A2Z.Net : A – Anonymous Types

This article marks the beginning of a new series of posts, which aims to dig deeper into 26 different features of .Net. For the first article, or the A – I have selected Anonymous Types. I would have chosen asynchronous programming, but I have whole dedicated series on asynchronous programming which could followed in this link.

Introduced in .Net 3.x, Anonymous Types has made the life easy for the developers by handling out the ability to work on an encapsulated type, without having to actually define one. I should say encapsulate is a wrong word here (but have used it so that I could highlight why) because you cannot place any behavior with the properties.

Anonymous Types and Syntax

Anonymous Type provides an easy way to encapsulate a set of properties without having to explicitly define a type. Notice that I used the word explicitly here – that is because the compiler creates the type for you under the hood. We will look into what happens behind the scenes a little later, to begin with, let us introduce the syntax of defining an anonymous type.

var user = new
{
    UserName ="Anu Viswan",
    Age = userRecord.Age,
    JoiningDate= DateTime.Now
};

The above code uses object initializer (without having to specify the data type) to create an instance of an anonymous type initialized with 3 properties – namely UserName(string),Age(int), and JoiningDate (DateTime). You could also create an anonymous type.

As I mentioned earlier, the compiler generates a type under the hood and ensure the code is still statically typed.

Since the type of anonymous Type is known statically, they must be stored using variables with var allowing type inference. You could also store them as object, but that would defeat the whole purpose of anonymous types.

In addition to the creation syntax mentioned above, framework also provides another way initializing an anonymous type. This syntax is known as projection initializer, and it allows you to use the property/field names from somewhere else.

var person = new Person
{
    FirstName = "Anu",
    LastName = "Viswan",
    Age = 36
};

var user = new
{
    person.FirstName,
    person.Age,
};

The above code uses a part of type Person to create an anonymous type. In this case, as you can notice, you are not explictly providing names for properties, instead, they uses the same name as the Person. This is of course equavalent of writing as

var user = new
{
    FirstName = person.FirstName,
    Age = person.Age
};

The projection initializer syntax is highly useful in Linq queries when you need to project a subset of a query or combine properties from multiple objects. For Example,

var personList = new List<Person>
{
	new Person {UserName="Anu Viswan",    Age = 36},
	// Other Persons
};

var employeeList = new List<Employee>
{
    new Employee{UserName="Anu Viswan",JoiningDate=new DateTime(2020,2,2)},
    // Other employee records //
};

var user = employeeList.Join(personList,
							e=>e.UserName,
							p=>p.UserName,
							(e,p)=> new
								{
									e.UserName,
									UserDetail = new  // An example of Nested Anonymous Type
									{
									e.JoiningDate,
									p.Age,
									QueryTimeStamp=DateTime.Now
									}
								});

As you can observe the creation of anonymous types are pretty simple. You could also create a type combining multiple objects as seen in above example. The above code also demonstrates that the anonymous types could be nested. UserDetail is an anonymous type within the user.

Behind the Scenes

Alright, enough of syntax, let us now look behind the scenes now. Let us bring back our first example of anonymous type and see how the type generated by the compiler looks like.

var user = new
{
    UserName = "Anu Viswan",
    Age = 36,
    JoiningDate= DateTime.Now
};

The compiler generated (_generated using ILSpy) type would look like

[CompilerGenerated]
[DebuggerDisplay("\\{ UserName = {UserName}, Age = {Age}, JoiningDate = {JoiningDate} }", Type = "<Anonymous Type>")]
internal sealed class <>f__AnonymousType0<<UserName>j__TPar, <Age>j__TPar, <JoiningDate>j__TPar>
{
	[DebuggerBrowsable(DebuggerBrowsableState.Never)]
	private readonly <UserName>j__TPar <UserName>i__Field;

	[DebuggerBrowsable(DebuggerBrowsableState.Never)]
	private readonly <Age>j__TPar <Age>i__Field;

	[DebuggerBrowsable(DebuggerBrowsableState.Never)]
	private readonly <JoiningDate>j__TPar <JoiningDate>i__Field;

	public <UserName>j__TPar UserName => <UserName>i__Field;

	public <Age>j__TPar Age => <Age>i__Field;

	public <JoiningDate>j__TPar JoiningDate => <JoiningDate>i__Field;

	[DebuggerHidden]
	public <>f__AnonymousType0(<UserName>j__TPar UserName, <Age>j__TPar Age, <JoiningDate>j__TPar JoiningDate)
	{
		<UserName>i__Field = UserName;
		<Age>i__Field = Age;
		<JoiningDate>i__Field = JoiningDate;
	}

	[DebuggerHidden]
	public override bool Equals(object value)
	{
		// Hidden : Will be discussed later
	}

	[DebuggerHidden]
	public override int GetHashCode()
	{
		// Hidden : Will be discussed later
	}

	[DebuggerHidden]
	public override string ToString()
	{
		// Hidden : Will be discussed later
	}
}


Characteristics of Generated Code

Let us break down the generated code and examine it in parts.

It is a Generic Class

As you would have already observed, the generated type is a class, name of which is generated by the compiler and cannot be accessed from your source code.

internal sealed class <>f__AnonymousType0<<UserName>j__TPar, <Age>j__TPar, <JoiningDate>j__TPar>
{
    public <>f__AnonymousType0(<UserName>j__TPar UserName, <Age>j__TPar Age, <JoiningDate>j__TPar JoiningDate)
	{
		<UserName>i__Field = UserName;
		<Age>i__Field = Age;
		<JoiningDate>i__Field = JoiningDate;
	}
}

It is also interesting to note that the generated type is a Generic class, with one type parameter for each of the properties. This would mean that if there is a second string, it would still have a different type parameter than the original one. For example,

var user = new
{
    UserName = "Anu Viswan",
    FirstName = "Anu"
};

The above would translate to

internal sealed class <>f__AnonymousType0<<UserName>j__TPar, <FirstName>j__TPar>
{
}

Note that the class is marked internal and sealed. Now both these characterstics aren’t guaranteered as per Jon Skeet’s book C# in Depth. This is actually interesting, as to know when and why the compiler could not guarante a consistent behavior.

