The suspend modifier — under the hood

Kotlin Vocabulary: Coroutines

Manuel Vivo
Mar 24, 2020 · 7 min read

Kotlin coroutines introduced the suspend modifier in our daily life as Android developers. Are you curious to see what’s happening under the hood? How is the compiler transforming the code to be able to suspend and resume the execution of coroutines?

Knowing this will help you better understand why a suspend function won’t return until all the work that it started has completed and how the code can suspend without blocking threads.

TL;DR; The Kotlin compiler will create a state machine for every suspend function that manages the coroutine’s execution for us!

📚 New to coroutines on Android? Check out these Coroutines codelabs:

If you prefer to watch a video about this, check this out:

Coroutines 101

Coroutines simplify asynchronous operations on Android. As explained in the documentation, we can use them to manage asynchronous tasks that might otherwise block the main thread and cause your app to freeze.

Coroutines are also helpful to replace callback-based APIs with imperative looking code. As example, check out this asynchronous code that uses callbacks:

Those callbacks can be converted to sequential function calls using coroutines:

In the coroutines code, we added the suspend modifier to the function. That tells the compiler that this function needs to be executed inside a coroutine. As a developer, you can think of a suspend function as a regular function whose execution might be suspended and resumed at some point.

Unlike callbacks, coroutines provide an easy way to swap between threads and handle exceptions.

But, what’s the compiler actually doing under the hood when we mark the function as suspend?

Suspend under the hood

Back to the loginUser suspend function, notice that the other functions it calls are also suspend functions:

In a nutshell, the Kotlin compiler will take suspend functions and convert them to an optimised version of callbacks using a finite state machine (which we’ll cover later).

You got it right, the compiler will write those callbacks for you!

The way suspend functions communicate with each other is with Continuation objects. A Continuation is just a generic callback interface with some extra information. As we will see later, it will represent the generated state machine of a suspend function.

Let’s take a look at its definition:

  • context will be the CoroutineContext to be used in that continuation.
  • resumeWith resumes execution of the coroutine with a Result, that can contain either a value which is the result of the computation that caused the suspension or an exception.

Note: From Kotlin 1.3 onwards, you can also use the extension functions resume(value: T) and resumeWithException(exception: Throwable) which are specialised versions of the resumeWith call.

The compiler will replace the suspend modifier with the extra parameter completion (of type Continuation) in the function signature that will be used to communicate the result of the suspend function to the coroutine that called it:

For simplicity, our example will return Unit instead of User. The User object will be “returned” in the added Continuation parameter.

The bytecode of suspend functions actually return Any? because it's a union type of T | COROUTINE_SUSPENDED. That allows the function to return synchronously when it can.

Note: If you mark a function that doesn’t call other suspend functions with the suspend modifier, the compiler will add the extra Continuation parameter but won’t do anything with it, the function body’s bytecode will look like a regular function.

You can also see the Continuation interface in other places:

  • When converting callback-based APIs to coroutines using suspendCoroutine or suspendCancellableCoroutine (which you should always prefer), you directly interact with a Continuation object to resume the coroutine that got suspended after running the block of code passed as a parameter.
  • You can start a coroutine with the startCoroutine extension function on a suspend function. It takes a Continuation object as a parameter that will get called when the new coroutine finishes with either a result or an exception.

You can swap between different Dispatchers to execute computations on different threads. How does Kotlin know where to resume a suspended computation?

There’s a subtype of Continuation called DispatchedContinuation whose resume function makes a dispatch call to the Dispatcher available in the CoroutineContext. All Dispatchers will call dispatch except Dispatchers.Unconfined whose isDispatchNeeded function override (that is called before dispatch) always returns false.

The generated State machine

Disclaimer: The code to be shown in the rest of the article will not fully match the bytecode generated by the compiler. It will be Kotlin code accurate enough to allow you to understand what’s really happening internally. This representation is generated by Coroutines version 1.3.3 and might change in future versions of the library.

The Kotlin compiler will identify when the function can suspend internally. Every suspension point will be represented as a state in the finite state machine. These states are represented with labels by the compiler:

For a better representation of the state machine, the compiler will use a when statement to implement the different states:

This code is incomplete since the different states have no way to share information. The compiler will use the same Continuation object in the function to do it. This is why the generic of the Continuation is Any? instead of the return type of the original function (i.e. User).

Furthermore, the compiler will create a private class that 1) holds the required data and 2) calls the loginUser function recursively to resume execution. You can check out an approximation of the generated class below.

Disclaimer: Comments are not generated by the compiler. I added them to explain what they do and make following the code easier to follow.

As invokeSuspend will call loginUser again with just the information of the Continuation object, the rest of parameters in the loginUser function signature become nullable. At this point, the compiler just needs to add information on how to move between states.

The first thing it needs to do is know if 1) it’s the first time the function is called or 2) the function has resumed from a previous state. It does it by checking if the continuation passed in is of type LoginUserStateMachine or not:

If it’s the first time, it will create a new LoginUserStateMachine instance and will store the completion instance received as a parameter so that it remembers how to resume the function that called this one. If it’s not, it will just carry on executing the state machine (the suspend function).

Now, let’s see the code that the compiler generates for moving between states and sharing information between them.

Spend some time going through the code above and see if you can spot the differences with the previous snippets of code. Let’s see what the compiler generates:

  • The when statement’s argument is the label from inside the LoginUserStateMachine instance.
  • Every time a new state is processed, there’s a check in case a failure happened when this function was suspended.
  • Before calling the next suspend function (i.e. logUserIn), the label of the LoginUserStateMachine instance is updated to the next state.
  • When inside this state machine there’s a call to another suspend function, the instance of the continuation (of type LoginUserStateMachine) is passed as a parameter. The suspend function to be called has also been transformed by the compiler and it’s another state machine like this one that takes a continuation object as a parameter! When the state machine of that suspend function finishes, it will resume the execution of this state machine.

The last state is different since it has to resume the execution of the function that called this one, as you can see in the code, it calls resume on the cont variable stored (at construction time) in LoginUserStateMachine:

As you see, the Kotlin compiler is doing a lot for us! From this suspend function:

The compiler generated all of this for us:

The Kotlin compiler transforms every suspend function to be a state machine, which optimises using callbacks every time a function needs to suspend.

Knowing what the compiler does under the hood now, you can better understand why a suspend function won’t return until all the work that it started has completed. Also, how the code can suspend without blocking threads: the information of what needs to be executed when the function is resumed is stored in the Continuation object!

Android Developers

The official Android Developers publication on Medium