Thread overview
A Modest Proposal: Final class instances
May 05, 2009
dsimcha
May 05, 2009
Jason House
May 05, 2009
Walter Bright
May 05, 2009
dsimcha
May 05, 2009
Walter Bright
May 05, 2009
div0
May 05, 2009
BCS
May 06, 2009
Stewart Gordon
May 05, 2009
Several people have griped in the past that D class methods are virtual by default.  I've pointed out to them that you can get around this by making the methods final.  However, this is a bit of a blunt instrument, because some use cases for a single class may call for polymorphism and other use cases for the same class may call for fast performance and no polymorphism.  A possible solution is, given a class:

class Foo {
    // Actual implementation.
}

final class FooFinal : Foo{
    // Dummy that just makes Foo final.
}

And then, when you need performance and not polymorphism, you would invoke FooFinal instead of Foo.  However, this requires manual forwarding of constructors and bloats the name space.  A simple syntactic sugar solution to this dilemma that would add very little complexity to the language would be to allow final to describe a class instance, as well as a class.  The following would apply to a final instance:

1.  Method calls don't need to be virtual.
2.  An instance of a subclass cannot be converted to a final instance of the
base class.
3.  A final instance can be implicitly converted to a non-final instance, but
the opposite would not work.

Using final as an instance attribute like this would also allow another useful
feature:  Storing class instances inline in arrays, structs, or other classes.
 Basically, by marking a class instance as final, you'd be telling the
compiler that you do not need and are not using polymorphism in this case,
even if the class hierarchy uses it for other use cases, and therefore, all
relevant optimizations can be made.
May 05, 2009
IMHO, this proposal does not go far enough. THose that worry about performance also care about object size. If a vtable or monitor are needed, then people will still avoid classes for performance.

dsimcha Wrote:

> Several people have griped in the past that D class methods are virtual by default.  I've pointed out to them that you can get around this by making the methods final.  However, this is a bit of a blunt instrument, because some use cases for a single class may call for polymorphism and other use cases for the same class may call for fast performance and no polymorphism.  A possible solution is, given a class:
> 
> class Foo {
>     // Actual implementation.
> }
> 
> final class FooFinal : Foo{
>     // Dummy that just makes Foo final.
> }
> 
> And then, when you need performance and not polymorphism, you would invoke FooFinal instead of Foo.  However, this requires manual forwarding of constructors and bloats the name space.  A simple syntactic sugar solution to this dilemma that would add very little complexity to the language would be to allow final to describe a class instance, as well as a class.  The following would apply to a final instance:
> 
> 1.  Method calls don't need to be virtual.
> 2.  An instance of a subclass cannot be converted to a final instance of the
> base class.
> 3.  A final instance can be implicitly converted to a non-final instance, but
> the opposite would not work.
> 
> Using final as an instance attribute like this would also allow another useful
> feature:  Storing class instances inline in arrays, structs, or other classes.
>  Basically, by marking a class instance as final, you'd be telling the
> compiler that you do not need and are not using polymorphism in this case,
> even if the class hierarchy uses it for other use cases, and therefore, all
> relevant optimizations can be made.

May 05, 2009
dsimcha wrote:
> A possible
> solution is, given a class:
> 
> class Foo {
>     // Actual implementation.
> }
> 
> final class FooFinal : Foo{
>     // Dummy that just makes Foo final.
> }

With dmd:

  final class A { }
  class B : A { }

Compiling:

  test.d(2): Error: class test.B cannot inherit from final class A
May 05, 2009
== Quote from Walter Bright (newshound1@digitalmars.com)'s article
> dsimcha wrote:
> > A possible
> > solution is, given a class:
> >
> > class Foo {
> >     // Actual implementation.
> > }
> >
> > final class FooFinal : Foo{
> >     // Dummy that just makes Foo final.
> > }
> With dmd:
>    final class A { }
>    class B : A { }
> Compiling:
>    test.d(2): Error: class test.B cannot inherit from final class A

Uh...I think you misread this.  I was inheriting from a non-final class to make it final, not the other way around.
May 05, 2009
Reply to dsimcha,