The generated type also has a constructor which accepts parameters which are in consistent with all the properties in the type.

Read Only Properties

The other characterstics of the generated type that is quite obvious is that all properties are read-only. One cannot reassign properties of an anonymous type. If the need arise, you would have to recreate the entire object.

var user = new
{
    UserName = "Anu Viswan",
    Age = 36,
    JoiningDate = DateTime.Now,
};

user.Age = 37; // This line would throw error

user = new  // The correct approach would be recreate
{
    UserName = user.UserName,
    Age = 37,
    JoiningDate = user.JoiningDate,
};

Derieved from System.Object & Overrides Equals()/GetHashCode()

The generated class for anonymous types is derieved from System.Object. It cannot be, hence, cast to any other type other than Object. The generated type also overrides the Equals() and GetHashcode() methods.

public override bool Equals(object value)
{
    <>f__AnonymousType0<<UserName>j__TPar, <Age>j__TPar, <JoiningDate>j__TPar> anon = value as <>f__AnonymousType0<<UserName>j__TPar, <Age>j__TPar, <JoiningDate>j__TPar>;

    return anon != null && EqualityComparer<<UserName>j__TPar>.Default.Equals(<UserName>i__Field, anon.<UserName>i__Field) && EqualityComparer<<Age>j__TPar>.Default.Equals(<Age>i__Field, anon.<Age>i__Field) && EqualityComparer<<JoiningDate>j__TPar>.Default.Equals(<JoiningDate>i__Field, anon.<JoiningDate>i__Field);
}

public override int GetHashCode()
{
    return ((459989953 * -1521134295 + EqualityComparer<<UserName>j__TPar>.Default.GetHashCode(<UserName>i__Field)) * -1521134295 + EqualityComparer<<Age>j__TPar>.Default.GetHashCode(<Age>i__Field)) * -1521134295 + EqualityComparer<<JoiningDate>j__TPar>.Default.GetHashCode(<JoiningDate>i__Field);
}

These ensures that two instances of AnonymousTypes are equal only if all the properties are equal. Do note that the exact implementation of GetHashCode() might also vary. For example,

var user = new
{
    UserName = "Anu Viswan",
    Age = 36,
    JoiningDate = new DateTime(2020,10,10),
};

var anotherInstance = new
{
    UserName = "Anu Viswan",
    Age = 36,
    JoiningDate = new DateTime(2020, 10, 10),
};

Console.WriteLine($"user & anotherInstance - Equality :{user.Equals(anotherInstance)}"); // True

Reuse of Anonymous Types

If two anonymoys types are created with same Property Names, in same order and type, then the compiler would treat the two objects as of same type, ie, essentially reuse the generated type.

For example, the preceeding code would reuse the same generated type.

Console.WriteLine($"user & anotherInstance - Type Equality :{user.GetType() == anotherInstance.GetType()}"); // True

This could verified when we check the generated code.

<>f__AnonymousType0<string, int, DateTime> user = new <>f__AnonymousType0<string, int, DateTime>("Anu Viswan", 36, new DateTime(2020, 10, 10));

<>f__AnonymousType0<string, int, DateTime> anotherInstance = new <>f__AnonymousType0<string, int, DateTime>("Anu Viswan", 36, new DateTime(2020, 10, 10));

However, if you change the order of Properties in either of the anonymous types, the compiler would generate another type. Consider the following code.

 var user = new
 {
     UserName = "Anu Viswan",
     Age = 36,
     JoiningDate = new DateTime(2020,10,10),
 };
  var differentInstance = new
 {
     UserName = "Anu Viswan",
     JoiningDate = new DateTime(2020, 10, 10),
     Age = 36,
 };

As you can observe, while the PropertyNames and Types are preserved in the two anonymous types, the order is changed.The generated code would be as following.

<>f__AnonymousType0<string, int, DateTime> user = new <>f__AnonymousType0<string, int, DateTime>("Anu Viswan", 36, new DateTime(2020, 10, 10));

f__AnonymousType1<string, DateTime, int> differentInstance = new <>f__AnonymousType1<string, DateTime, int>("Anu Viswan", new DateTime(2020, 10, 10), 36);

As seen in the generated code, the compiler generates and use different internal types for the two anonymous types.

This holds true only when the two anonymous types are in the same assembly. If the two anonymous types are defined in two different assemblies, compiler would generate two different types for each. We will explore this later in this article.

Points Of Interest

The above section outlined the characterstics of Anonymous Types. There are however, certain points which is worth noting while learning about Anonymous Types.

Pass as Parameter/ Return Type

Since the generated type is not accessible/addressed by the developer in his code, it to an extend limits the transfer an anonymous types from one method to another as you cannot express the type as Parameter. However, you could make use of Generics to pass the anonymous types around. For example,

 public void Print<T>(T data)
 {
     var properties = typeof(T).GetProperties();
     foreach(var property in properties)
     {
         Console.WriteLine($"{property.Name} = { property.GetValue(data)}");
     }
 }

The compiler here treats the generated type as any reference type in the system. This allows us to use generics to pass anonymous types around.

You could also pass the anonymous type as an object or dynamic parameter as well, however, that would take away any benefits associated to anonymous types.

Similiar Anonymous Types in Two Assemblies

In the earlier section we noticed how the compiler would reuse the same generated type if the Property Names, Order and Types are the same. However, there is a caveat. The generated types are localized to the assembly.

If the two types resides in different assembly, then the compiler would generate different Anonymous Types for each of them.

Consider the following scenario.

