A Unified Executors Proposal for C++ | P0443R12

Title: A Unified Executors Proposal for C++
Authors: Jared Hoberock, jhoberock@nvidia.com
Michael Garland, mgarland@nvidia.com
Chris Kohlhoff, chris@kohlhoff.com
Chris Mysen, mysen@google.com
Carter Edwards, hcedwar@sandia.gov
Gordon Brown, gordon@codeplay.com
David Hollman, dshollm@sandia.gov
Lee Howes, lwh@fb.com
Kirk Shoop, kirkshoop@fb.com
Lewis Baker, lbaker@fb.com
Eric Niebler, eniebler@fb.com
Other Contributors: Hans Boehm, hboehm@google.com
Thomas Heller, thom.heller@gmail.com
Bryce Lelbach, brycelelbach@gmail.com
Hartmut Kaiser, hartmut.kaiser@gmail.com
Bryce Lelbach, brycelelbach@gmail.com
Gor Nishanov, gorn@microsoft.com
Thomas Rodgers, rodgert@twrodgers.com
Michael Wong, michael@codeplay.com
Document Number: P0443R12
Date: 2020-01-13
Audience: SG1 - Concurrency and Parallelism, LEWG
Reply-to: sg1-exec@googlegroups.com
Abstract: This paper proposes a programming model for executors, which are modular components for creating execution, and senders, which are lazy descriptions of execution.

1 Design Document

1.1 Motivation

When we imagine the future of C++ programs, we envision elegant compositions of networked, asynchronous parallel computations accelerated by diverse hardware, ranging from tiny mobile devices to giant supercomputers. In the present, hardware diversity is greater than ever, but C++ programmers lack satisfying parallel programming tools for them. Industrial-strength concurrency primitives like std::thread and std::atomic are powerful but hazardous. std::async and std::future suffer from well-known problems. And the standard algorithms library, though parallelized, remains inflexible and non-composable.

To address these temporary challenges and build toward the future, C++ must lay a foundation for controlling program execution. First, C++ must provide flexible facilities to control where and when work happens. This paper proposes a design for those facilities. After much discussion and collaboration, SG1 adopted this design by universal consensus at the Cologne meeting in 2019.

1.2 Usage Example

This proposal defines requirements for two key components of execution: a work execution interface and a representation of work and their interrelationships. Respectively, these are executors and senders and receivers:

// make P0443 APIs in namespace std::execution available
using namespace std::execution;

// get an executor from somewhere, e.g. a thread pool
std::static_thread_pool pool(16);
executor auto ex = pool.executor();

// use the executor to describe where some high-level library should execute its work
perform_business_logic(ex);

// alternatively, use primitive P0443 APIs directly

// immediately submit work to the pool
execute(ex, []{ std::cout << "Hello world from the thread pool!"; });

// immediately submit work to the pool and require this thread to block until completion
execute(std::require(ex, blocking.always), foo);

// describe a chain of dependent work to submit later
sender auto begin    = schedule(ex);
sender auto hi_again = then(begin, []{ std::cout << "Hi again! Have an int."; return 13; });
sender auto work     = then(hi_again, [](int arg) { return arg + 42; });

// prints the final result
receiver auto print_result = as_receiver([](int arg) { std::cout << "Received " << std::endl; });

// submit the work for execution on the pool by combining with the receiver 
submit(work, print_result);

// Blue: proposed by P0443. Teal: possible extensions.

1.3 Executors Execute Work

As lightweight handles, executors impose uniform access to execution contexts.

Executors provide a uniform interface for work creation by abstracting underlying resources where work physically executes. The previous code example’s underlying resource was a thread pool. Other examples include SIMD units, GPU runtimes, or simply the current thread. In general, we call such resources execution contexts. As lightweight handles, executors impose uniform access to execution contexts. Uniformity enables control over where work executes, even when it is executed indirectly behind library interfaces.

The basic executor interface is the execute function through which clients execute work:

// obtain an executor
executor auto ex = ...

// define our work as a nullary invocable
invocable auto work = []{ cout << "My work" << endl; };

// execute our work via the execute customization point
execute(ex, work);

On its own, execute is a primitive “fire-and-forget”-style interface. It accepts a single nullary invocable, and returns nothing to identify or interact with the work it creates. In this way, it trades convenience for universality. As a consequence, we expect most programmers to interact with executors via more convenient higher-level libraries, our envisioned asynchronous STL being such an example.

Consider how std::async could be extended to interoperate with executors enabling client control over execution:

template<class Executor, class F, class Args...>
future<invoke_result_t<F,Args...>> async(const Executor& ex, F&& f, Args&&... args) {
  // package up the work
  packaged_task work(forward<F>(f), forward<Args>(args)...);

  // get the future
  auto result = work.get_future();

  // execute work on the given executor
  execution::execute(ex, move(work));

  return result;
}

The benefit of such an extension is that a client can select from among multiple thread pools to control exactly which pool std::async uses simply by providing a corresponding executor. Inconveniences of work packaging and submission become the library’s responsibility.

Authoring executors. Programmers author custom executor types by defining a type with an execute function. Consider the implementation of an executor whose execute function executes the client’s work “inline”:

struct inline_executor {
  // define execute
  template<class F>
  void execute(F&& f) const noexcept {
    std::invoke(std::forward<F>(f));
  }

  // enable comparisons
  auto operator<=>(const inline_executor&) const = default;
};

Additionally, a comparison function determines whether two executor objects refer to the same underlying resource and therefore execute with equivalent semantics. Concepts executor and executor_of summarize these requirements. The former validates executors in isolation; the latter, when both executor and work are available.

Executor customization can accelerate execution or introduce novel behavior. The previous example demonstrated custom execution at the granularity of a new executor type, but finer-grained and coarser-grained customization techniques are also possible. These are executor properties and control structures, respectively.

Executor properties communicate optional behavioral requirements beyond the minimal contract of execute, and this proposal specifies several. We expect expert implementors to impose these requirements beneath higher-level abstractions. In principle, optional, dynamic data members or function parameters could communicate these requirements, but C++ requires the ability to introduce customization at compile time. Moreover, optional parameters lead to combinatorially many function variants.

Instead, statically-actionable properties factor such requirements and thereby avoid a combinatorial explosion of executor APIs. For example, consider the requirement to execute blocking work with priority. An unscalable design might embed these options into the execute interface by multiplying individual factors into separate functions: execute, blocking_execute, execute_with_priority, blocking_execute_with_priority, etc.

Executors avoid this unscalable situation by adopting P1393’s properties design based on require and prefer:

// obtain an executor
executor auto ex = ...;

// require the execute operation to block
executor auto blocking_ex = std::require(ex, execution::blocking.always);

// prefer to execute with a particular priority p
executor auto blocking_ex_with_priority = std::prefer(blocking_ex, execution::priority(p));

// execute my blocking, possibly prioritized work
execution::execute(blocking_ex_with_priority, work);

Each application of require or prefer transforms an executor into one with the requested property. In this example, if ex cannot be transformed into a blocking executor, the call to require will fail to compile. prefer is a weaker request used to communicate hints and consequently always succeeds because it may ignore the request.

Consider a version of std::async which never blocks the caller:

template<executor E, class F, class... Args>
auto really_async(const E& ex, F&& f, Args&&... args) {
  using namespace execution;

  // package up the work
  packaged_task work(forward<F>(f), forward<Args>(args)...);

  // get the future
  auto result = work.get_future();

  // execute the nonblocking work on the given executor
  execute(require(ex, blocking.never), move(work));

  return result;
}

Such an enhancement could address a well-known hazard of std::async:

// confusingly, always blocks in the returned but discarded future's destructor
std::async(foo);

// *never* blocks
really_async(foo);

Control structures permit customizations at a higher level of abstraction by allowing executors to “hook” them and is useful when an efficient implementation is possible on a particular execution context. The first such control structure this proposal defines is bulk_execute, which creates a group of function invocations in a single operation. This pattern permits a wide range of efficient implementations and is of fundamental importance to C++ programs and the standard library.

By default, bulk_execute invokes execute repeatedly, but repeatedly executing individual work items is inefficient at scale. Consequently, many platforms provide APIs that explicitly and efficiently execute bulk work. In such cases, a custom bulk_execute avoids inefficient platform interactions via direct access to these accelerated bulk APIs while also optimizing the use of scalar APIs.

bulk_execute receives an invocable and an invocation count. Consider a possible implementation:

struct simd_executor : inline_executor { // first, satisfy executor requirements via inheritance
  template<class F>
  simd_sender bulk_execute(F f, size_t n) const {
    #pragma simd
    for(size_t i = 0; i != n; ++i) {
      std::invoke(f, i);
    }

    return {};
  }
};

To accelerate bulk_execute, simd_executor uses a SIMD loop.

bulk_execute should be used in cases where multiple pieces of work are available at once:

template<class Executor, class F, class Range>
void my_for_each(const Executor& ex, F f, Range rng) {
  // request bulk execution, receive a sender
  sender auto s = execution::bulk_execute(ex, [=](size_t i) {
    f(rng[i]);
  });

  // initiate execution and wait for it to complete
  execution::sync_wait(s);
}

simd_executor’s particular bulk_execute implementation executes “eagerly”, but bulk_execute’s semantics do not require it. As my_for_each demonstrates, unlike execute, bulk_execute is an example of a “lazy” operation whose execution may be optionally postponed. The token this bulk_execute returns is an example of a sender a client may use to initiate execution or otherwise interact with the work. For example, calling sync_wait on the sender ensures that the bulk work completes before the caller continues. Senders and receivers are the subject of the next section.

1.4 Senders and Receivers Represent Work

The executor concept addresses a basic need of executing a single operation in a specified execution context. The expressive power of executor is limited, however: since execute returns void instead of a handle to the just-scheduled work, the executor abstraction gives no generic way to chain operations and thereby propagate values, errors, and cancellation signals downstream; no way to handle scheduling errors occurring between when work submission and execution; and no convenient way to control the allocation and lifetime of state associated with an operation.

Without such controls, it is not possible to define Generic (in the Stepanov sense) asynchronous algorithms that compose efficiently with sensible default implementations. To fill this gap, this paper proposes two related abstractions, sender and receiver, concretely motivated below.

1.4.1 Generic async algorithm example: retry

retry is the kind of Generic algorithm senders and receivers enable. It has simple semantics: schedule work on an execution context; if the execution succeeds, done; otherwise, if the user requests cancellation, done; otherwise, if a scheduling error occurs, try again.

template<invocable Fn>
void retry(executor_of<Fn> auto ex, Fn fn) {
  // ???
}

Executors alone prohibit a generic implementation because they lack a portable way to intercept and react to scheduling errors. Later we show how this algorithm might look when implemented with senders and receivers.