> Several people have griped in the past that D class methods are
> virtual by default.  I've pointed out to them that you can get around
> this by making the methods final.  However, this is a bit of a blunt
> instrument, because some use cases for a single class may call for
> polymorphism and other use cases for the same class may call for fast
> performance and no polymorphism.  A possible solution is, given a
> class:
> 
> class Foo {
> // Actual implementation.
> }
> final class FooFinal : Foo{
> // Dummy that just makes Foo final.
> }
> And then, when you need performance and not polymorphism, you would
> invoke FooFinal instead of Foo.  However, this requires manual
> forwarding of constructors and bloats the name space.  A simple
> syntactic sugar solution to this dilemma that would add very little
> complexity to the language would be to allow final to describe a class
> instance, as well as a class.  The following would apply to a final
> instance:
> 
> 1.  Method calls don't need to be virtual.
> 2.  An instance of a subclass cannot be converted to a final instance
> of the
> base class.
> 3.  A final instance can be implicitly converted to a non-final
> instance, but
> the opposite would not work.
> Using final as an instance attribute like this would also allow
> another useful
> feature:  Storing class instances inline in arrays, structs, or other
> classes.
> Basically, by marking a class instance as final, you'd be telling the
> compiler that you do not need and are not using polymorphism in this
> case,
> even if the class hierarchy uses it for other use cases, and
> therefore, all relevant optimizations can be made.
> 

You can't make an instance final without creating a new type (code bloat) because even if you can generate non virtual calls into the object, interior calls will still be virtual. One solution to this would be to rerun the code gen for the classes methods (and all base class methods) without virtual calls.

OTOH you could do that for even non-final types: generate a set of methods with (where possible) non-virtual internal calls (that could require re-running the code gen on base class methods) that is used only for this class and a set with virtual calls for types that derive from it. This would result in a lot more code but it could (depending on usage) be a lot faster code.


May 05, 2009
dsimcha wrote:
> == Quote from Walter Bright (newshound1@digitalmars.com)'s article
>> dsimcha wrote:
>>> A possible
>>> solution is, given a class:
>>>
>>> class Foo {
>>>     // Actual implementation.
>>> }
>>>
>>> final class FooFinal : Foo{
>>>     // Dummy that just makes Foo final.
>>> }
>> With dmd:
>>    final class A { }
>>    class B : A { }
>> Compiling:
>>    test.d(2): Error: class test.B cannot inherit from final class A
> 
> Uh...I think you misread this.  I was inheriting from a non-final class to make it
> final, not the other way around.


Since all classes inherit from Object, this does indeed do what you asked for.

Try it. You'll find it "finalizes" the class. I'd also suggest trying out some calls to virtual functions in a final class, and running the result through obj2asm to see what happens.
May 05, 2009
Walter Bright wrote:
> dsimcha wrote:
>> == Quote from Walter Bright (newshound1@digitalmars.com)'s article
>>> dsimcha wrote:
>>>> A possible
>>>> solution is, given a class:
>>>>
>>>> class Foo {
>>>>     // Actual implementation.
>>>> }
>>>>
>>>> final class FooFinal : Foo{
>>>>     // Dummy that just makes Foo final.
>>>> }
>>> With dmd:
>>>    final class A { }
>>>    class B : A { }
>>> Compiling:
>>>    test.d(2): Error: class test.B cannot inherit from final class A
>>
>> Uh...I think you misread this.  I was inheriting from a non-final
>> class to make it
>> final, not the other way around.
> 
> 
> Since all classes inherit from Object, this does indeed do what you asked for.
> 
> Try it. You'll find it "finalizes" the class. I'd also suggest trying out some calls to virtual functions in a final class, and running the result through obj2asm to see what happens.

Hey thats great.

Now all we need is the 'virtual' keyword to go with some the methods of the final class... :O
May 06, 2009
dsimcha wrote:
<snip>
> 1.  Method calls don't need to be virtual.
> 2.  An instance of a subclass cannot be converted to a final instance of the
> base class.
> 3.  A final instance can be implicitly converted to a non-final instance, but
> the opposite would not work.
> 
> Using final as an instance attribute like this would also allow another useful
> feature:  Storing class instances inline in arrays, structs, or other classes.
>  Basically, by marking a class instance as final, you'd be telling the
> compiler that you do not need and are not using polymorphism in this case,
> even if the class hierarchy uses it for other use cases, and therefore, all
> relevant optimizations can be made.

So the class would be contained by value, and method calls would bypass the vtable.  Effectively, the class would become more like a struct in these instances.

There are two difficulties I can see:

- either the final instance would still need a vtable to support methods calling each other, or there'd have to be internally two versions of each method
- containment by value means that one would have to deal with copying of instances, in ways that you wouldn't have to worry about if you were using the same class normally.

Stewart.