* a = (A*)o; > a->foo(); Firstly, try putting your code through a C++ compiler. Secondly, your code goes through void*, whereby it's obvious that you're intent on subverting the type system. Moreover, it may break if multiple inheritance is involved. That said, as I try it C++ allows the cast straight from B* to A*, but it must be explicit.

(In reply to comment #12) > (In reply to comment #10) > > But I can still cast to the interface. Protection attributes are compile-time > > entities, they are not flagged at runtime: > > > > SoundManager s; > > Object o = s; > > > > INetworkListener inl = cast(INetworkListener)o; > > > > The compiler just looks in the object's list of interfaces to see if it finds > > that interface. There is no protection associated with it at runtime. > > But the compiler is capable of checking to see whether the inheritance is > public or private. It just doesn't at the moment. No runtime protection > checking needed in this instance. > Casting to Object is valid, since the class did not inherit from Object privately. Once you get to object, you need runtime information to cast to a subclass or interface. If you're suggesting that an optimization can "see" that the object originated from a SoundManager instance, then I can easily generate a case where the cast to object and the cast to the interface are hidden in opaque functions, eliminating the possibility of optimization. I still contend that this cannot be checked at compile-time. (In reply to comment #11) > So what? How does this differ from the following in C++: > > [snip] > > You just hijacked type system, that's it. No, I did not. Casting to an Object is a legal move within the type system. It is equivalent to casting to a base class. Casting to an interface is also a legal move, it is like a C++ dynamic_cast to a derived class. There is no hijacking going on. > > One possible solution would be to add protection flag into class typeinfo which > will be taken into account during dynamic cast. Or remove interface from > classinfo.interfaces altogether (it should still be possible to cast known > class to its known base class/interface statically, i.e. without dynamic cast). I'm not sure this is possible, but I don't know completely the inner workings of an interface cast. I think it's required that the interface vtable be present in the class instance, and I think it's probably also required that it be in the classinfo. > I know the issue is not of high value but it should either be fixed for > consistency, or protection inheritance attributes should be removed from > language. I would agree wholeheartedly that inheritance protection should be disallowed for interfaces. I am not sure about inheritance protection for the base class, I've never used it, and the documentation is completely absent (save for the mention that it's valid syntax). I can't see a use case that couldn't be solved by composition, or using private inner classes.

(In reply to comment #13) >> But the compiler is capable of checking to see whether the inheritance is >> public or private. It just doesn't at the moment. No runtime protection >> checking needed in this instance. > > Casting to Object is valid, since the class did not inherit from Object > privately. Once you get to object, you need runtime information to cast to a > subclass or interface. Sorry, I misread what your code was doing. It's the same as the C++ example indeed.

This is a compiler bug. You can override an interface function.

Commit pushed to https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/f595bde35576dd0165dd0e0964d478c7db31955a fix Issue 2524 - final override inconsistent when implementing interfaces