1.4.2 Goal: an asynchronous STL

Suitably chosen concepts driving the definition of Generic async algorithms like retry streamline the creation of efficient, asynchronous graphs of work. Here is some sample syntax for the sorts of async programs we envision (borrowed from P1897):

sender auto s = just(3) |                                  // produce '3' immediately
                via(scheduler1) |                          // transition context
                then([](int a){return a+1;}) |             // chain continuation
                then([](int a){return a*2;}) |             // chain another continuation
                via(scheduler2) |                          // transition context
                handle_error([](auto e){return just(3);}); // with default value on errors
int r = sync_wait(s);                                      // wait for the result

It should be possible to replace just(3) with a call to any asynchronous API whose return type satisfies the correct concept and maintain this program’s correctness. Generic algorithms like when_all and when_any would permit users to express fork/join concurrency in their DAGs. As with STL’s iterator abstraction, the cost of satisfying the conceptual requirements are offset by the expressivity of a large reusable and composable library of algorithms.

1.4.3 Current techniques

There are many techniques for creating chains of dependent asynchronous execution. Ordinary callbacks have enjoyed success in C++ and elsewhere for years. Modern codebases have switched to variations of future abstractions that support continuations (e.g., std::experimental::future::then). In C++20 and beyond, we could imagine standardizing on coroutines, so that launching an async operation returns an awaitable. Each of these approaches has strengths and weaknesses.

Futures, as traditionally realized, require the dynamic allocation and management of a shared state, synchronization, and typically type-erasure of work and continuation. Many of these costs are inherent in the nature of “future” as a handle to an operation that is already scheduled for execution. These expenses rule out the future abstraction for many uses and makes it a poor choice for a basis of a Generic mechanism.

Coroutines suffer many of the same problems but can avoid synchronizing when chaining dependent work because they typically start suspended. In many cases, coroutine frames require unavoidable dynamic allocation. Consequently, coroutines in embedded or heterogeneous environments require great attention to detail. Neither are coroutines good candidates for cancellation because the early and safe termination of coordinating coroutines requires unsatisfying solutions. On the one hand, exceptions are inefficient and disallowed in many environments. Alternatively, clumsy ad hoc mechanisms, whereby co_yield returns a status code, hinder correctness. P1662 provides a complete discussion.

Callbacks are the simplest, most powerful, and most efficient mechanism for creating chains of work, but suffer problems of their own. Callbacks must propagate either errors or values. This simple requirement yields many different interface possibilities, but the lack of a standard obstructs Generic design. Additionally, few of these possibilities accomodate cancellation signals when the user requests upstream work to stop and clean up.

1.5 Receiver, sender, and scheduler

With the preceding as motivation, we introduce primitives to address the needs of Generic asynchronous programming in the presence of value, error, and cancellation propagation.

1.5.1 Receiver

A receiver is simply a callback with a particular interface and semantics. Unlike a traditional callback which uses function-call syntax and a single signature handling both success and error cases, a receiver has three separate channels for value, error, and “done” (aka cancelled).

These channels are specified as customization points, and a type R modeling receiver_of<R,Ts...> supports them:

std::execution::set_value(r, ts...); // signal success, but set_value itself may fail
std::execution::set_error(r, ep);    // signal error (ep is std::exception_ptr), never fails
std::execution::set_done(r);         // signal stopped, never fails

Exactly one of the three functions must be called on a receiver before it is destroyed. Each of these interfaces is considered “terminal”. That is, a particular receiver may assume that if one is called, no others ever will be. The one exception being if set_value exits with an exception, the receiver is not yet complete. Consequently, another function must be called on it before it is destroyed. After a failed call to set_value, correctness requires a subsequent call either to set_error or set_done; a receiver need not guarantee that a second call to set_value is well-formed. Collectively, these requirements are the “receiver contract”.

Although receiver’s interface appears novel at first glance, it remains just a callback. Moreover, receiver’s novelty disappears when recognizing that std::promise’s set_value and set_exception provide essentially the same interface. This choice of interface and semantics, along with sender, facilitate the Generic implementation of many useful async algorithms like retry.

1.5.2 Sender

A sender represents work that has not been scheduled for execution yet, to which one must add a continuation (a receiver) and then “launch”, or enqueue for execution. A sender’s duty to its connected receiver is to fulfill the receiver contract by ensuring that one of the three receiver functions returns normally.

This proposal currently fuses these two operations — attach a continuation and launch for execution — into the single operation submit. P2006 proposes to split submit into a connect operation that packages a sender and a receiver into an executable state, and a start operation that enqueues the state for execution.

// P0443R12
std::execution::submit(snd, rec);

// P0443R12 + P2006R0
auto state = std::execution::connect(snd, rec);
// ... later
std::execution::start(state);

Such a split offers interesting opportunities for optimization, and will harmonize senders with coroutines.

The sender concept itself places no requirements on the execution context on which a sender’s work executes. Instead, specific models of the sender concept may offer stronger guarantees about the context from which the receiver’s methods will be invoked. This is particularly true of the senders created by a scheduler.

1.5.3 Scheduler

Many generic async algorithms create multiple execution agents on the same execution context. Therefore, it is insufficient to parameterize these algorithms with a single-shot sender completing in a known context. Rather, it makes sense to pass these algorithms a factory of single-shot senders. Such a factory is called a “scheduler”, and it has a single basis operation: schedule:

sender auto s = std::execution::schedule(sched);
// OK, s is a single-shot sender of void that completes in sched's execution context

Like executors, schedulers act as handles to an execution context. Unlike executors, schedulers submit execution lazily, but a single type may simultaneously model both concepts. We envision that subsumptions of the scheduler concept will add the ability to postpone or cancel execution until after some time period has elapsed.

1.6 Senders, receivers, and generic algorithms

Useful concepts constrain generic algorithms while allowing default implementations via those concepts’ basis operations. Below, we show how these sender and receiver provide efficient default implementations of common async algorithms. We envision that most generic async algorithms will be implemented as taking a sender and returning a sender whose connect method wraps its receiver an adaptor that implements the algorithm’s logic. The then algorithm below, which chains a continuation function on a sender, is a simple demonstration.

1.6.1 Algorithm then

The following code implements a then algorithm that, like std::experimental::future::then, schedules a function to be applied to the result of an asynchronous operation when available. This code demonstrates how an algorithm can adapt receivers to codify the algorithm’s logic.

template<receiver R, class F>
struct _then_receiver : R { // for exposition, inherit set_error and set_done from R
    F f_;

    // Customize set_value by invoking the callable and passing the result to the base class
    template<class... As>
      requires receiver_of<R, invoke_result_t<F, As...>>
    void set_value(Args&&... args) && noexcept(/*...*/) {
        ::set_value((R&&) *this, invoke((F&&) f_, (As&&) as...));
    }

    // Not shown: handle the case when the callable returns void
};

template<sender S, class F>
struct _then_sender : _sender_base {
    S s_;
    F f_;

    template<receiver R>
      requires sender_to<S, _then_receiver<R, F>>
    state_t<S, _then_receiver<R, F>> connect(R r) && {
        return ::connect((S&&)s_, _then_receiver<R, F>{(R&&)r, (F&&)f_});
    }
};

template<sender S, class F>
sender auto then(S s, F f) {
    return _then_sender{{}, (S&&)s, (F&&)f};
}

Given some asynchronous, sender-returning API async_foo, a user of then can execute some code once the async result is available:

sender auto s = then(async_foo(args...), [](auto result) {/* stuff... */});

This builds a composed asynchronous operation. When the user wants to schedule this operation for execution, they would connect a receiver, and then call start on the resulting operation state.

Scheduling work on an execution context can also be done with then. Given a static_thread_pool object pool that satisfied the scheduler concept, a user may do the following:

sender auto s = then(
    std::execution::schedule( pool ),
    []{ std::printf("hello world"); } );

This creates a sender that, when submitted, will call printf from a thread in the thread pool.

There exist heterogeneous computing environments that are unable to execute arbitrary code. For those, an implementation of then as shown above would either not work or would incur the cost of a transition to the host in order to execute the unknown code. Therefore, then itself and several other fundamental algorithmic primitives, would themselves need to be customizable on a per-execution context basis.

A full working example of then can be found here: https://godbolt.org/z/dafqM-

1.6.2 Algorithm retry

As described above, the idea of retry is to retry the async operation on failure, but not on success or cancellation. Key to a correct generic implementation of retry is the ability to distinguish the error case from the cancelled case.

As with the then algorithm, the retry algorithm places the logic of the algorithm into a custom receiver to which the sender to be retried is connect-ed. That custom receiver has set_value and set_done members that simply pass their signals through unmodified. The set_error member, on the other hand, reconstructs the operation state in-place by making another call to connect with the original sender and a new instance of the custom receiver. That new operation state is then start-ed again, which effectively causes the original sender to be retried.

The appendix lists the source of the retry algorithm. Note that the signature of the retry algorithm is simply:

sender auto retry(sender auto s);

That is, it is not parameterized on an execution context on which to retry the operation. That is because we can assume the existence of a function on which schedules a sender for execution on a specified execution context:

sender auto on(sender auto s, scheduler auto sched);

Given these two functions, a user can simply do retry(on(s, sched)) to retry an operation on a particular execution context.

1.6.3 Toward an asynchronous STL

The algorithms then and retry are only two of many interesting Generic asynchronous algorithms that are expressible in terms of senders and receivers. Two other important algorithms are on and via, the former which schedules a sender for execution on a particular scheduler, and the latter which causes a sender’s continuations to be run on a particular scheduler. In this way, chains of asynchronous computation can be created that transition from one execution context to another.

Other important algorithms are when_all and when_any, encapsulating fork/join semantics. With these algorithms and others, entire DAGs of async computation can be created and executed. when_any can in turn be used to implement a generic timeout algorithm, together with a sender that sleeps for a duration and then sends a “done” signal, and so these algorithms compose.

In short, sender/receiver permits a rich set of Generic asynchronous algorithms to sit alongside Stepanov’s sequence algorithms in the STL. Asynchronous APIs that return senders would be usable with these Generic algorithms, increasing reusability. P1897 suggest an initial set of these algorithms.

1.7 Summary

We envision a future when C++ programmers can express asynchronous, parallel execution of work on diverse hardware resources through elegant standard interfaces. This proposal provides a foundation for flexible exeuction and is our initial step towards that goal. Executors represent hardware resources that execute work. Senders and receivers represent lazily-constructed asynchronous DAGs of work. These primitives empower programmers to control where and when work happens.

2 Proposed Wording

2.1 Execution Support Library

2.1.1 General

This Clause describes components supporting execution of function objects [function.objects].

(The following definition appears in working draft N4762 [thread.req.lockable.general])

An execution agent is an entity such as a thread that may perform work in parallel with other execution agents. [Note: Implementations or users may introduce other kinds of agents such as processes or thread-pool tasks. –end note] The calling agent is determined by context; e.g., the calling thread that contains the call, and so on.

An execution agent invokes a function object within an execution context such as the calling thread or thread-pool. An executor submits a function object to an execution context to be invoked by an execution agent within that execution context. [Note: Invocation of the function object may be inlined such as when the execution context is the calling thread, or may be scheduled such as when the execution context is a thread-pool with task scheduler. –end note] An executor may submit a function object with execution properties that specify how the submission and invocation of the function object interacts with the submitting thread and execution context, including forward progress guarantees [intro.progress].

