Create a future with coroutine in C++

I didn’t write on this blog for years now. Here is what I am doing right now. I am a C++ software engineer in Diagdev, a hematology-related company based in the south of France near Montpellier. I also developed a little library: Little Type Library. The goal of this library is to add range and concept like to C++17. It also brings other functional features with the pipe notation or the >> notation. We will see how to create a future using coroutine

What is a coroutine?

A coroutine, even if the name is complicated, is a simple thing.
Functions begin, do things, and end. A coroutine is a kind of generalization over functions and they can be stopped and resumed in the same, or better, in another thread.

A coroutine can be stopped either at the beginning, during the processing, or at the end.

What is the problem with conventionnal future?

Let’s say we have this code.

#include <future>

int main() {
    auto performSomething = []{};
    auto future = std::async(performSomething);
    
    doSomething();
    
    future.wait();
}

If performSomething and doSomething last the same duration, there is no issue, but if performSomething lasts longer, the future.wait() will make the main thread stalls during some instant, and so, you lose performance. If the performSomething lasts longer, there is a thread that does not do anything and so, we lose performance.

Monadic expression is a way to solve it. The idea is to give the future one callback that will be automatically called when the asynchronous function ends. If you want to know more about this, I wrote an article a few years ago.

How coroutines can solve the problem

Actually, my solution using coroutine is like monadic expression. However, instead of giving a callback, you write the function normally.

Here is the code we want to achieve.

int square(int x) {
  return x * x;
}

task f() {
  auto squared6 = co_await async(square, 6);

  std::cout << squared6 << std::endl;
}

int main() {
  std::jthread thread;
  ::thread = &thread;

  f();

  doSomethingHeavy();
  return 0;
}

I don’t manage in a clean way my threads to be simpler. I used jthread instead of thread for its auto-join. This code is simple. The main function call f which is a coroutine. The coroutine f calls async with square and 6 as arguments. Async returns an awaitable awaited in f. The await launch square in another thread. f saves its state and returns to main. Once the thread finishes computing square, it resumes the coroutine f, and prints the squared value.

Yes you read it properly, the std::cout << squared6 << std::endl is executed in the same thread as square.

So, let’s create a future using coroutine !

Async coroutine

The job of the async coroutine is to return a future and launch the given function when we co_await the future. So here is the code :

template <typename F, typename... Args>
future<std::invoke_result_t<F, Args...>> async(F f, Args... args) {
  std::cout << "async" << std::endl;
  co_return f(args...);
}

Since we use the co_return keyword, async is a coroutine. We don’t want to execute f directly, so we need to stop the coroutine at the beginning. We also need to stop at the end to avoid use after free. Since the coroutine is stopped before its beginning, we need to save the result of the function. To finish, the future has the handle of the coroutine to be able to resume it.

Here is the full code of future:

template <typename T> struct future {
  struct promise_type {
    T value;
    future get_return_object() {
      return {std::coroutine_handle<promise_type>::from_promise(*this)};
    }
    std::suspend_always initial_suspend() noexcept {
      return {};
    }
    std::suspend_always final_suspend() noexcept {
      return {};
    }
    void return_value(T x) {
      value = std::move(x);
    }
    void unhandled_exception() noexcept {}
  };

  std::coroutine_handle<promise_type> coro;

  future(std::coroutine_handle<promise_type> coro) : coro{coro} {}

  ~future() {
    if (coro)
      coro.destroy();
  }

  AwaitableFuture operator co_await() {
    return {*this};
  }
};

We see that the future has an co_await operator. This operator needs an awaitable object. An awaitable has 3 methods.

• await_ready: to know if we must stop the coroutine or not
• await_suspend: Called when we suspend the coroutine, it is here we will launch the other thread.
• await_resume: To return the result to the called of co_await.

For our future, the awaitable will look like that:

  struct AwaitableFuture {
    future &m_future;
    bool await_ready() const noexcept { return false; }

    void await_suspend(std::coroutine_handle<> handle) {
      *thread = std::jthread([this, handle] {
        m_future.coro.resume();
        handle.resume();
      });
    }

    T await_resume() {
      return m_future.coro.promise().value;
    }
  };

Why do we need to call resume twice? Let’s remind the usage of the future :

task f() {
  auto squared6 = co_await async(square, 6);

  std::cout << squared6 << std::endl;
}

f is a coroutine, but async also! The future coroutine is to resume the async one, and the await_suspend one is to resume f. The order is important, else, you will not compute the value before resuming f.

Since f is a coroutine, here is the simple code for the task :

struct task {
  struct promise_type {
    task get_return_object() { return {}; }
    std::suspend_never initial_suspend() noexcept { return {}; }
    std::suspend_never final_suspend() noexcept { return {}; }
    void return_void() {}
    void unhandled_exception() noexcept {}
  };
};

Here is a full code with an online compiler if you want to try it.

Conclusion

I hope you understand this article well, and if you have some questions, don’t hesitate to ask :). For the future, I could write an article about how to create a when_all using coroutines, or how to use a thread pool with coroutines.

Thanks for reading !