Home > Articles > Programming > Windows Programming

Using Object Oriented State Machines in .NET

  • Print
  • + Share This
Maybe you haven't worked with state machines since your college computer science courses. Jon Shemitz offers a reason to dust off the technique with .NET: object-oriented state machines can be easier to read and debug than their
Like this article? We recommend

State machines are old technology, but they have many uses even for modern programmers. Probably the most common reason to build a state machine on .NET is emulating coroutines in recursive IEnumerator, such as one that must return a FileSystemInfo for every file and directory under a starting directory. Since this only takes three states, that's exactly what the sample code for this article ( StateMachines.zip ) can do. In Whidbey, iterators will eliminate the need to emulate coroutines, but Whidbey may still be as much as a year out—and even when Whidbey arrives, there will still be uses for state machines in .NET programming.

The basic idea is that you execute a state machine iteratively, when you need to extract the next value from an input stream, or when a new input comes in. The Current state controls which code is executed on each iteration; the Current state's handler may or may not change the Current state. If the Current state handler doesn't change the Current state, the same handler is invoked each time the state machine is iterated, until something changes.

The canonical way to implement a state machine is via an enum containing the name of every state, and a giant switch statement containing the handler for each enum. This compiles to a nice efficient jump table, but it's hard to maintain. When you add a new state, you have to change both the enum and the switch statement; plus, tracing execution means lots of jumping around the switch statement.

After writing an IEnumerator as an enum and switch state machine, I got to wondering: if an object is just a state packet, could you swap state by changing the CurrentState object, instead of the CurrentEnum value? Each state's handler would stand by itself as a standard method, which is easier to read than a long switch statement. Adding a new state just means adding a new class, which is easier than editing both the enum and the switch. And when a handler changes the CurrentState, you can use F12 to jump to the implementation of the next state, which is easier than having to find it by hand in a long switch statement.

The interface between IEnumerator and a state machine is pretty obvious: Reset() sets CurrentState to an initial state. MoveNext() iterates the state machine, either setting an internal CurrentValue field and returning true, or returning false to indicate that there are no (more) items. And the Current property exposes the internal CurrentValue field.

Each state handler needs to be able to change the enumerator's private CurrentState field. That is, we need a circular reference where the state machine refers to the current state object, and the state handler has a reference back to the state machine.

A simple implementation gives the Enumerator object a private reference to a CurrentState object; each CurrentState object is an instance of a private class, nested within the Enumerator class, that has a reference to the Enumerator instance so that it can change the CurrentState.

The only problem, here, is that each different enumeration needs to build the state machine infrastructure all over again: reusability demands looser coupling.

Similarly, you could implement the current state handler as a virtual method, but this requires that all state classes descend from a common, abstract ancestor. While this is probably fine in at least 99.9% of all cases, using interfaces presents a bit more flexibility.

Thus, StateMachines.cs includes an abstract Enumerator class which expects state handler classes to support IStateHandler:

public interface IStateHandler 
{
	bool MoveNext(IStateMachine StateMachine) ;
}

The abstract Enumerator class supports IEnumerator via a generic state machine:

private IStateHandler  CurrentState;
private object CurrentValue ;

protected abstract IStateHandler  StartState() ;

public void Reset()
{
  CurrentState = /*abstract*/ StartState();
  CurrentValue  = null;
}

public object Current
{  
  get  {  return CurrentValue ;  }  
}

public bool MoveNext()
{
  return CurrentState.MoveNext(this);
}

Reset() sets an initial state handler, and MoveNext() iterates the state machine by calling the current state's handler. Now, sometimes you have to cycle the state machine through a few iterations before you get to an output state. For example, a given directory may have subdirectories but no files, or you may have just processed the last subdirectory of a last subdirectory, and so on. In an enum and switch implementation, this might be done with code like this:

bool Deliver = false;
do
  switch (CurrentState)
  {
  }
while (! Deliver);

This loops until a state handler sets Deliver to true.