For the intent of this library and extensions to this library, the lifetime of an execution agent begins before the function object is invoked and ends after this invocation completes, either normally or having thrown an exception.

2.1.2 Header <execution> synopsis

namespace std {
namespace execution {

  // Exception types:

  extern runtime_error const invocation-error; // exposition only
  struct receiver_invocation_error : runtime_error, nested_exception {
    receiver_invocation_error() noexcept
      : runtime_error(invocation-error), nested_exception() {}
  };

  // Invocable archetype

  using invocable_archetype = unspecified;

  // Customization points:

  inline namespace unspecified{
    inline constexpr unspecified set_value = unspecified;

    inline constexpr unspecified set_done = unspecified;

    inline constexpr unspecified set_error = unspecified;

    inline constexpr unspecified execute = unspecified;

    inline constexpr unspecified submit = unspecified;

    inline constexpr unspecified schedule = unspecified;

    inline constexpr unspecified bulk_execute = unspecified;
  }

  // Concepts:

  template<class T, class E = exception_ptr>
    concept receiver = see-below;

  template<class T, class... An>
    concept receiver_of = see-below;

  template<class S>
    concept sender = see-below;

  template<class S>
    concept typed_sender = see-below;

  template<class S, class R>
    concept sender_to = see-below;

  template<class S>
    concept scheduler = see-below;

  template<class E>
    concept executor = see-below;

  template<class E, class F>
    concept executor_of = see-below;

  // Sender and receiver utilities type
  class sink_receiver;

  template<class S> struct sender_traits;

  // Associated execution context property:

  struct context_t;

  constexpr context_t context;

  // Blocking properties:

  struct blocking_t;

  constexpr blocking_t blocking;

  // Properties to allow adaptation of blocking and directionality:

  struct blocking_adaptation_t;

  constexpr blocking_adaptation_t blocking_adaptation;

  // Properties to indicate if submitted tasks represent continuations:

  struct relationship_t;

  constexpr relationship_t relationship;

  // Properties to indicate likely task submission in the future:

  struct outstanding_work_t;

  constexpr outstanding_work_t outstanding_work;

  // Properties for bulk execution guarantees:

  struct bulk_guarantee_t;

  constexpr bulk_guarantee_t bulk_guarantee;

  // Properties for mapping of execution on to threads:

  struct mapping_t;

  constexpr mapping_t mapping;

  // Memory allocation properties:

  template <typename ProtoAllocator>
  struct allocator_t;

  constexpr allocator_t<void> allocator;

  // Executor type traits:

  template<class Executor> struct executor_shape;
  template<class Executor> struct executor_index;

  template<class Executor> using executor_shape_t = typename executor_shape<Executor>::type;
  template<class Executor> using executor_index_t = typename executor_index<Executor>::type;

  // Polymorphic executor support:

  class bad_executor;

  template <class... SupportableProperties> class any_executor;

  template<class Property> struct prefer_only;

} // namespace execution
} // namespace std

2.2 Requirements

2.2.1 ProtoAllocator requirements

A type A meets the ProtoAllocator requirements if A is CopyConstructible (C++Std [copyconstructible]), Destructible (C++Std [destructible]), and allocator_traits<A>::rebind_alloc<U> meets the allocator requirements (C++Std [allocator.requirements]), where U is an object type. [Note: For example, std::allocator<void> meets the proto-allocator requirements but not the allocator requirements. –end note] No comparison operator, copy operation, move operation, or swap operation on these types shall exit via an exception.

2.2.2 Invocable archetype

The name execution::invocable_archetype is an implementation-defined type that, along with any argument pack, models invocable.

A program that creates an instance of execution::invocable_archetype is ill-formed.

2.2.3 Customization points

2.2.3.1 execution::set_value

The name execution::set_value denotes a customization point object. The expression execution::set_value(R, Vs...) for some subexpressions R and Vs... is expression-equivalent to:

[Editorial note: We should probably define what “send the value(s) Vs... to the receiver R’s value channel” means more carefully. –end editorial note]

2.2.3.2 execution::set_done

The name execution::set_done denotes a customization point object. The expression execution::set_done(R) for some subexpression R is expression-equivalent to:

[Editorial note: We should probably define what “signal receiver R’s done channel” means more carefully. –end editorial note]

2.2.3.3 execution::set_error

The name execution::set_error denotes a customization point object. The expression execution::set_error(R, E) for some subexpressions R and E are expression-equivalent to:

[Editorial note: We should probably define what “send the error E to the receiver R’s error channel” means more carefully. –end editorial note]

2.2.3.4 execution::execute

The name execution::execute denotes a customization point object.

For some subexpressions e and f, let E be a type such that decltype((e)) is E and let F be a type such that decltype((f)) is F. The expression execution::execute(e, f) is ill-formed if F does not model invocable, or if E does not model either executor or sender. Otherwise, it is expression-equivalent to:

[Editorial note: We should probably define what “execute the function object F on the executor E” means more carefully. –end editorial note]

2.2.3.5 execution::submit

The name execution::submit denotes a customization point object.

A receiver object is submitted for execution via a sender by scheduling the eventual evaluation of one of the receiver’s value, error, or done channels.

For some subexpressions s and r, let S be a type such that decltype((s)) is S and let R be a type such that decltype((r)) is R. The expression execution::submit(s, r) is ill-formed if R does not model receiver, or if S does not model either sender or executor. Otherwise, it is expression-equivalent to:

[Editorial note: We should probably define what “submit the receiver object R via the sender S” means more carefully. –end editorial note]

2.2.3.6 execution::schedule

The name execution::schedule denotes a customization point object. The expression execution::schedule(S) for some subexpression S is expression-equivalent to:

2.2.3.7 execution::bulk_execute

The name execution::bulk_execute denotes a customization point object. If is_convertible_v<N, size_t> is true, then the expression execution::bulk_execute(S, F, N) for some subexpressions S, F, and N is expression-equivalent to:

[Editorial note: We should probably define what “execute N invocations of the function object F on the executor S in bulk” means more carefully. –end editorial note]

2.2.4 Concept receiver

XXX TODO The receiver concept…

// exposition only:
template<class T>
inline constexpr bool is-nothrow-move-or-copy-constructible =
  is_nothrow_move_constructible<T> ||
   copy_constructible<T>;

template<class T, class E = exception_ptr>
concept receiver =
  move_constructible<remove_cvref_t<T>> &&
  (is-nothrow-move-or-copy-constructible<remove_cvref_t<T>>) &&
  requires(T&& t, E&& e) {
    { execution::set_done((T&&) t) } noexcept;
    { execution::set_error((T&&) t, (E&&) e) } noexcept;
  };

2.2.5 Concept receiver_of

XXX TODO The receiver_of concept…

template<class T, class... An>
concept receiver_of =
  receiver<T> &&
  requires(T&& t, An&&... an) {
    execution::set_value((T&&) t, (An&&) an...);
  };

2.2.6 Concepts sender and sender_to

XXX TODO The sender and sender_to concepts…

Let sender-to-impl be the exposition-only concept

template<class S, class R>
concept sender-to-impl =
  requires(S&& s, R&& r) {
    execution::submit((S&&) s, (R&&) r);
  };

Then,

template<class S>
concept sender =
  move_constructible<remove_cvref_t<S>> &&
  sender-to-impl<S, sink_receiver>;

template<class S, class R>
concept sender_to =
  sender<S> &&
  receiver<R> &&
  sender-to-impl<S, R>;

None of these operations shall introduce data races as a result of concurrent invocations of those functions from different threads.

An sender type’s destructor shall not block pending completion of the submitted function objects. [Note: The ability to wait for completion of submitted function objects may be provided by the associated execution context. –end note]

In addition to the above requirements, types S and R model sender_to only if they satisfy the requirements from the Table below.

In the Table below,

Expression Return Type Operational semantics
execution::submit(s, r) void If execution::submit(s, r) exits without throwing an exception, then the implementation shall invoke exactly one of execution::set_value(rc, values...), execution::set_error(rc, error) or execution::set_done(rc) where rc is either r or an object moved from r. If any of the invocations of set_value or set_error exits via an exception then it is valid to call to either set_done(rc) or set_error(rc, E), where E is an exception_ptr pointing to an unspecified exception object.
submit may or may not block pending the successful transfer of execution to one of the three receiver operations.
The start of the invocation of submit strongly happens before [intro.multithread] the invocation of one of the three receiver operations.

2.2.7 Concept typed_sender

A sender is typed if it declares what types it sends through a receiver’s channels. The typed_sender concept is defined as:

template<template<template<class...> class Tuple, template<class...> class Variant> class>
struct has-value-types; // exposition only

template<template<class...> class Variant>
struct has-error-types; // exposition only

template<class S>
comcept has-sender-types = // exposition only
  requires {
    typename has-value-types<S::template value_types>;
    typename has-error-types<S::template error_types>;
    typename bool_constant<S::sends_done>;
  };

template<class S>
concept typed_sender =
  sender<S> &&
  has-sender-types<sender_traits<S>>;

2.2.8 Concept scheduler

XXX TODO The scheduler concept…

template<class S>
concept scheduler =
  copy_constructible<remove_cvref_t<S>> &&
  equality_comparable<remove_cvref_t<S>> &&
  requires(E&& e) {
    execution::schedule((S&&)s);
  }; // && sender<invoke_result_t<execution::schedule, S>>

None of a scheduler’s copy constructor, destructor, equality comparison, or swap operation shall exit via an exception.

None of these operations, nor an scheduler type’s schedule function, or associated query functions shall introduce data races as a result of concurrent invocations of those functions from different threads.

For any two (possibly const) values x1 and x2 of some scheduler type X, x1 == x2 shall return true only if x1.query(p) == x2.query(p) for every property p where both x1.query(p) and x2.query(p) are well-formed and result in a non-void type that is EqualityComparable (C++Std [equalitycomparable]). [Note: The above requirements imply that x1 == x2 returns true if x1 and x2 can be interchanged with identical effects. An scheduler may conceptually contain additional properties which are not exposed by a named property type that can be observed via execution::query; in this case, it is up to the concrete scheduler implementation to decide if these properties affect equality. Returning false does not necessarily imply that the effects are not identical. –end note]

An scheduler type’s destructor shall not block pending completion of any receivers submitted to the sender objects returned from schedule. [Note: The ability to wait for completion of submitted function objects may be provided by the execution context that produced the scheduler. –end note]

In addition to the above requirements, type S models scheduler only if it satisfies the requirements in the Table below.