// Assembly 1
 var user = new
 {
     UserName = "Anu Viswan",
     Age = 36,
     JoiningDate = new DateTime(2020,10,10),
 };
 var bar = new Bar();
 bar.Print(user);


// Assembly 2
public class Bar
{
    public void Print(object data)
    {
        var typeSchema = new { UserName = string.Empty, Age = default(int), JoiningDate = default(DateTime) };
        var castInstance = Cast(typeSchema, data);
        Console.WriteLine(data.GetType() == castInstance.GetType());
    }
    private static T Cast<T>(T typeHolder, Object x) where T : class
    {
        return (T)x;
    }
}

The above casting would raise an InvalidCastException as the two types resides in different assemblies, resulting in compiler generating two different anonymous types.

This would have been a succesful cast if the code resided in the same assembly.

Conclusion

As observed, the implement of Anonymous Types is pretty simple, and so is its usage. The compiler generates a generic class for each unique anonymous type within the assembly. The feature has been extensively used since Linq came along as it allows to create subsets from queries without having to define an explicit type.

The complete code citd in this article is available in my Github.

C# 8 : Nullable Reference Types, Null Forgiving and Null Coalescing Assignment

There are some properties that ought to be learned together to understand the complete picture. In this blog post, we will address two of such features.

Nullable Reference Types

Let us begin by writing some code.

public class Foo
{
    public Bar Bar { get; set; }
}

public class Bar
{
    public int Id { get; set; }
}

// Client Code
var foo = new Foo();
Console.WriteLine(foo.Bar.Id);

Notice anything wrong ? How often have been guilty of attempting to access something that could potentially be null ? In the above case, we haven’t initialized the Foo.Bar property, but in the client code, we are attempting to access the Bar.Id property.

This would throw a NullReferenceException, as you are dereferencing a null value (Bar hasn’t been assigned yet). We should have checked for null in the first place.

But as the code base increases, there are times when we forget to do so. And unfortunately, prior to C# 8, there was no way compiler could warn us to do the null check even if we might have wondered many a times if it could do so.

Nullable Reference Types in C# 8 brings us pretty much the same. It allows us to mark the Reference Types which can/cannot be null, using the same syntax we used to mark Nullable Value Types – using the ? operator.

In C# 8, with Nullable Reference Types enable, if you were to compile the code, the compiler warns you with the following.

Warning	CS8618	Non-nullable property 'Bar' is uninitialized. Consider declaring the property as nullable.

Depending on your business logic, you have couple of ways to proceed from here.

Case 1 – Bar is Nullable

If your business logic requires that the Bar property could be null, then you could mark it explicitly. For example,

public Bar? Bar { get; set; }

The ? operator tells the compiler that the property could be null and it should warn the developer for possible null checks if he hasn’t done so yet. The compiler duely warns you with the following.

Warning	CS8602	Dereference of a possibly null reference.

Case 2 – Bar is non-nullable

If your application logic requires that the Bar property cannot be null, then you could let the compiler know by using the following syntax (basically omitting the ? operator).

public Bar Bar { get; set; }

The above code tells the compiler that the Bar cannot be null. Interestingly, we still get a warning now, let us first examine what it is before understanding why compiler throws the warning.

Warning	CS8618	Non-nullable property 'Bar' is uninitialized. Consider declaring the property as nullable.

We have declared the property as Non-nullable, yet we have left it uninitialized. We could of course remove this warning by initializing the property with a default value (either in the property declaration or in the constructor), but this gives us an oppurtunity to introduce another feature.

Null Forgiving Operator

You could have a situation where, as per the business logic, you know that a particular property can never be null at runtime. However, you are in no position to assign a default value. In such situations, you could use the Null-forgivingoperator to disable the warnings. Null-Forgiving is denoted by using the postfix operator !. For example,

public Bar Bar { get; set; } = null!;

The above code informs the compiler that it would skip the warning about Bar being uninitialized and it would not be null in runtime. Do note the null-forgiving operator has not runtime influence.

Null Coalescing Assignment

C# 8 also introduces the Null Coalescing assignment via the the assignment operator ??=. For demonstration, let us further develop our code and introduce a CreateBarInstance to assign instances of Bar. For example,

public void CreateBarInstance(Bar instance)
{
    Bar = instance;
}

Due to some business reasons, we have requirement that the method should assign Bar instance, only if it is null. Of course, one could do this an if condition, but that would introduce a lot of boiler plate code. We could eliminate these extraneous code using the null coalescing operator.

public void CreateBarInstance(Bar instance)
{
    Bar ??= instance;
}

The line Bar ??= instance ensures that Bar is assigned only when the it is null. Or in other words, the value of right hand operand is assigned to left hand operand only if the left hand operand is null.

Demonstrating the code in action,

var foo = new Foo();
foo.CreateBarInstance(new Bar(2));
Console.WriteLine(foo.Bar.Id);
foo.CreateBarInstance(new Bar(3));
Console.WriteLine(foo.Bar.Id);

Output of the above code

2
2

That’s all for now, we will continue our exploration of C# 8 features.

Asynchronous Code – Behind the Scenes – 004

During this series of deep dive into the asynchronous calls, we have so far looked into

  • [x] General Structure of generated code.
  • [x] Role of Stub/Worker method.
  • [x] Structure of State Machine and role of Fields.
  • [x] Implementation of the SetStateMachine method.
  • [ ] Implementation of the MoveNext method.

It is now time to look at the most important piece of the puzzle – the MoveNext() method.

Before we begin exploring the MoveNext(), let us remind ourself that the method is called when the async method is first invoked and then, each time it is resumed. The Method would be responsible for the following.

  • Ensure the method starts/resumes execution at the right place when it starts for the first time or resumes after a pause.
  • Preserve the state of State Machine when it needs to pause.
  • Schedule a continuation when the awaited expression hasn’t been completed yet.
  • Retrieve values from the awaiter.
  • Propagate the return values or method completion via the Builder.
  • Propagate the exceptions if any via the Builder.

The last 2 points are curious if you were to consider that the MoveNext method has a void return Type. So how does the MoveNextreturn the result or exceptions ? Of course via the Builder instance. It is the role of the Stub method to return the Task to the Caller method.

