C++ coroutines: Defining the co_await operator

Raymond Chen

At the start of this series, I noted that there are three steps in obtaining an awaiter for an awaitable object. The first two were marked “We’re not read to talk about this yet.”

Well, now we’re ready to talk about one of them.

When you do a co_await x, the compiler tries to come up with a thing called an awaiter.

  1. (We’re not ready to talk about step 1 yet.)
  2. If there is a defined operator co_await for x, then invoke it to obtain the awaiter.
  3. Otherwise, x is its own awaiter.

The search for a operator co_await follows the usual rules for operator overloading: See if the operator is overloaded by the object itself. If not, then look for a free definition.

The case of an awaitable object implementing its own operator co_await is rather unusual. After all, if you can add an operator co_await to a class, then you may as well just add the await_* methods to the class while you’re at it.

The more common case is a free operator co_await, because that lets you add co_await support to a type that wasn’t initially defined with coroutines in mind. For example, the C++/WinRT library defines an operator co_await(std::chrono::duration): If you co_await 30s;, it will pause the coroutine for 30 seconds.

The operator co_await is a unary operator, so the member function version is defined with no parameters (in which case, this is the object being awaited), and the free function has one parameter (in which case, the parameter is the object being awaited). In both cases, the return value is the awaiter to use.

Just as an exercise, let’s define a co_await operator that takes a Core­Dispatcher and switches to that dispatcher’s thread.

We already wrote an awaitable to do this last time:

auto ensure_dispatcher_thread(CoreDispatcher dispatcher)
  struct awaiter : std::experimental::suspend_always
    CoreDispatcher dispatcher;

    bool await_ready() { return dispatcher.HasThreadAccess(); }

    void await_suspend(
        std::experimental::coroutine_handle<> handle)
                           [handle]{ handle(); });
  return awaiter{ {}, std::move(dispatcher) };

We can add co_await support to Core­Dispatcher by defining an operator co_await:

auto operator co_await(CoreDispatcher dispatcher)
  return ensure_dispatcher_thread(std::move(dispatcher));

Now you can co_await a Core­Dispatcher directly.

co_await this.Dispatcher();

The search for an operator co_await finds the operator we defined above, so it is invoked to produce the awaiter. The return value is the awaiter inside the ensure_dispatcher_thread function, so that’s what ends up being used to control the suspension and resumption of the coroutine.

Next time, we’ll look a bit more at the consequences of the operator co_await search algorithm.


1 comment

Discussion is closed. Login to edit/delete existing comments.

  • Kalle Niemitalo 0

    I had difficulty finding operator co_await(std::chrono::duration). C++/WinRT apparently defines it as winrt::operator co_await(winrt::Windows::Foundation::TimeSpan), where winrt::Windows::Foundation::TimeSpan is the same as std::chrono::duration, and there are conversions to this type from other types of durations.

    The other operator co_await functions in C++/WinRT are in namespace winrt::Windows::Foundation, like their parameters (such as IAsyncAction), so they can be found with argument-dependent lookup.

Feedback usabilla icon