In the Table below,

Expression Return Type Operational semantics
execution::schedule(s) N Evaluates execution::schedule(s) on the calling thread to create N.

execution::submit(N, r), for some receiver object r, is required to eagerly submit r for execution on an execution agent that s creates for it. Let rc be r or an object created by copy or move construction from r. The semantic constraints on the sender N returned from a scheduler s’s schedule function are as follows:

[ Note: The senders returned from a scheduler’s schedule function have wide discretion when deciding which of the three receiver functions to call upon submission. — end note ]

2.2.9 Concepts executor and executor_of

XXX TODO The executor and executor_of concepts…

Let executor-of-impl be the exposition-only concept

template<class E, class F>
concept executor-of-impl =
  invocable<F> &&
  is_nothrow_copy_constructible_v<E> &&
  is_nothrow_destructible_v<E> &&
  equality_comparable<E> &&
  requires(const E& e, F&& f) {
    execution::execute(e, (F&&)f);
  };

Then,

template<class E>
concept executor = executor-of-impl<E, execution::invocable_archetype>;

template<class E, class F>
concept executor_of = executor-of-impl<E, F>;

Neither of an executor’s equality comparison or swap operation shall exit via an exception.

None of an executor type’s copy constructor, destructor, equality comparison, swap function, execute function, or associated query functions shall introduce data races as a result of concurrent invocations of those functions from different threads.

For any two (possibly const) values x1 and x2 of some executor type X, x1 == x2 shall return true only if std::query(x1,p) == std::query(x2,p) for every property p where both std::query(x1,p) and std::query(x2,p) are well-formed and result in a non-void type that is equality_comparable (C++Std [equalitycomparable]). [Note: The above requirements imply that x1 == x2 returns true if x1 and x2 can be interchanged with identical effects. An executor may conceptually contain additional properties which are not exposed by a named property type that can be observed via std::query; in this case, it is up to the concrete executor implementation to decide if these properties affect equality. Returning false does not necessarily imply that the effects are not identical. –end note]

An executor type’s destructor shall not block pending completion of the submitted function objects. [Note: The ability to wait for completion of submitted function objects may be provided by the associated execution context. –end note]

In addition to the above requirements, types E and F model executor_of only if they satisfy the requirements of the Table below.

In the Table below,

Expression Return Type Operational semantics
execution::execute(e, f) void Evaluates DECAY_COPY(std::forward<F>(f)) on the calling thread to create cf that will be invoked at most once by an execution agent.
May block pending completion of this invocation.
Synchronizes with [intro.multithread] the invocation of f.
Shall not propagate any exception thrown by the function object or any other function submitted to the executor. [Note: The treatment of exceptions thrown by one-way submitted functions is implementation-defined. The forward progress guarantee of the associated execution agent(s) is implementation-defined. –end note.]

[Editorial note: The operational semantics of execution::execute should be specified with the execution::execute CPO rather than the executor concept. –end note.]

2.2.10 Sender and receiver traits

2.2.10.1 Class sink_receiver

XXX TODO The class sink_receiver

    class sink_receiver {
    public:
      void set_value(auto&&...) {}
      [[noreturn]] void set_error(auto&&) noexcept {
        std::terminate();
      }
      void set_done() noexcept {}
    };

2.2.10.2 Class template sender_traits

XXX TODO The class templatesender_traits

The class template sender_traits can be used to query information about a sender; in particular, what values and errors it sends through a receiver’s value and error channel, and whether or not it ever calls set_done on a receiver.

    template<class S>
    struct sender-traits-base {}; // exposition-only

    template<class S>
      requires (!same_as<S, remove_cvref_t<S>>)
    struct sender-traits-base : sender_traits<remove_cvref_t<S>> {};

    template<class S>
      requires same_as<S, remove_cvref_t<S>> &&
      sender<S> && has-sender-traits<S>
    struct sender-traits-base<S> {
      template<template<class...> class Tuple, template<class...> class Variant>
      using value_types = typename S::template value_types<Tuple, Variant>;

      template<template<class...> class Variant>
      using error_types = typename S::template error_types<Variant>;

      static constexpr bool sends_done = S::sends_done;
    };

    template<class S>
    struct sender_traits : sender-traits-base<S> {};

Because a sender may send one set of types or another to a receiver based on some runtime condition, sender_traits may provide a nested value_types template that is parameterized on a tuple-like class template and a variant-like class template that are used to hold the result.

[Example: If a sender type S sends types As... or Bs... to a receiver’s value channel, it may specialize sender_traits such that typename sender_traits<S>::value_types<tuple, variant> names the type variant<tuple<As...>, tuple<Bs...>>end example]

Because a sender may send one or another type of error types to a receiver, sender_traits may provide a nested error_types template that is parameterized on a variant-like class template that is used to hold the result.

[Example: If a sender type S sends error types exception_ptr or error_code to a receiver’s error channel, it may specialize sender_traits such that typename sender_traits<S>::error_types<variant> names the type variant<exception_ptr, error_code>end example]

A sender type can signal that it never calls set_done on a receiver by specializing sender_traits such that sender_traits<S>::sends_done is false; conversely, it may set sender_traits<S>::sends_done to true to indicate that it does call set_done on a receiver.

Users may specialize sender_traits on program-defined types.

2.2.11 Query-only properties

2.2.11.1 Associated execution context property

struct context_t
{
  template <class T>
    static constexpr bool is_applicable_property_v = executor<T>;

  static constexpr bool is_requirable = false;
  static constexpr bool is_preferable = false;

  using polymorphic_query_result_type = any;

  template<class Executor>
    static constexpr decltype(auto) static_query_v
      = Executor::query(context_t());
};

The context_t property can be used only with query, which returns the execution context associated with the executor.

The value returned from std::query(e, context_t), where e is an executor, shall not change between invocations.

2.2.11.2 Polymorphic executor wrappers

The any_executor class template provides polymorphic wrappers for executors.

In several places in this section the operation CONTAINS_PROPERTY(p, pn) is used. All such uses mean std::disjunction_v<std::is_same<p, pn>...>.

In several places in this section the operation FIND_CONVERTIBLE_PROPERTY(p, pn) is used. All such uses mean the first type P in the parameter pack pn for which std::is_convertible_v<p, P> is true. If no such type P exists, the operation FIND_CONVERTIBLE_PROPERTY(p, pn) is ill-formed.

template <class... SupportableProperties>
class any_executor
{
public:
  // construct / copy / destroy:

  any_executor() noexcept;
  any_executor(nullptr_t) noexcept;
  any_executor(const any_executor& e) noexcept;
  any_executor(any_executor&& e) noexcept;
  template<class... OtherSupportableProperties>
    any_executor(any_executor<OtherSupportableProperties...> e);
  template<class... OtherSupportableProperties>
    any_executor(any_executor<OtherSupportableProperties...> e) = delete;
  template<executor Executor>
    any_executor(Executor e);

  any_executor& operator=(const any_executor& e) noexcept;
  any_executor& operator=(any_executor&& e) noexcept;
  any_executor& operator=(nullptr_t) noexcept;
  template<executor Executor>
    any_executor& operator=(Executor e);

  ~any_executor();

  // any_executor modifiers:

  void swap(any_executor& other) noexcept;

  // any_executor operations:

  template <class Property>
  any_executor require(Property) const;

  template <class Property>
  typename Property::polymorphic_query_result_type query(Property) const;

  template<class Function>
    void execute(Function&& f) const;

  // any_executor capacity:

  explicit operator bool() const noexcept;

  // any_executor target access:

  const type_info& target_type() const noexcept;
  template<executor Executor> Executor* target() noexcept;
  template<executor Executor> const Executor* target() const noexcept;
};

// any_executor comparisons:

template <class... SupportableProperties>
bool operator==(const any_executor<SupportableProperties...>& a, const any_executor<SupportableProperties...>& b) noexcept;
template <class... SupportableProperties>
bool operator==(const any_executor<SupportableProperties...>& e, nullptr_t) noexcept;
template <class... SupportableProperties>
bool operator==(nullptr_t, const any_executor<SupportableProperties...>& e) noexcept;
template <class... SupportableProperties>
bool operator!=(const any_executor<SupportableProperties...>& a, const any_executor<SupportableProperties...>& b) noexcept;
template <class... SupportableProperties>
bool operator!=(const any_executor<SupportableProperties...>& e, nullptr_t) noexcept;
template <class... SupportableProperties>
bool operator!=(nullptr_t, const any_executor<SupportableProperties...>& e) noexcept;

// any_executor specialized algorithms:

template <class... SupportableProperties>
void swap(any_executor<SupportableProperties...>& a, any_executor<SupportableProperties...>& b) noexcept;

template <class Property, class... SupportableProperties>
any_executor prefer(const any_executor<SupportableProperties>& e, Property p);

The any_executor class satisfies the executor concept requirements.

[Note: To meet the noexcept requirements for executor copy constructors and move constructors, implementations may share a target between two or more any_executor objects. –end note]

Each property type in the SupportableProperties... pack shall provide a nested type polymorphic_query_result_type.

The target is the executor object that is held by the wrapper.

2.2.11.2.1 any_executor constructors
any_executor() noexcept;

Postconditions: !*this.

any_executor(nullptr_t) noexcept;

Postconditions: !*this.

any_executor(const any_executor& e) noexcept;

Postconditions: !*this if !e; otherwise, *this targets e.target() or a copy of e.target().

any_executor(any_executor&& e) noexcept;

Effects: If !e, *this has no target; otherwise, moves e.target() or move-constructs the target of e into the target of *this, leaving e in a valid state with an unspecified value.

template<class... OtherSupportableProperties>
  any_executor(any_executor<OtherSupportableProperties...> e);

Remarks: This function shall not participate in overload resolution unless: * CONTAINS_PROPERTY(p, OtherSupportableProperties) , where p is each property in SupportableProperties....

Effects: *this targets a copy of e initialized with std::move(e).

template<class... OtherSupportableProperties>
  any_executor(any_executor<OtherSupportableProperties...> e) = delete;

Remarks: This function shall not participate in overload resolution unless CONTAINS_PROPERTY(p, OtherSupportableProperties) is false for some property p in SupportableProperties....

template<executor Executor>
  any_executor(Executor e);

Remarks: This function shall not participate in overload resolution unless:

Effects: *this targets a copy of e.

2.2.11.2.2 any_executor assignment
any_executor& operator=(const any_executor& e) noexcept;

Effects: any_executor(e).swap(*this).

Returns: *this.

any_executor& operator=(any_executor&& e) noexcept;

Effects: Replaces the target of *this with the target of e, leaving e in a valid state with an unspecified value.

Returns: *this.

any_executor& operator=(nullptr_t) noexcept;

Effects: any_executor(nullptr).swap(*this).

Returns: *this.

template<executor Executor>
  any_executor& operator=(Executor e);

Requires: As for template<executor Executor> any_executor(Executor e).

Effects: any_executor(std::move(e)).swap(*this).

Returns: *this.

2.2.11.2.3 any_executor destructor
~any_executor();

Effects: If *this != nullptr, releases shared ownership of, or destroys, the target of *this.

2.2.11.2.4 any_executor modifiers
void swap(any_executor& other) noexcept;

Effects: Interchanges the targets of *this and other.