Without taking any time longer, let us take a peek at the generated code. We will then proceed to split it into parts and find how it works

The Whole Code

private void MoveNext()
{
    int num = <>1__state;
    try
    {
        TaskAwaiter awaiter;
        if (num != 0)
        {
            if (num == 1)
            {
                awaiter = <>u__1;
                <>u__1 = default(TaskAwaiter);
                num = (<>1__state = -1);
                goto IL_00cc;
            }
            awaiter = Task.Delay(delay).GetAwaiter();
            if (!awaiter.IsCompleted)
            {
                num = (<>1__state = 0);
                <>u__1 = awaiter;
                <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
                return;
            }
        }
        else
        {
            awaiter = <>u__1;
            <>u__1 = default(TaskAwaiter);
            num = (<>1__state = -1);
        }
        awaiter.GetResult();
        Console.WriteLine(delay);
        awaiter = Bar().GetAwaiter();
        if (!awaiter.IsCompleted)
        {
            num = (<>1__state = 1);
            <>u__1 = awaiter;
            <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
            return;
        }
        goto IL_00cc;
        IL_00cc:
        awaiter.GetResult();
    }
    catch (Exception exception)
    {
        <>1__state = -2;
        <>t__builder.SetException(exception);
        return;
    }
    <>1__state = -2;
    <>t__builder.SetResult();
}

That does look a bit scary to begin with. But, have no worries. We will break it down and understand it better.

Exception Handling

Now we already know that the associated Task object returned by the async method would contain any exception if any. It also sets the status to faulted. So how does the State Machine help in doing so ? That’s the first part we will explore. Let us have birds-eye view of the MoveNext() method – for time being, we will ignore all code within the try block.

private void MoveNext()
{
    int num = <>1__state;
    try
    {
        // Ignore this code for the moment
    }
    catch (Exception exception)
    {
        <>1__state = -2;
        <>t__builder.SetException(exception);
        return;
    }
    <>1__state = -2;
    <>t__builder.SetResult();
}

As you can observe the entire MoveNext() method has a big try catch wrapping the code within. The interesting part for the moment would be the catch block. If any exceptions occurs in the try block, the MoveNext() method sets the state to -2 to indicate the method has completed (-2 indicates completion, irrespective of success or failure). It then uses the Builder to set the exception using the Builder.SetException method.

Only special exceptions like the ThreadAbortException or the StackOverflowException can cause the MoveNext() method to end with an exception.

High Level Flow of State Machine

At a higher level, one can observe that the MoveNext() method returns if any of the following are true

  • Each time the state machine needs to be pause (for an await statement to complete).
  • Execution reaches the end of the method
  • Exception is thrown, but not caught in the async method.

A High level flow of the State Machine could be summarized as follows.

  1. The Stub Method (Worker Method) initiates the State Machine using the Builder Object (AsyncTaskMethodBuilder).
  2. Jump to the correct place in State Machine based on the State Field.
  3. Execute the State Machine until the code reaches await statement or end of the method (return statement).
  4. Fetch the awaiter.
    • If the awater is completed, go back to the Step 2.
    • If not, attach a continuation to the awaiter.
    • If this is the first awaiter, return the Task.
  5. The Task returned in Step 5, would be returned the caller via the Builder.

The Try Block

The Try blocks starts with a switch/if condition depending on the number of await statements within the method. If it has 3 or more awaits, usually one could notice a switch case, in all other cases, an if statement is used.

Irrespective of the approach, the condition to check resolves around the State of the State Machine. If the state is negative, it indicates the first call to the MoveNext() method. If the value of State is a positive number, then it indicates the State Machine is resuming from a pause.

public int <>1__state;
public AsyncTaskMethodBuilder <>t__builder;
public int delay;
private TaskAwaiter <>u__1;

private void MoveNext()
{
    int num = <>1__state;
    try
    {
        TaskAwaiter awaiter;
        if (num != 0)
        {
            if (num == 1)
            {
                awaiter = <>u__1;
                <>u__1 = default(TaskAwaiter);
                num = (<>1__state = -1);
                goto IL_00cc;
            }
            awaiter = Task.Delay(delay).GetAwaiter();
            if (!awaiter.IsCompleted)
            {
                num = (<>1__state = 0);
                <>u__1 = awaiter;
                <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
                return;
            }
        }
        else
        {
            awaiter = <>u__1;
            <>u__1 = default(TaskAwaiter);
            num = (<>1__state = -1);
        }
        awaiter.GetResult();
        Console.WriteLine(delay);
        awaiter = Bar().GetAwaiter();
        if (!awaiter.IsCompleted)
        {
            num = (<>1__state = 1);
            <>u__1 = awaiter;
            <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
            return;
        }
        goto IL_00cc;
        IL_00cc:
        awaiter.GetResult();
    }
    catch (Exception exception)
    {
        // Not displayed for clarity
    }
    <>1__state = -2;
    <>t__builder.SetResult();
}

One of the first things you notice in the code above is that State is stored in a local variable. I guess this is done for optimization purposes. We could use a dedicated post later for understanding different optimizations techniques used by compiler here, for now let us stick to the task in hand.

As one can observe, when the Method is invoked for the first time, as the state would be -1, the code would proceed and hit the first await statement.

awaiter = Task.Delay(delay).GetAwaiter();
if (!awaiter.IsCompleted)
{
    num = (<>1__state = 0);
    <>u__1 = awaiter;
    <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
    return;
}

It fetches the Awaiter using the GetAwaiter method. If the awaiter is already completed, it would proceed to the next step in the original method. If not, it would set the State to 0 (indicating the first instance where the MoveNext() method had to await – zero based index), store the awaiter in the field and schedules the state machine to proceed to the next action when the specified awaiter completes using the Builder.AwaitUnsafeOnCompleted method.