In my loosely-coupled, object-oriented implementation, the abstract Enumerator class supports the IStateMachine interface, which provides state classes with a couple of utilities: bool Yield() sets the internal CurrentValue (and, optionally, the CurrentState) and returns true; bool Recurse() sets the CurrentState, and recursively returns whatever the new state's IMoveNext implementation returns:

bool IStateMachine.Yield(object NewValue)
{
  CurrentValue  = NewValue;
  return true;
}

bool IStateMachine.Yield(object NewValue, IStateHandler  NextState)
{
  CurrentValue  = NewValue;
  CurrentState = NextState;
  return true;
}

bool IStateMachine.Recurse(IStateHandler NextState)
{
  if (NextState == null)
    return false; // no next state, no next item
  CurrentState = NextState;
  return NextState.MoveNext(this); // recurse
}

That is, return Yield() sets a new CurrentValue and suspends the state machine; return Recurse() passes control to a new state handler (which may itself either Yield() or Recurse()) and passes the result back out to the "outer process."

The actual state classes are nested two deep in the AllFiles class, which implements IEnumerable by returning a new instance of the nested private class DirectoryEnumerator: Enumerator.DirectoryEnumerator overrides StartState() and contains the actual state handler classes, ShowDirectory, ShowFiles, and ShowSubdirectories.

The start state is ShowDirectory, with a null next state, and the DirectoryInfo passed to the AllFiles constructor:

protected override IStateHandler  StartState()
{
  return new ShowDirectory(null, RootDirectory);
}

Thus, calling Enumerator.MoveNext() after a Reset() calls the ShowDirectory state handler:

public bool MoveNext(IStateMachine StateMachine)
{
  IStateHandler  subdirectories = 
    new ShowSubdirectories(NextState, Directory.GetDirectories());
  IStateHandler  files = 
    new ShowFiles(subdirectories, Directory.GetFiles());


  // set Current to this directory; next time, scan files
  return StateMachine.Yield(Directory, files);
}

This first creates a ShowSubdirectories state that "returns" to the ShowDirectory state's NextState. Then, it creates a ShowFiles state that "forwards" to the ShowSubdirectories state. Finally, it sets the CurrentValue to Directory and returns true.

The ShowFiles constructor calls GetEnumerator() on its second, IEnumerable parameter, and ShowFiles.MoveNext() either steps this array enumerator or passes control to the next, ShowSubdirectories state:

public bool MoveNext(IStateMachine StateMachine)
{
  if (ArrayEnumerator.MoveNext())
    return StateMachine.Yield((FileInfo) ArrayEnumerator.Current);
  else
    return StateMachine.Recurse(NextState);
}

The ShowSubdirectories constructor similarly calls GetEnumerator() on its second, IEnumerable parameter, and ShowSubdirectories.MoveNext() either steps this array enumerator and recursively sets the NewDirectory state to iterate the subdirectory, or "returns" to the parent directory state, which may be null:

public bool MoveNext(IStateMachine StateMachine)
{
  if (ArrayEnumerator.MoveNext())
    // at least one more subdirectory - recurse
    return StateMachine.Recurse(
      new ShowDirectory(this, 
        (DirectoryInfo) ArrayEnumerator.Current));
  else
    // no (more) subdirectories - pop
    return StateMachine.Recurse(NextState);
}

As you can see, ShowSubdirectories is a strictly internal state that never yields control to the "outer process." It either invokes the ShowDirectory state with itself as the parent state, or it returns control to its own parent.

State machines often require more state data than simply which handler to invoke next. For example, the whole point of IEnumerator is the CurrentValue. In a one-off state machine, state handlers might be nested within the class that implements IEnumerator, and can thus freely access private fields and private properties. In a more loosely-coupled implementation, you can do as I've done, and build access to the extra state variables into the handlers' interface to the state machine.

One thing you don't need to do is to build an explicit state stack, as you often need to do in an an enum and switch implementation. Each state class contains a NextState field, which can be a "return address" (as in the ShowDirectory and ShowSubdirectories states) or a "forwarding address" (as in the ShowFiles state). In effect, the state classes maintain a sort of distributed stack, in the linked list of ShowDirectory state objects.

Overall, I find that object-oriented state machines are a lot easier to read and debug than their enum and switch equivalents. While there's a certain amount of programming overhead involved in creating a class for each state, this seems to me to be roughly comparable to the amount of overhead involved in updating both an enum and a switch.

While implementing a state machine via state objects is a bit more expensive than using a switch statement, I don't think this matters all that much since most uses of state machines focus on their ability to maintain context across multiple invocations, and speed is not a primary concern.

Jon Shemitz is an independent developer and author in Santa Cruz, California.

Download the code from this article.

  • + Share This
  • 🔖 Save To Your Account