2.2.11.2.5 any_executor operations
template <class Property>
any_executor require(Property p) const;

Remarks: This function shall not participate in overload resolution unless FIND_CONVERTIBLE_PROPERTY(Property, SupportableProperties)::is_requirable is well-formed and has the value true.

Returns: A polymorphic wrapper whose target is the result of std::require(e, p), where e is the target object of *this.

template <class Property>
typename Property::polymorphic_query_result_type query(Property p) const;

Remarks: This function shall not participate in overload resolution unless FIND_CONVERTIBLE_PROPERTY(Property, SupportableProperties) is well-formed.

Returns: If std::query(e, p) is well-formed, static_cast<Property::polymorphic_query_result_type>(std::query(e, p)), where e is the target object of *this. Otherwise, Property::polymorphic_query_result_type{}.

template<class Function>
  void execute(Function&& f) const;

Effects: Performs execution::execute(e, f2), where:

2.2.11.2.6 any_executor capacity
explicit operator bool() const noexcept;

Returns: true if *this has a target, otherwise false.

2.2.11.2.7 any_executor target access
const type_info& target_type() const noexcept;

Returns: If *this has a target of type T, typeid(T); otherwise, typeid(void).

template<executor Executor> Executor* target() noexcept;
template<executor Executor> const Executor* target() const noexcept;

Returns: If target_type() == typeid(Executor) a pointer to the stored executor target; otherwise a null pointer value.

2.2.11.2.8 any_executor comparisons
template<class... SupportableProperties>
bool operator==(const any_executor<SupportableProperties...>& a, const any_executor<SupportableProperties...>& b) noexcept;

Returns:

template<class... SupportableProperties>
bool operator==(const any_executor<SupportableProperties...>& e, nullptr_t) noexcept;
template<class... SupportableProperties>
bool operator==(nullptr_t, const any_executor<SupportableProperties...>& e) noexcept;

Returns: !e.

template<class... SupportableProperties>
bool operator!=(const any_executor<SupportableProperties...>& a, const any_executor<SupportableProperties...>& b) noexcept;

Returns: !(a == b).

template<class... SupportableProperties>
bool operator!=(const any_executor<SupportableProperties...>& e, nullptr_t) noexcept;
template<class... SupportableProperties>
bool operator!=(nullptr_t, const any_executor<SupportableProperties...>& e) noexcept;

Returns: (bool) e.

2.2.11.2.9 any_executor specialized algorithms
template<class... SupportableProperties>
void swap(any_executor<SupportableProperties...>& a, any_executor<SupportableProperties...>& b) noexcept;

Effects: a.swap(b).

template <class Property, class... SupportableProperties>
any_executor prefer(const any_executor<SupportableProperties...>& e, Property p);

Remarks: This function shall not participate in overload resolution unless FIND_CONVERTIBLE_PROPERTY(Property, SupportableProperties)::is_preferable is well-formed and has the value true.

Returns: A polymorphic wrapper whose target is the result of std::prefer(e, p), where e is the target object of *this.

2.2.12 Behavioral properties

Behavioral properties define a set of mutually-exclusive nested properties describing executor behavior.

Unless otherwise specified, behavioral property types S, their nested property types S::Ni, and nested property objects S::ni conform to the following specification:

struct S
{
  template <class T>
    static constexpr bool is_applicable_property_v = executor<T>;

  static constexpr bool is_requirable = false;
  static constexpr bool is_preferable = false;
  using polymorphic_query_result_type = S;

  template<class Executor>
    static constexpr auto static_query_v
      = see-below;

  template<class Executor>
  friend constexpr S query(const Executor& ex, const Property& p) noexcept(see-below);

  friend constexpr bool operator==(const S& a, const S& b);
  friend constexpr bool operator!=(const S& a, const S& b) { return !operator==(a, b); }

  constexpr S();

  struct N1
  {
    static constexpr bool is_requirable = true;
    static constexpr bool is_preferable = true;
    using polymorphic_query_result_type = S;

    template<class Executor>
      static constexpr auto static_query_v
        = see-below;

    static constexpr S value() { return S(N1()); }
  };

  static constexpr N1 n1;

  constexpr S(const N1);

  ...

  struct NN
  {
    static constexpr bool is_requirable = true;
    static constexpr bool is_preferable = true;
    using polymorphic_query_result_type = S;

    template<class Executor>
      static constexpr auto static_query_v
        = see-below;

    static constexpr S value() { return S(NN()); }
  };

  static constexpr NN nN;

  constexpr S(const NN);
};

Queries for the value of an executor’s behavioral property shall not change between invocations unless the executor is assigned another executor with a different value of that behavioral property.

S() and S(S::Ei()) are all distinct values of S. [Note: This means they compare unequal. –end note.]

The value returned from std::query(e1, p1) and a subsequent invocation std::query(e1, p1), where

shall compare equal unless

The value of the expression S::N1::static_query_v<Executor> is

[Note: These rules automatically enable the S::N1 property by default for executors which do not provide a query function for properties S::Ni. –end note]

The value of the expression S::Ni::static_query_v<Executor>, for all 1 < i <= N, is

The value of the expression S::static_query_v<Executor> is

[Note: These rules automatically enable the S::N1 property by default for executors which do not provide a query function for properties S or S::Ni. –end note]

Let k be the least value of i for which can_query_v<Executor,S::Ni> is true, if such a value of i exists.

template<class Executor>
  friend constexpr S query(const Executor& ex, const Property& p) noexcept(noexcept(std::query(ex, std::declval<const S::Nk>())));

Returns: std::query(ex, S::Nk()).

Remarks: This function shall not participate in overload resolution unless is_same_v<Property,S> && can_query_v<Executor,S::Ni> is true for at least one S::Ni`.

bool operator==(const S& a, const S& b);

Returns: true if a and b were constructed from the same constructor; false, otherwise.

2.2.12.1 Blocking properties

The blocking_t property describes what guarantees executors provide about the blocking behavior of their execution functions.

blocking_t provides nested property types and objects as described below.

Nested Property Type Nested Property Object Name Requirements
blocking_t::possibly_t blocking.possibly Invocation of an executor’s execution function may block pending completion of one or more invocations of the submitted function object.
blocking_t::always_t blocking.always Invocation of an executor’s execution function shall block until completion of all invocations of submitted function object.
blocking_t::never_t blocking.never Invocation of an executor’s execution function shall not block pending completion of the invocations of the submitted function object.
2.2.12.1.1 blocking_t::always_t customization points

In addition to conforming to the above specification, the blocking_t::always_t property provides the following customization:

struct always_t
{
  static constexpr bool is_requirable = true;
  static constexpr bool is_preferable = false;

  template <class T>
    static constexpr bool is_applicable_property_v = executor<T>;

  template<class Executor>
    friend see-below require(Executor ex, blocking_t::always_t);
};

If the executor has the blocking_adaptation_t::allowed_t property, this customization uses an adapter to implement the blocking_t::always_t property.

template<class Executor>
  friend see-below require(Executor ex, blocking_t::always_t);

Returns: A value e1 of type E1 that holds a copy of ex. E1 provides an overload of require such that e1.require(blocking.always) returns a copy of e1, an overload of query such that std::query(e1,blocking) returns blocking.always, and functions execute and bulk_execute shall block the calling thread until the submitted functions have finished execution. e1 has the same executor properties as ex, except for the addition of the blocking_t::always_t property, and removal of blocking_t::never_t and blocking_t::possibly_t properties if present.

Remarks: This function shall not participate in overload resolution unless blocking_adaptation_t::static_query_v<Executor> is blocking_adaptation.allowed.

2.2.12.2 Properties to indicate if blocking and directionality may be adapted

The blocking_adaptation_t property allows or disallows blocking or directionality adaptation via std::require.

blocking_adaptation_t provides nested property types and objects as described below.

Nested Property Type Nested Property Object Name Requirements
blocking_adaptation_t::disallowed_t blocking_adaptation.disallowed The require customization point may not adapt the executor to add the blocking_t::always_t property.
blocking_adaptation_t::allowed_t blocking_adaptation.allowed The require customization point may adapt the executor to add the blocking_t::always_t property.
2.2.12.2.1 blocking_adaptation_t::allowed_t customization points

In addition to conforming to the above specification, the blocking_adaptation_t::allowed_t property provides the following customization:

struct allowed_t
{
  static constexpr bool is_requirable = true;
  static constexpr bool is_preferable = false;

  template <class T>
    static constexpr bool is_applicable_property_v = executor<T>;

  template<class Executor>
    friend see-below require(Executor ex, blocking_adaptation_t::allowed_t);
};

This customization uses an adapter to implement the blocking_adaptation_t::allowed_t property.

template<class Executor>
  friend see-below require(Executor ex, blocking_adaptation_t::allowed_t);

Returns: A value e1 of type E1 that holds a copy of ex. In addition, blocking_adaptation_t::static_query_v<E1> is blocking_adaptation.allowed, and e1.require(blocking_adaptation.disallowed) yields a copy of ex. e1 has the same executor properties as ex, except for the addition of the blocking_adaptation_t::allowed_t property.

2.2.12.3 Properties to indicate if submitted tasks represent continuations

The relationship_t property allows users of executors to indicate that submitted tasks represent continuations.

relationship_t provides nested property types and objects as indicated below.

Nested Property Type Nested Property Object Name Requirements
relationship_t::fork_t relationship.fork Function objects submitted through the executor do not represent continuations of the caller.
relationship_t::continuation_t relationship.continuation Function objects submitted through the executor represent continuations of the caller. Invocation of the submitted function object may be deferred until the caller completes.

2.2.12.4 Properties to indicate likely task submission in the future

The outstanding_work_t property allows users of executors to indicate that task submission is likely in the future.

outstanding_work_t provides nested property types and objects as indicated below.

Nested Property Type Nested Property Object Name Requirements
outstanding_work_t::untracked_t outstanding_work.untracked The existence of the executor object does not indicate any likely future submission of a function object.
outstanding_work_t::tracked_t outstanding_work.tracked The existence of the executor object represents an indication of likely future submission of a function object. The executor or its associated execution context may choose to maintain execution resources in anticipation of this submission.

[Note: The outstanding_work_t::tracked_t and outstanding_work_t::untracked_t properties are used to communicate to the associated execution context intended future work submission on the executor. The intended effect of the properties is the behavior of execution context’s facilities for awaiting outstanding work; specifically whether it considers the existance of the executor object with the outstanding_work_t::tracked_t property enabled outstanding work when deciding what to wait on. However this will be largely defined by the execution context implementation. It is intended that the execution context will define its wait facilities and on-destruction behaviour and provide an interface for querying this. An initial work towards this is included in P0737r0. –end note]

2.2.12.5 Properties for bulk execution guarantees

Bulk execution guarantee properties communicate the forward progress and ordering guarantees of execution agents associated with the bulk execution.

bulk_guarantee_t provides nested property types and objects as indicated below.