On resumption after the pause, it moves to else part (remember, the state is having a value 0 now). It restores the awaiter stored in the field and clears the fields so that GC could take care of it. It also sets the State to -1.

else
{
    awaiter = <>u__1;
    <>u__1 = default(TaskAwaiter);
    num = (<>1__state = -1);
}
awaiter.GetResult();
Console.WriteLine(delay);
awaiter = Bar().GetAwaiter();
if (!awaiter.IsCompleted)
{
    num = (<>1__state = 1);
    <>u__1 = awaiter;
    <>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
    return;
}

It then proceeds to fetch the Result using the TaskAwaiter.GetResult() method and then executes the remaining steps untill it hits the next await or the method completes. On completion (either finished or faulted), it sets the State to -2 and sets the Result using the Builder.

<>1__state = -2;
<>t__builder.SetResult();

Over the last few posts, we have traced through the generated source code behind the asynchronous methods. We noticed how the method gets translated to a pair of Stub/Working method and a State Machine. We also explored the State Machine in detail and understood how the MoveNext method method navigates the original method while maitaining the states.

The whole process, starting from the moment your code hits the await expression could be summarized as,

  1. Get the awaiter from the awaitable expression using the GetAwaiter() method.
  2. Check if the awaiter has been comepleted
    • If Yes, Go to Step 8. (Fast Path)
    • If No, remember where you have reached using the State Field. (Slow Path)
  3. Store the awaiter in a field.
  4. Schedule a continuation with the awaiter, such that when the continuation is executed, you are back at the right place.
  5. Return from the MoveNext, either to the original caller if it is the first pause, or to whatever has scheduled the continuation.
  6. When the continuation fires, set the State to -1 to indicate running.
  7. Restore the Awaiter from the field and store it back in the Stack. Remember to reset the field so that GC could take care of it.
  8. Fetch the result using GetResult() method.
  9. Continue with rest of the code.

This, was a simple asynchornous method devoid of any controls methodologies like the loops. In the next part of this series, we will use the knowledge we have gained so far to understand more complex scenarios in depth.

Once again, I would like to thank the wonderful Jon Skeets for his brillant book – C# in Depth. You ought to rename it to “C# Bible” Jon !!

Asynchronous Code – Behind the Scenes – 003

Okay, I wasn’t quite realistic in the earlier post when I mentioned we would look at MoveNext in this one. I missed an important clog of the wheel. The SetStateMachine() method.

IAsyncStateMachine.SetStateMachine

We will only breifly visit the SetStateMachine method here, as the complete picture becomes more clear when we look to details of the MoveNext() method.

So how does the SetStateMachine method looks like in the generated code. Interestingly, it has two different implementation depending on whether you are in Release or Debug mode.

// Release Mode
[DebuggerHidden]
private void SetStateMachine(IAsyncStateMachine stateMachine)
{
    <>t__builder.SetStateMachine(stateMachine);
}

void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine)
{
    this.SetStateMachine(stateMachine);
}


// Debug Mode
[DebuggerHidden]
private void SetStateMachine(IAsyncStateMachine stateMachine)
{
}

void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine)
{
    this.SetStateMachine(stateMachine);
}

As one can observe, in the Debug Mode, the method is empty. Hence the following explanation is more relavant for the Release mode.

Let’s go back a bit and think about our little Stub method.

private static Task Bar()
{
	<Bar>d__2 stateMachine = new <Bar>d__2();
	stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
	stateMachine.<>1__state = -1;
	stateMachine.<>t__builder.Start(ref stateMachine);
	return stateMachine.<>t__builder.Task;
}

When the State Machine is started by the Stub Method, it is residing on the Stack as a local variable of the Stub Method.

This is where the whole crazy stuff starts. When the State Machine pauses and resumes again, it needs a lot of information. For this to happen, when it pauses, the state machine has to box itself and store in heap, so that when it resumes, it has all the necessary informations. After it is boxed, the state machine is called on the box value using box value as arguement.

Do note that the boxing happens only once. The State machine also ensures that the builder has a reference to the single boxed version of the state machine.

This can be noticed if you dig a deep into the code of AsyncMethodBuilderCore.SetStateMachine

public void SetStateMachine(IAsyncStateMachine stateMachine)
{
	if (stateMachine == null)
	{
		throw new ArgumentNullException("stateMachine");
	}
	if (m_stateMachine != null)
	{
		throw new InvalidOperationException(Environment.GetResourceString("AsyncMethodBuilder_InstanceNotInitialized"));
	}
	m_stateMachine = stateMachine;
}

We will leave the SetStateMachine here because that’s all it does. Its role would be more visible once we examine the MoveNext method in detail.

Asynchronous Code – Behind the Scenes – 002

In the earlier part of this series, we reviewed the generic structure of decompiled async code, especially the stub method. In this part, we would continue our explore of async code and look into the State Machine. We would not delve deep into the most important MoveNext() method yet, we will first familiar with the different parts of the State Machine first.

State Machine

Let us go back to the ILSpy and see how State Machine looks like.

[StructLayout(LayoutKind.Auto)]
[CompilerGenerated]
private struct <Foo>d__1 : IAsyncStateMachine
{
	public int <>1__state;

	public AsyncTaskMethodBuilder <>t__builder;

	public int delay;

	private TaskAwaiter <>u__1;

	private void MoveNext()
	{
		// To be discussed later
	}

	void IAsyncStateMachine.MoveNext()
	{
		//ILSpy generated this explicit interface implementation from .override directive in MoveNext
		this.MoveNext();
	}

	[DebuggerHidden]
	private void SetStateMachine(IAsyncStateMachine stateMachine)
	{
		<>t__builder.SetStateMachine(stateMachine);
	}

	void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine)
	{
		//ILSpy generated this explicit interface implementation from .override directive in SetStateMachine
		this.SetStateMachine(stateMachine);
	}
}

  • IAsyncStateMachine Interface

One of the first things we would notice is the implementation of IAsyncStateMachine interface. The IAsyncStateMachine interface, which is defined under the System.Runtime.CompilerServices namespace, represents the state machine generated for the async method. The interface itself is a simple one, with just two methods in it.

public interface IAsyncStateMachine
{
    /// <summary>Moves the state machine to its next state.</summary>
    void MoveNext();
    /// <summary>Configures the state machine with a heap-allocated replica.</summary>
    /// <param name="stateMachine">The heap-allocated replica.</param>
    void SetStateMachine(IAsyncStateMachine stateMachine);
}

The MoveNext() as explained earlier, represents the heart of asynchronous code. We would, for time being, delay visiting the method for a bit longer. However, the key point to remember at this point of time is that each time the State Machine is starts or resumes (after a pause), the MoveNext() method would be called. The SetStateMachine() method associates the builder with the specific state machine.

