First-Class Slots for .NET

Yesterday, I posted about the new extension of IRef<T> to arbitrary reference semantics for the CLR, including referencing inner fields, properties, and array slots. First-class references make it simple to operate on specific mutable data without caring about the underlying type of that data.

I just pushed another abstraction that handles a related, but different case: first-class slots.

ISlot<T>

An object "slot" is a value that designates a mutable location of a specific class of values, not a mutable location of a specific instance like first-class references. Where first-class references hide the underlying object type, slots expose the object type and allow you to mutate the slots of multiple objects at once, as long as they are subtypes of the slot's object type. Here's the declaration of ISlot<T>:

public interface ISlot<TObj, T>
{
    T Get(TObj obj);
    T Set(TObj obj, T value);
}

As you can see, the object we're manipulating is passed in while operating on it, so an ISlot is actually a direct descriptor that can access members of any instance of a subtype of TObj. If you're wondering when you would ever need this, rest assured that there are applications that need to write algorithms of this type. For instance, consider an object-relational mapper (ORM), which uses reflection to extract the members that need to be get/set on object flush/load from a database. Essentially, the ORM is reflecting over all slots of an object being flushed or loaded from a database, but it does so in a manner that isn't very reusable, and the object hydration code is coupled tightly to the database access code as a result.

Reifying slots as a distinct, first-class abstraction makes them independently testable, and the reflection code and database access code is now very loosely coupled. An ORM is but one example of an application that makes use of generic object slots.

Similar to first-class references, the Slot class exposes some static constructors for creating slots:

public abstract class Slot
{
  public static Member<TObj, T> Create<TObj, T>(Func<TObj, T> get, Action<TObj, T> set);
  public static Array<T> Create<T>(int index);
  public static Member<TObj, T> Create<TObj, T>(MemberInfo member)
  public static Member<TObj, T> Create<TObj, T>(Expression<Func<TObj, T>> member)

  public sealed class Array<T> : Slot, ISlot<T[], T> { ... }
  public sealed class Member<TObj, T> : Slot, ISlot<TObj, T> { ... }
}

These operations are exactly what you saw in the last article, where you can create slots from object members and array indices.

It might seem at first that slots are strict generalizations of first-class references, but this is deceptive. It's true that any algorithm you'd could write using references could be rewritten to use slots, but the number of type parameters and value parameters could increase non-linearly to the point where it's unwieldy and could easily obfuscate the underlying algorithm.

Limitations

There is one limitation at the moment. Specifically, value types need to be passed by reference during update in order for assignment to be visible to callers of ISlot.Set, but this isn't currently possible given the interface. As a result, there's currently a type constraint on TObj to restrict it to reference types.

A simple solution would simply be to return TObj from ISlot.Set, so the calling context can simply overwrite its own local value with the one modified by the slot. Another possibility is to make the TObj parameter to ISlot.Set a by-ref parameter. I'm considering these and a few other options, and Sasa's v0.12.0 release will probably contain the final solution.