Nested Property Type Nested Property Object Name Requirements
bulk_guarantee_t::unsequenced_t bulk_guarantee.unsequenced Execution agents within the same bulk execution may be parallelized and vectorized.
bulk_guarantee_t::sequenced_t bulk_guarantee.sequenced Execution agents within the same bulk execution may not be parallelized.
bulk_guarantee_t::parallel_t bulk_guarantee.parallel Execution agents within the same bulk execution may be parallelized.

Execution agents associated with the bulk_guarantee_t::unsequenced_t property may invoke the function object in an unordered fashion. Any such invocations in the same thread of execution are unsequenced with respect to each other. [Note: This means that multiple execution agents may be interleaved on a single thread of execution, which overrides the usual guarantee from [intro.execution] that function executions do not interleave with one another. –end note]

Execution agents associated with the bulk_guarantee_t::sequenced_t property invoke the function object in sequence in lexicographic order of their indices.

Execution agents associated with the bulk_guarantee_t::parallel_t property invoke the function object with a parallel forward progress guarantee. Any such invocations in the same thread of execution are indeterminately sequenced with respect to each other. [Note: It is the caller’s responsibility to ensure that the invocation does not introduce data races or deadlocks. –end note]

[Editorial note: The descriptions of these properties were ported from [algorithms.parallel.user]. The intention is that a future standard will specify execution policy behavior in terms of the fundamental properties of their associated executors. We did not include the accompanying code examples from [algorithms.parallel.user] because the examples seem easier to understand when illustrated by std::for_each. –end editorial note]

2.2.12.6 Properties for mapping of execution on to threads

The mapping_t property describes what guarantees executors provide about the mapping of execution agents onto threads of execution.

mapping_t provides nested property types and objects as indicated below.

Nested Property Type Nested Property Object Name Requirements
mapping_t::thread_t mapping.thread Execution agents are mapped onto threads of execution.
mapping_t::new_thread_t mapping.new_thread Each execution agent is mapped onto a new thread of execution.
mapping_t::other_t mapping.other Mapping of each execution agent is implementation-defined.

[Note: A mapping of an execution agent onto a thread of execution implies the execution agent runs as-if on a std::thread. Therefore, the facilities provided by std::thread, such as thread-local storage, are available. mapping_t::new_thread_t provides stronger guarantees, in particular that thread-local storage will not be shared between execution agents. –end note]

2.2.13 Properties for customizing memory allocation

template <typename ProtoAllocator>
struct allocator_t;

The allocator_t property conforms to the following specification:

template <typename ProtoAllocator>
struct allocator_t
{
    template <class T>
      static constexpr bool is_applicable_property_v = executor<T>;

    static constexpr bool is_requirable = true;
    static constexpr bool is_preferable = true;

    template<class Executor>
    static constexpr auto static_query_v
      = Executor::query(allocator_t);

    template <typename OtherProtoAllocator>
    allocator_t<OtherProtoAllocator> operator()(const OtherProtoAllocator &a) const;

    static constexpr ProtoAllocator value() const;

private:
    ProtoAllocator a_; // exposition only
};
Property Notes Requirements
allocator_t<ProtoAllocator> Result of allocator_t<void>::operator(OtherProtoAllocator). The executor shall use the encapsulated allocator to allocate any memory required to store the submitted function object.
allocator_t<void> Specialisation of allocator_t<ProtoAllocator>. The executor shall use an implementation defined default allocator to allocate any memory required to store the submitted function object.

If the expression std::query(E, P) is well formed, where P is an object of type allocator_t<ProtoAllocator>, then: * the type of the expression std::query(E, P) shall satisfy the ProtoAllocator requirements; * the result of the expression std::query(E, P) shall be the allocator currently established in the executor E; and * the expression std::query(E, allocator_t<void>{}) shall also be well formed and have the same result as std::query(E, P).

2.2.13.1 allocator_t members

template <typename OtherProtoAllocator>
allocator_t<OtherProtoAllocator> operator()(const OtherProtoAllocator &a) const;

Returns: An allocator object whose exposition-only member a_ is initialized as a_(a).

Remarks: This function shall not participate in overload resolution unless ProtoAllocator is void.

[Note: It is permitted for a to be an executor’s implementation-defined default allocator and, if so, the default allocator may also be established within an executor by passing the result of this function to require. –end note]

static constexpr ProtoAllocator value() const;

Returns: The exposition-only member a_.

Remarks: This function shall not participate in overload resolution unless ProtoAllocator is not void.

2.3 Executor type traits

2.3.1 Associated shape type

template<class Executor>
struct executor_shape
{
  private:
    // exposition only
    template<class T>
    using helper = typename T::shape_type;

  public:
    using type = std::experimental::detected_or_t<
      size_t, helper, decltype(std::require(declval<const Executor&>(), execution::bulk))
    >;

    // exposition only
    static_assert(std::is_integral_v<type>, "shape type must be an integral type");
};

2.3.2 Associated index type

template<class Executor>
struct executor_index
{
  private:
    // exposition only
    template<class T>
    using helper = typename T::index_type;

  public:
    using type = std::experimental::detected_or_t<
      executor_shape_t<Executor>, helper, decltype(std::require(declval<const Executor&>(), execution::bulk))
    >;

    // exposition only
    static_assert(std::is_integral_v<type>, "index type must be an integral type");
};

2.4 Polymorphic executor support

2.4.1 Class bad_executor

An exception of type bad_executor is thrown by polymorphic executor member functions execute and bulk_execute when the executor object has no target.

class bad_executor : public exception
{
public:
  // constructor:
  bad_executor() noexcept;
};

2.4.1.1 bad_executor constructors

bad_executor() noexcept;

Effects: Constructs a bad_executor object.

Postconditions: what() returns an implementation-defined NTBS.

2.4.2 Struct prefer_only

The prefer_only struct is a property adapter that disables the is_requirable value.

[Example:

Consider a generic function that performs some task immediately if it can, and otherwise asynchronously in the background.

template<class Executor, class Callback>
void do_async_work(
    Executor ex,
    Callback callback)
{
  if (try_work() == done)
  {
    // Work completed immediately, invoke callback.
    std::require(ex,
        execution::single,
        execution::oneway,
      ).execute(callback);
  }
  else
  {
    // Perform work in background. Track outstanding work.
    start_background_work(
        std::prefer(ex,
          execution::outstanding_work.tracked),
        callback);
  }
}

This function can be used with an inline executor which is defined as follows:

struct inline_executor
{
  constexpr bool operator==(const inline_executor&) const noexcept
  {
    return true;
  }

  constexpr bool operator!=(const inline_executor&) const noexcept
  {
    return false;
  }

  template<class Function> void execute(Function f) const noexcept
  {
    f();
  }
};

as, in the case of an unsupported property, invocation of std::prefer will fall back to an identity operation.

The polymorphic executor wrapper should be able to simply swap in, so that we could change do_async_work to the non-template function:

void do_async_work(
    executor<
      execution::single,
      execution::oneway,
      execution::outstanding_work_t::tracked_t> ex,
    std::function<void()> callback)
{
  if (try_work() == done)
  {
    // Work completed immediately, invoke callback.
    std::require(ex,
        execution::single,
        execution::oneway,
      ).execute(callback);
  }
  else
  {
    // Perform work in background. Track outstanding work.
    start_background_work(
        std::prefer(ex,
          execution::outstanding_work.tracked),
        callback);
  }
}

with no change in behavior or semantics.

However, if we simply specify execution::outstanding_work.tracked in the executor template parameter list, we will get a compile error due to the executor template not knowing that execution::outstanding_work.tracked is intended for use with prefer only. At the point of construction from an inline_executor called ex, executor will try to instantiate implementation templates that perform the ill-formed std::require(ex, execution::outstanding_work.tracked).

The prefer_only adapter addresses this by turning off the is_requirable attribute for a specific property. It would be used in the above example as follows:

void do_async_work(
    executor<
      execution::single,
      execution::oneway,
      prefer_only<execution::outstanding_work_t::tracked_t>> ex,
    std::function<void()> callback)
{
  ...
}

– end example]

template<class InnerProperty>
struct prefer_only
{
  InnerProperty property;

  static constexpr bool is_requirable = false;
  static constexpr bool is_preferable = InnerProperty::is_preferable;

  using polymorphic_query_result_type = see-below; // not always defined

  template<class Executor>
    static constexpr auto static_query_v = see-below; // not always defined

  constexpr prefer_only(const InnerProperty& p);

  constexpr auto value() const
    noexcept(noexcept(std::declval<const InnerProperty>().value()))
      -> decltype(std::declval<const InnerProperty>().value());

  template<class Executor, class Property>
  friend auto prefer(Executor ex, const Property& p)
    noexcept(noexcept(std::prefer(std::move(ex), std::declval<const InnerProperty>())))
      -> decltype(std::prefer(std::move(ex), std::declval<const InnerProperty>()));

  template<class Executor, class Property>
  friend constexpr auto query(const Executor& ex, const Property& p)
    noexcept(noexcept(std::query(ex, std::declval<const InnerProperty>())))
      -> decltype(std::query(ex, std::declval<const InnerProperty>()));
};

If InnerProperty::polymorphic_query_result_type is valid and denotes a type, the template instantiation prefer_only<InnerProperty> defines a nested type polymorphic_query_result_type as a synonym for InnerProperty::polymorphic_query_result_type.

If InnerProperty::static_query_v is a variable template and InnerProperty::static_query_v<E> is well formed for some executor type E, the template instantiation prefer_only<InnerProperty> defines a nested variable template static_query_v as a synonym for InnerProperty::static_query_v.

constexpr prefer_only(const InnerProperty& p);

Effects: Initializes property with p.

constexpr auto value() const
  noexcept(noexcept(std::declval<const InnerProperty>().value()))
    -> decltype(std::declval<const InnerProperty>().value());

Returns: property.value().

Remarks: Shall not participate in overload resolution unless the expression property.value() is well-formed.

template<class Executor, class Property>
friend auto prefer(Executor ex, const Property& p)
  noexcept(noexcept(std::prefer(std::move(ex), std::declval<const InnerProperty>())))
    -> decltype(std::prefer(std::move(ex), std::declval<const InnerProperty>()));

Returns: std::prefer(std::move(ex), p.property).

Remarks: Shall not participate in overload resolution unless std::is_same_v<Property, prefer_only> is true, and the expression std::prefer(std::move(ex), p.property) is well-formed.

template<class Executor, class Property>
friend constexpr auto query(const Executor& ex, const Property& p)
  noexcept(noexcept(std::query(ex, std::declval<const InnerProperty>())))
    -> decltype(std::query(ex, std::declval<const InnerProperty>()));

Returns: std::query(ex, p.property).

Remarks: Shall not participate in overload resolution unless std::is_same_v<Property, prefer_only> is true, and the expression std::query(ex, p.property) is well-formed.

2.5 Thread pools

Thread pools manage execution agents which run on threads without incurring the overhead of thread creation and destruction whenever such agents are needed.