The importance of the implementation of the interface and how it binds the state machine with the stub method could be understood by looking at the signature of the AsyncTaskMethodBuilder.Start(). The method accepts a single generic parameter, which has a constraint of having implemented the IAsyncStateMachine.

public void Start<TStateMachine>(ref TStateMachine stateMachine) where TStateMachine : IAsyncStateMachine
{
	if (stateMachine == null)
	{
		throw new ArgumentNullException("stateMachine");
	}
	ExecutionContextSwitcher ecsw = default(ExecutionContextSwitcher);
	RuntimeHelpers.PrepareConstrainedRegions();
	try
	{
		ExecutionContext.EstablishCopyOnWriteScope(ref ecsw);
		stateMachine.MoveNext();
	}
	finally
	{
		ecsw.Undo();
	}
}

We would not go too deep into AsyncTaskMethodBuilder.Start(), but key take away would be

  • The constraint applied to parameter where TStateMachine : IAsyncStateMachine
  • The method is responsible for calling IAsyncStateMachine.MoveNext()

There is another interesting fact to this look at this point. The generated State Machine has a small but significant difference depending on whether your are looking at debug/release mode code. When in release mode, the compiler optimizes the code and creates a stuct based State Machine, while in debug mode, it creates a class. This is supposed to an optimization done to so that the compiler would skip allocating memory when the awaitable has already been completed awaited. The following code displays the State Machine when decompiled in debug mode.

[CompilerGenerated]
private sealed class <Foo>d__1 : IAsyncStateMachine
{
	public int <>1__state;

	public AsyncTaskMethodBuilder <>t__builder;

	public int delay;

	private TaskAwaiter <>u__1;

	private void MoveNext()
	{
		// To be discussed later
	}

	void IAsyncStateMachine.MoveNext()
	{
		//ILSpy generated this explicit interface implementation from .override directive in MoveNext
		this.MoveNext();
	}

	[DebuggerHidden]
	private void SetStateMachine(IAsyncStateMachine stateMachine)
	{
	}

	void IAsyncStateMachine.SetStateMachine(IAsyncStateMachine stateMachine)
	{
		//ILSpy generated this explicit interface implementation from .override directive in SetStateMachine
		this.SetStateMachine(stateMachine);
	}
}


  • Fields

public int <>1__state;
public AsyncTaskMethodBuilder <>t__builder;
public int delay;
private TaskAwaiter <>u__1;

The next thing one would notice with the generated code is the presence of certains fields in the state machine. The fields could be broadly categorized into

  • Current State : As discussed in earlier post, this could have any of the following values
-1 : Not Started
-2 : Completed
Any other Value : Paused

  • Method Builder : Communicates with the async infrastructure and returns the task
  • TaskAwaiter
  • Parameters and local variables
  • Temporary Stack Variables

TaskAwaiter and parameters are used to remember the values when the State Machine resumes after a Pause. If the state machine requires a variable which it doesn’t need to remember after resuming, then it remains as private variable.

Temporary stack variables are used as a part of larger expression, when the compiler needs to remember intermediate results. For example,

int result = x + y + await task;

The most important point to remember about the Fields and Variables is that the compiler ensures it uses minimum fields/variables as possible by reusing them.

If your code have multiple await that was supposed to return

  • Task<int>
  • Task<string>
  • Task

Then the compiler would most likely create just 3 awaiters, one each for the different types involved.

That is it about the general structure of the State Machine. We would now proceed to the most important part, which is of course the MoveNext() method. We will do it in the next post.

Asynchronous Code – Behind the Scenes – 001

If you were to ask me what was the biggest milestone in .Net development, then my choice would definetly be .Net 5.0 – especially the introduction of the async/await. The more you learn about the underlying working, you cannot but stop and admire the efforts done by the lang uage developers to make our life easier.

In this mini series on asynchronous programming in .Net, we will delve deeper into the fascinating world of async await and learn more about what happens behind the scenes. I thought i would structure it as a mini-series rather than a single monolythic post as this is a vast topic (at least for an average developer like me).

Setting the Stage

Let us begin by writing a simple async method, which we would then decompile using ILSpy to know what happens beneath. We will keep the base code as simple as possible.

class Program
{
    static void Main(string[] args)
    {
        Foo(10);
    }

    static async Task Foo(int delay)
    {
        await Task.Delay(delay);
        Console.WriteLine(delay);
        await Bar();
    }

    static async Task Bar()
    {
        await Task.Delay(100);
    }
}

For the demonstration purpose, we will use ILSpy for decompiling our code, but please feel free to choose any decompiler you are comfortable with. In case you are using ILSpy, please ensure you have the following settings unchecked.

View->Options->Decompiler->C# 5.0-> Decompile async methods

This would ensure we could view the decompiled async code and the associated state machine. Okay, now let us see what ILSpy has to offer us after decompiling our code. We will read the code in parts, so that it is easier for us to understand the whole structure.

  1. Stub Method
  2. State Machine structure
  3. MoveNext Method method

Stub Method

As you would be already aware, the async methods are implemented with the help of State Machines internally.

Stubs are methods which has the same signature as your original async method and is responsible for creating the state machine. Let us check the decompiled stub method.

[AsyncStateMachine(typeof(<Foo>d__1))]
[DebuggerStepThrough]
private static Task Foo(int delay)
{
	<Foo>d__1 stateMachine = new <Foo>d__1();
	stateMachine.delay = delay;
	stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
	stateMachine.<>1__state = -1;
	stateMachine.<>t__builder.Start(ref stateMachine);
	return stateMachine.<>t__builder.Task;
}

[AsyncStateMachine(typeof(<Bar>d__2))]
[DebuggerStepThrough]
private static Task Bar()
{
	<Bar>d__2 stateMachine = new <Bar>d__2();
	stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
	stateMachine.<>1__state = -1;
	stateMachine.<>t__builder.Start(ref stateMachine);
	return stateMachine.<>t__builder.Task;
}

Let us skip the attributes for a moment (we will come back to very shortly), and first concentrate on signature of both Stub methods. As mentioned earlier, both shares the same signature with their original async methods.

The Stub Method is responsible for creating/initializing the State Machine and Starting it. The state machines are initialized with following

  • Parameters as Fields

Any parameter in the original async method are added as Fields in the State Machine. For example, if you inspect the following code from Foo Stub Method.

stateMachine.delay = delay;

The Foo method, if you remember, accepted a single parameter of Type int and was called delay. This parameter would be now added a field in the State Machine.

  • Type of Builder

The type of Builder varies depending on the return type of the Method in question.

Return TypeBuilder
TaskAsyncTaskMethodBuilder
Task<TResult>AsyncTaskMethodBuilder<T>
voidAsyncVoidMethodBuilder
Custom Task TypeBuilder specified by AsyncTaskMethodBuilderAttribute

In the above scenario, both methods returns Task. For the same reason, the both uses an AsyncTaskMethodBuilder. Please note that the Custom Task Type was introduced only with C# 7. Prior to C# 7, only the first 3 builders were applicable.

The attribute AsyncStateMachine points to the method’s particular state machine and aids in tooling.

  • State of the State Machine

A async method could be in either the following states – Not Started – Executing – Paused – Completed (Successfully or Faulted)

Out of these, the most important state for the State Machine is the Paused State. While in the Executing State, the async method is pretty much like synchronous code as it passes through each instruction. The CPU would keep track of the currently executing step via Instruction Pointer.

However, the state machine comes into picture immediately as the method pauses when it reaches an await (incomplete) expression. Please note that this is applicable only for async expression that has not been completed. In case the awaited expression is completed, the code works similiar to synchronous code and the state machine would not be brought into picture.

Each time the state has to be paused, the state is recorded so that once the operation awaited is completed, the method could be continued.

As one can expected, the Stub method would like to set the initial state of the state machine to Not Started.

The State Property of the state machine handles the current state with following value codes

ValueDescription
-1Not Started
-2Completed (Successfully/Faulted)
Any other valuePaused at an await Expression

As one can observe, these are the values which the Stub method initializes the State Machine with.

<Foo>d__1 stateMachine = new <Foo>d__1();
stateMachine.delay = delay;
stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
stateMachine.<>1__state = -1;
stateMachine.<>t__builder.Start(ref stateMachine);

Additional, it also starts the State Machine using the Builder.Start method, passing a reference to the State Machine created. The important point to note here is that the stateMachine is passed as reference to the method. This is because both StateMachine and AsyncTaskBuilder are mutable value types, and passing the instance by reference ensures no local copy are created. One can notice a lot of optimization done by the compiler.

The final step of the Stub method is of course, to return the Task object. The Task object is created by the Builder, who also ensures the Task’s state is changed accordingly as the method execution progresses.

return stateMachine.<>t__builder.Task;

When the Builder.Start method is invoked, it begins executing the MoveNext method (we will discuss it later) untill the method reaches an incomplete await expression. At this point, the MoveNext would return the Task, following up which, the Start method also returns. The task is then returned to the Caller method.

That’s it for the Stub Method, in the next section, we will look into the State Machine structure.

Json : Fetch all keys using Linq

Given a Json, how do we get all the keys in it ? Just the keys. For example, Consider the Json below.

{
Person:['Jia Anu', 'Anu Viswan','Sreena Anu'],
Details:[
{
Name: 'Jia Anu',
Age:3,
},
{
Name: 'Anu Viswan',
Age:35,
},
{
Name: 'Sreena Anu',
Age:34,
}
],
Address : 'Address here',
State : 'Kerala',
Country : 'India'
}

Given the above Json, you would like to get all the keys. You might be later interested to show this as a Menu or Dropdown, that’s down to the requirements. But, how do we get the keys in first place ? May be a Linq One liner would be great ?

Person
Details
Details.Name
Details.Age
Address
State
Country

Of course, you do have the option to loop through the JTokens, but the easiest way out would be using Linq.

var data = JObject.Parse(jsonString);
var result = data.Descendants()
.OfType<JProperty>()
.Select(f=>Regex.Replace(f.Path,@"\[[0-9]\]",string.Empty))
.Distinct();

The Regex involed is to ensure the arrays indices are removed and keys in arrays aren’t duplicated. The output of above Linq would give you collection of all keys in the Json as expected.

Why avoid Conversion Operators

No doubt that the Conversation Operators makes developers life easier by bring in a kind of substitutability, but it comes with a price. If not properly designed/used, it could result in subtle errors that can be easily ignored.
Consider the following code.
public class Car
{
  public string MakerName { get; set; }
  public Car(string Maker) => MakerName = Maker;
}
public class ElectricCar
{
  public string BrandName { get; set; }
  public ElectricCar(string Brand) => BrandName = Brand;
}
We have a class called Car, which exposes a single property – MakerName. We have another class ElectricCar, which exposes a property called BrandName. As seen, both classes are inherently incompatible, and you cannot implicitly/explicitly convert one to other. In order to solve this problem, we would be tempted to introduce an implicit conversion operator between the two classes.