2.5.1 Header <thread_pool> synopsis

namespace std {

  class static_thread_pool;

} // namespace std

2.5.2 Class static_thread_pool

static_thread_pool is a statically-sized thread pool which may be explicitly grown via thread attachment. The static_thread_pool is expected to be created with the use case clearly in mind with the number of threads known by the creator. As a result, no default constructor is considered correct for arbitrary use cases and static_thread_pool does not support any form of automatic resizing.

static_thread_pool presents an effectively unbounded input queue and the execution functions of static_thread_pool’s associated executors do not block on this input queue.

[Note: Because static_thread_pool represents work as parallel execution agents, situations which require concurrent execution properties are not guaranteed correctness. –end note.]

class static_thread_pool
{
  public:
    using scheduler_type = see-below;
    using executor_type = see-below;
    
    // construction/destruction
    explicit static_thread_pool(std::size_t num_threads);
    
    // nocopy
    static_thread_pool(const static_thread_pool&) = delete;
    static_thread_pool& operator=(const static_thread_pool&) = delete;

    // stop accepting incoming work and wait for work to drain
    ~static_thread_pool();

    // attach current thread to the thread pools list of worker threads
    void attach();

    // signal all work to complete
    void stop();

    // wait for all threads in the thread pool to complete
    void wait();

    // placeholder for a general approach to getting schedulers from 
    // standard contexts.
    scheduler_type scheduler() noexcept;

    // placeholder for a general approach to getting executors from 
    // standard contexts.
    executor_type executor() noexcept;
};

For an object of type static_thread_pool, outstanding work is defined as the sum of:

The static_thread_pool member functions scheduler, executor, attach, wait, and stop, and the associated schedulers’, senders` and executors’ copy constructors and member functions, do not introduce data races as a result of concurrent invocations of those functions from different threads of execution.

A static_thread_pool’s threads run execution agents with forward progress guarantee delegation. [Note: Forward progress is delegated to an execution agent for its lifetime. Because static_thread_pool guarantees only parallel forward progress to running execution agents; i.e., execution agents which have run the first step of the function object. –end note]

2.5.2.1 Types

using scheduler_type = see-below;

A scheduler type conforming to the specification for static_thread_pool scheduler types described below.

using executor_type = see-below;

An executor type conforming to the specification for static_thread_pool executor types described below.

2.5.2.2 Construction and destruction

static_thread_pool(std::size_t num_threads);

Effects: Constructs a static_thread_pool object with num_threads threads of execution, as if by creating objects of type std::thread.

~static_thread_pool();

Effects: Destroys an object of class static_thread_pool. Performs stop() followed by wait().

2.5.2.3 Worker management

void attach();

Effects: Adds the calling thread to the pool such that this thread is used to execute submitted function objects. [Note: Threads created during thread pool construction, or previously attached to the pool, will continue to be used for function object execution. –end note] Blocks the calling thread until signalled to complete by stop() or wait(), and then blocks until all the threads created during static_thread_pool object construction have completed. (NAMING: a possible alternate name for this function is join().)

void stop();

Effects: Signals the threads in the pool to complete as soon as possible. If a thread is currently executing a function object, the thread will exit only after completion of that function object. Invocation of stop() returns without waiting for the threads to complete. Subsequent invocations to attach complete immediately.

void wait();

Effects: If not already stopped, signals the threads in the pool to complete once the outstanding work is 0. Blocks the calling thread (C++Std [defns.block]) until all threads in the pool have completed, without executing submitted function objects in the calling thread. Subsequent invocations of attach() complete immediately.

Synchronization: The completion of each thread in the pool synchronizes with (C++Std [intro.multithread]) the corresponding successful wait() return.

2.5.2.4 Scheduler creation

scheduler_type scheduler() noexcept;

Returns: A scheduler that may be used to create sender objects that may be used to submit receiver objects to the thread pool. The returned scheduler has the following properties already established:

2.5.2.5 Executor creation

executor_type executor() noexcept;

Returns: An executor that may be used to submit function objects to the thread pool. The returned executor has the following properties already established:

2.5.3 static_thread_pool scheduler types

All scheduler types accessible through static_thread_pool::scheduler(), and subsequent invocations of the member function require, conform to the following specification.

class C
{
  public:

    // types:

    using sender_type = see-below;

    // construct / copy / destroy:

    C(const C& other) noexcept;
    C(C&& other) noexcept;

    C& operator=(const C& other) noexcept;
    C& operator=(C&& other) noexcept;

    // scheduler operations:

    see-below require(const execution::allocator_t<void>& a) const;
    template<class ProtoAllocator>
    see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

    see-below query(execution::context_t) const noexcept;
    see-below query(execution::allocator_t<void>) const noexcept;
    template<class ProtoAllocator>
    see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

    bool running_in_this_thread() const noexcept;
};

bool operator==(const C& a, const C& b) noexcept;
bool operator!=(const C& a, const C& b) noexcept;

Objects of type C are associated with a static_thread_pool.

2.5.3.1 Constructors

C(const C& other) noexcept;

Postconditions: *this == other.

C(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

2.5.3.2 Assignment

C& operator=(const C& other) noexcept;

Postconditions: *this == other.

Returns: *this.

C& operator=(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

Returns: *this.

2.5.3.3 Operations

see-below require(const execution::allocator_t<void>& a) const;

Returns: require(execution::allocator(x)), where x is an implementation-defined default allocator.

template<class ProtoAllocator>
  see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

Returns: An scheduler object of an unspecified type conforming to these specifications, associated with the same thread pool as *this, with the execution::allocator_t<ProtoAllocator> property established such that allocation and deallocation associated with function submission will be performed using a copy of a.alloc. All other properties of the returned scheduler object are identical to those of *this.

static_thread_pool& query(execution::context_t) const noexcept;

Returns: A reference to the associated static_thread_pool object.

see-below query(execution::allocator_t<void>) const noexcept;
see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

Returns: The allocator object associated with the executor, with type and value as either previously established by the execution::allocator_t<ProtoAllocator> property or the implementation defined default allocator established by the execution::allocator_t<void> property.

bool running_in_this_thread() const noexcept;

Returns: true if the current thread of execution is a thread that was created by or attached to the associated static_thread_pool object.

2.5.3.4 Comparisons

bool operator==(const C& a, const C& b) noexcept;

Returns: true if &a.query(execution::context) == &b.query(execution::context) and a and b have identical properties, otherwise false.

bool operator!=(const C& a, const C& b) noexcept;

Returns: !(a == b).

2.5.3.5 static_thread_pool scheduler functions

In addition to conforming to the above specification, static_thread_pool schedulers shall conform to the following specification.

class C
{
  public:
    sender_type schedule() noexcept;
};

C is a type satisfying the scheduler requirements.

2.5.3.6 Sender creation

  sender_type schedule() noexcept;

Returns: A sender that may be used to submit function objects to the thread pool. The returned sender has the following properties already established:

2.5.4 static_thread_pool sender types

All sender types accessible through static_thread_pool::scheduler().schedule(), and subsequent invocations of the member function require, conform to the following specification.

class C
{
  public:

    // construct / copy / destroy:

    C(const C& other) noexcept;
    C(C&& other) noexcept;

    C& operator=(const C& other) noexcept;
    C& operator=(C&& other) noexcept;

    // sender operations:

    see-below require(execution::blocking_t::never_t) const;
    see-below require(execution::blocking_t::possibly_t) const;
    see-below require(execution::blocking_t::always_t) const;
    see-below require(execution::relationship_t::continuation_t) const;
    see-below require(execution::relationship_t::fork_t) const;
    see-below require(execution::outstanding_work_t::tracked_t) const;
    see-below require(execution::outstanding_work_t::untracked_t) const;
    see-below require(const execution::allocator_t<void>& a) const;
    template<class ProtoAllocator>
    see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

    static constexpr execution::bulk_guarantee_t query(execution::bulk_guarantee_t::parallel_t) const;
    static constexpr execution::mapping_t query(execution::mapping_t::thread_t) const;
    execution::blocking_t query(execution::blocking_t) const;
    execution::relationship_t query(execution::relationship_t) const;
    execution::outstanding_work_t query(execution::outstanding_work_t) const;
    see-below query(execution::context_t) const noexcept;
    see-below query(execution::allocator_t<void>) const noexcept;
    template<class ProtoAllocator>
    see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

    bool running_in_this_thread() const noexcept;
};

bool operator==(const C& a, const C& b) noexcept;
bool operator!=(const C& a, const C& b) noexcept;

Objects of type C are associated with a static_thread_pool.

2.5.4.1 Constructors

C(const C& other) noexcept;

Postconditions: *this == other.

C(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

2.5.4.2 Assignment

C& operator=(const C& other) noexcept;

Postconditions: *this == other.

Returns: *this.

C& operator=(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

Returns: *this.

2.5.4.3 Operations

see-below require(execution::blocking_t::never_t) const;
see-below require(execution::blocking_t::possibly_t) const;
see-below require(execution::blocking_t::always_t) const;
see-below require(execution::relationship_t::continuation_t) const;
see-below require(execution::relationship_t::fork_t) const;
see-below require(execution::outstanding_work_t::tracked_t) const;
see-below require(execution::outstanding_work_t::untracked_t) const;

Returns: An sender object of an unspecified type conforming to these specifications, associated with the same thread pool as *this, and having the requested property established. When the requested property is part of a group that is defined as a mutually exclusive set, any other properties in the group are removed from the returned sender object. All other properties of the returned sender object are identical to those of *this.

see-below require(const execution::allocator_t<void>& a) const;

Returns: require(execution::allocator(x)), where x is an implementation-defined default allocator.

template<class ProtoAllocator>
  see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

Returns: An sender object of an unspecified type conforming to these specifications, associated with the same thread pool as *this, with the execution::allocator_t<ProtoAllocator> property established such that allocation and deallocation associated with function submission will be performed using a copy of a.alloc. All other properties of the returned sender object are identical to those of *this.

static constexpr execution::bulk_guarantee_t query(execution::bulk_guarantee_t) const;

Returns: execution::bulk_guarantee.parallel

static constexpr execution::mapping_t query(execution::mapping_t) const;

Returns: execution::mapping.thread.

execution::blocking_t query(execution::blocking_t) const;
execution::relationship_t query(execution::relationship_t) const;
execution::outstanding_work_t query(execution::outstanding_work_t) const;

Returns: The value of the given property of *this.

static_thread_pool& query(execution::context_t) const noexcept;

Returns: A reference to the associated static_thread_pool object.

see-below query(execution::allocator_t<void>) const noexcept;
see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

Returns: The allocator object associated with the sender, with type and value as either previously established by the execution::allocator_t<ProtoAllocator> property or the implementation defined default allocator established by the execution::allocator_t<void> property.

bool running_in_this_thread() const noexcept;

Returns: true if the current thread of execution is a thread that was created by or attached to the associated static_thread_pool object.

2.5.4.4 Comparisons

bool operator==(const C& a, const C& b) noexcept;

Returns: true if &a.query(execution::context) == &b.query(execution::context) and a and b have identical properties, otherwise false.

bool operator!=(const C& a, const C& b) noexcept;

Returns: !(a == b).