We will modify ElectricCar as the following.

public class ElectricCar
{
  public string BrandName { get; set; }
  public ElectricCar(string Brand) => BrandName = Brand;
  public static implicit operator Car(ElectricCar car) => new  Car(car.BrandName);
}

We have added an implicit operator that supports conversion between ElectricCar and Car Classes. This works fine and you can now convert between the two classes. Let’s now create a method which updates the BrandName for the class Car.
public static void UpdateMarkerName(Car car, string MakerName) => car.MakerName =  MakerName;
By virtue of our implicit operator, we should be able to pass instances of both Car and Electric Car to this method. Let’s attempt to do so.
class Program
{
        
  static void Main(string[] args)
  {
     var _car = new Car("Toyota");
     var _electricCar = new ElectricCar("Toyota");

     UpdateMarkerName(_car, "Tesla");
     UpdateMarkerName(_electricCar, "Tesla");

     Console.WriteLine($"Car : {_car.MakerName}");
     Console.WriteLine($"Electric Car : {_electricCar.BrandName}");

     Console.ReadLine();
}
   public static void UpdateMarkerName(Car car, string MakerName) =>  car.MakerName = MakerName;
}
The output of above code would be curiously different, despite the implicit conversion.

Car : Tesla
Electric Car : Toyota

Of course, it is not hard to understand why. The instance of ElectricCar, thanks to the implicit operator implementation, would be converted to an instance of Car when passed to the UpdateMarkerName method. However, the scope of converted object is within the UpdateMarkerName class. The Method, updates the MakerName, but eventually throws away the instance. It has no reflection on the original instance. This is in stark contrast to when an instance of Car is being passed.

This kind of error can be easily overlooked by the developer and can lead to issues that are hard to trace. This is one of the reasons why we need to keep away from the Conversion Operators.

Nominal Vs Structural Type System

Anonymous Types and Tuples might look very similar, but there is one significant difference which separates them. As always, nothing can be more explanatory than writing code.
var anonymous1 = new { x = 1, y = 1 };
var anonymous2 = new { y = 1, x = 1 };

var tuple1 = (x : 1, y : 1);
var tuple2 = (y: 1, x: 1);

Console.WriteLine($"Anonymous Type :  {anonymous1.Equals(anonymous2)}");
Console.WriteLine($"Tuple Type : {tuple1.Equals(tuple2)}");
Consider the above code. What could be the output ? Are both false or are both true ? Interestingly, the output is as follows.
Anonymous Type : False
Tuple Type : True
Though syntactically similar, both Anonymous Types and Tuples vary in a significant factor – the associated type system.  While Anonymous Types is based on Nominal Type System, Tuple is based on Structural Type System. This results in the behavior seen in the above code, with both tuples being equal despite the obvious naming pattern difference.
That brings us to another scenario, what if the sub types for tuples are differently named. For example, consider following code.
var tuple1 = (x : 1, y : 1);
var tuple3 = (a: 1, b: 1);
Console.WriteLine($"Tuple Type : {tuple1.Equals(tuple3)}");
As you would have rightly guessed, the output would be “true” as the tuple is only interested in the types and not the names.

Benchmarking Span<T> Performance

Span<T> is yet another addition to C# 7.x and is particularly useful in developing memory intensive applications.  So what is Span all about ?

 

As Microsoft describes it, Span<T> is a new value Type which enables the representation of contiguous regions of arbitrary memory, regardless of whether the memory is associated with a managed object, is provided by native code  or is on the stack, with a performance characteristics like that of an array.

 

It feels like you can use pointers, but without entering ‘unsafe’ code. That definitely is interesting. We will be looking into Span<T> closely in upcoming posts, but for this post, we will compare Span with Strings when parsing sub strings.  We will begin by how sub strings work. We will keep it simple for sake of the example. Consider the following code.
public void DummyMethod(string data) { }

string keyString = "234567";
DummyMethod(keyString.Substring(0, 2));
DummyMethod(keyString.Substring(2, 2));
DummyMethod(keyString.Substring(4, 2));

Remember string is immutable, and each time we are invoking the Substring method, we are allocating a new memory location is allocated with the substring. Now that isn’t a scenario you wouldn’t want to be if you are working on a memory intensive application and you want to parse a really long string. But what is we could parse the string right from the memory location allocated for keyString. That would be extremely efficient right ?

That’s where Span<T> comes in. It would allow us to point to a contiguous region of memory, and allow us to parse through it without needing a different memory allocation. Let’s rewrite the above code using Span<T>.

public void DummyMethod(ReadOnlySpan data) { }

ReadOnlySpan keyString = "234567".AsReadOnlySpan();
DummyMethod(keyString.Slice(0, 2));
DummyMethod(keyString.Slice(2, 2));
DummyMethod(keyString.Slice(4, 2));

Notice that the string has a nice little extension method to create a ReadOnlySpan. We are also using the Slice method (instead of Substring) to access the specific part of memory location. We will do a bit of bench marking to understand the performance implication. Let’s create a BenchmarkDemo Class for our demonstrative purpose. Complete source code is available at my Github.

[Benchmark]
public void UsingSubString()
{
    string keyString = "long string";
    for (int i = 0; i < IterationLimit; i++)
        DummyStringMethod(keyString.Substring(0, i));
}
void DummyStringMethod(string _) { }

[Benchmark]
public void UsingSpan()
{
    ReadOnlySpan keyString = "long string".AsReadOnlySpan();
    for (int i = 0; i < IterationLimit; i++)
        DummySpanMethod(keyString.Slice(0, i));
}

void DummySpanMethod(ReadOnlySpan _) { }

The above code is for demonstrative purpose, hence we are using a dummy string "long string". In the example in Github, you can find the same replaced with lorem ipsum.

Alright, now lets run the Benchmark code and analyse the memory allocation against iteration loops of 10,100, and 400. We are using BenchmarkDotNet for example.

Benchmark Span

As you can see the “UsingSubString” method uses a huge amount of memory as the calls to substring increases. On other hand, Span based method hardly uses any and doesn’t have any different as the number of calls increases.

That’s it for now, we will investigate more on Span<T> in upcoming posts.