2.5.4.5 static_thread_pool sender execution functions

In addition to conforming to the above specification, static_thread_pool executors shall conform to the following specification.

class C
{
  public:
    template<class Receiver>
      void submit(Receiver&& r) const;
};

C is a type satisfying the sender requirements.

template<class Receiver>
  void submit(Receiver&& r) const;

Effects: Submits the receiver r for execution on the static_thread_pool according to the the properties established for *this. let e be an object of type exception_ptr, then static_thread_pool will evaluate one of set_value(r), set_error(r, e), or set_done(r).

2.5.5 static_thread_pool executor types

All executor types accessible through static_thread_pool::executor(), and subsequent invocations of the member function require, conform to the following specification.

class C
{
  public:

    // types:

    using shape_type = size_t;
    using index_type = size_t;

    // construct / copy / destroy:

    C(const C& other) noexcept;
    C(C&& other) noexcept;

    C& operator=(const C& other) noexcept;
    C& operator=(C&& other) noexcept;

    // executor operations:

    see-below require(execution::blocking_t::never_t) const;
    see-below require(execution::blocking_t::possibly_t) const;
    see-below require(execution::blocking_t::always_t) const;
    see-below require(execution::relationship_t::continuation_t) const;
    see-below require(execution::relationship_t::fork_t) const;
    see-below require(execution::outstanding_work_t::tracked_t) const;
    see-below require(execution::outstanding_work_t::untracked_t) const;
    see-below require(const execution::allocator_t<void>& a) const;
    template<class ProtoAllocator>
    see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

    static constexpr execution::bulk_guarantee_t query(execution::bulk_guarantee_t::parallel_t) const;
    static constexpr execution::mapping_t query(execution::mapping_t::thread_t) const;
    execution::blocking_t query(execution::blocking_t) const;
    execution::relationship_t query(execution::relationship_t) const;
    execution::outstanding_work_t query(execution::outstanding_work_t) const;
    see-below query(execution::context_t) const noexcept;
    see-below query(execution::allocator_t<void>) const noexcept;
    template<class ProtoAllocator>
    see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

    bool running_in_this_thread() const noexcept;
};

bool operator==(const C& a, const C& b) noexcept;
bool operator!=(const C& a, const C& b) noexcept;

Objects of type C are associated with a static_thread_pool.

2.5.5.1 Constructors

C(const C& other) noexcept;

Postconditions: *this == other.

C(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

2.5.5.2 Assignment

C& operator=(const C& other) noexcept;

Postconditions: *this == other.

Returns: *this.

C& operator=(C&& other) noexcept;

Postconditions: *this is equal to the prior value of other.

Returns: *this.

2.5.5.3 Operations

see-below require(execution::blocking_t::never_t) const;
see-below require(execution::blocking_t::possibly_t) const;
see-below require(execution::blocking_t::always_t) const;
see-below require(execution::relationship_t::continuation_t) const;
see-below require(execution::relationship_t::fork_t) const;
see-below require(execution::outstanding_work_t::tracked_t) const;
see-below require(execution::outstanding_work_t::untracked_t) const;

Returns: An executor object of an unspecified type conforming to these specifications, associated with the same thread pool as *this, and having the requested property established. When the requested property is part of a group that is defined as a mutually exclusive set, any other properties in the group are removed from the returned executor object. All other properties of the returned executor object are identical to those of *this.

see-below require(const execution::allocator_t<void>& a) const;

Returns: require(execution::allocator(x)), where x is an implementation-defined default allocator.

template<class ProtoAllocator>
  see-below require(const execution::allocator_t<ProtoAllocator>& a) const;

Returns: An executor object of an unspecified type conforming to these specifications, associated with the same thread pool as *this, with the execution::allocator_t<ProtoAllocator> property established such that allocation and deallocation associated with function submission will be performed using a copy of a.alloc. All other properties of the returned executor object are identical to those of *this.

static constexpr execution::bulk_guarantee_t query(execution::bulk_guarantee_t) const;

Returns: execution::bulk_guarantee.parallel

static constexpr execution::mapping_t query(execution::mapping_t) const;

Returns: execution::mapping.thread.

execution::blocking_t query(execution::blocking_t) const;
execution::relationship_t query(execution::relationship_t) const;
execution::outstanding_work_t query(execution::outstanding_work_t) const;

Returns: The value of the given property of *this.

static_thread_pool& query(execution::context_t) const noexcept;

Returns: A reference to the associated static_thread_pool object.

see-below query(execution::allocator_t<void>) const noexcept;
see-below query(execution::allocator_t<ProtoAllocator>) const noexcept;

Returns: The allocator object associated with the executor, with type and value as either previously established by the execution::allocator_t<ProtoAllocator> property or the implementation defined default allocator established by the execution::allocator_t<void> property.

bool running_in_this_thread() const noexcept;

Returns: true if the current thread of execution is a thread that was created by or attached to the associated static_thread_pool object.

2.5.5.4 Comparisons

bool operator==(const C& a, const C& b) noexcept;

Returns: true if &a.query(execution::context) == &b.query(execution::context) and a and b have identical properties, otherwise false.

bool operator!=(const C& a, const C& b) noexcept;

Returns: !(a == b).

2.5.5.5 static_thread_pool executor execution functions

In addition to conforming to the above specification, static_thread_pool executors shall conform to the following specification.

class C
{
  public:
    template<class Function>
      void execute(Function&& f) const;

    template<class Function>
      void bulk_execute(Function&& f, size_t n) const;
};

C is a type satisfying the Executor requirements.

template<class Function>
  void execute(Function&& f) const;

Effects: Submits the function f for execution on the static_thread_pool according to the the properties established for *this. If the submitted function f exits via an exception, the static_thread_pool invokes std::terminate().

template<class Function>
  void bulk_execute(Function&& f, size_t n) const;

Effects: Submits the function f for bulk execution on the static_thread_pool according to properties established for *this. If the submitted function f exits via an exception, the static_thread_pool invokes std::terminate().

2.6 Changelog

2.6.1 Revision 12

Introduced introductory design discussion which replaces the obsolete P0761. No normative changes.

2.6.2 Revision 11

As directed by SG1 at the 2019-07 Cologne meeting, we have implemented the following changes suggested by P1658 and P1660 which incorporate “lazy” execution:

2.6.3 Revision 10

As directed by LEWG at the 2018-11 San Diego meeting, we have migrated the property customization mechanism to namespace std and moved all of the details of its specification to a separate paper, P1393. This change also included the introduction of a separate customization point for interface-enforcing properties, require_concept. The generalization also necessitated the introduction of is_applicable_property_v in the properties paper, which in turn led to the introduction of is_executor_v to express the applicability of properties in this paper.

2.6.4 Revision 9

As directed by the SG1/LEWG straw poll taken during the 2018 Bellevue executors meeting, we have separated The Unified Executors programming model proposal into two papers. This paper contains material related to one-way execution which the authors hope to standardize with C++20 as suggested by the Bellevue poll. P1244 contains remaining material related to dependent execution. We expect P1244 to evolve as committee consensus builds around a design for dependent execution.

This revision also contains bug fixes to the allocator_t property which were originally scheduled for Revision 7 but were inadvertently omitted.

2.6.5 Revision 8

Revision 8 of this proposal makes interface-changing properties such as oneway mutually exclusive in order to simplify implementation requirements for executor adaptors such as polymorphic executors. Additionally, this revision clarifies wording regarding execution agent lifetime.

2.6.6 Revision 7

Revision 7 of this proposal corrects wording bugs discovered by the authors after Revision 6’s publication.

2.6.7 Revision 6

Revision 6 of this proposal corrects bugs and omissions discovered by the authors after Revision 5’s publication, and introduces an enhancement improving the safety of the design.

2.6.8 Revision 5

Revision 5 of this proposal responds to feedback requested during the 2017 Albuquerque ISO C++ Standards Committee meeting and introduces changes which allow properties to better interoperate with polymorphic executor wrappers and also simplify execution::require’s behavior.

2.6.9 Revision 4

2.6.10 Revision 3

2.6.11 Revision 2

2.6.12 Revision 1

2.6.13 Revision 0

2.7 Appendix: Executors Bibilography

Paper Notes Date introduced
N3378 - A preliminary proposal for work executors
N3562 - Executors and schedulers, revision 1
N3731 - Executors and schedulers, revision 2
N3785 - Executors and schedulers, revision 3
N4143 - Executors and schedulers, revision 4
N4414 - Executors and schedulers, revision 5
P0008 - C++ Executors
Initial executors proposal from Google, based on an abstract base class. 2012-02-24
N4046 - Executors and Asynchronous Operations Initial executors proposal from Kohlhoff, based on extensions to ASIO. 2014-05-26
N4406 - Parallel Algorithms Need Executors
P0058 - An interface for abstracting execution
Initial executors proposal from Nvidia, based on a traits class. 2015-04-10
P0285 - Using customization points to unify executors Proposes unifying various competing executors proposals via customization points. 2016-02-14
P0443 - A Unified Executors Proposal for C++ This proposal. 2016-10-17
P0688 - A Proposal to Simplify the Executors Design Proposes simplifying this proposal’s APIs using properties. 2017-06-19
P0761 - Executors Design Document Describes the design of this proposal circa 2017. 2017-07-31
P1055 - A Modest Executor Proposal Initial executors proposal from Facebook, based on lazy execution. 2018-04-26
P1194 - The Compromise Executors Proposal: A lazy simplification of P0443 Initial proposal to integrate senders and receivers into this proposal. 2018-10-08
P1232 - Integrating executors with the standard library through customization Proposes to allow executors to customize standard algorithms directly. 2018-10-08
P1244 - Dependent Execution for a Unified Executors Proposal for C++ Vestigal futures-based dependent execution functionality excised from later revisions of this proposal. 2018-10-08
P1341 - Unifying asynchronous APIs in C++ standard Library Proposes enhancements making senders awaitable. 2018-11-25
P1393 - A General Property Customization Mechanism Standalone paper proposing the property customization used by P0443 executors. 2019-01-13
P1677 - Cancellation is serendipitous-success Motivates the need for done in addition to error. 2019-05-18
P1678 - Callbacks and Composition Argues for callbacks/receivers as a universal design pattern in the standard library. 2019-05-18
P1525 - One-Way execute is a Poor Basis Operation Identifies deficiencies of execute as a basis operation. 2019-06-17
P1658 - Suggestions for Consensus on Executors Suggests progress-making changes to this proposal circa 2019. 2019-06-17
P1660 - A Compromise Executor Design Sketch Proposes concrete changes to this proposal along the lines of P1525, P1658, and P1738. 2019-06-17
P1738 - The Executor Concept Hierarchy Needs a Single Root Identifies problems caused by a multi-root executor concept hierarchy. 2019-06-17
P1897 - Towards C++23 executors: A proposal for an initial set of algorithms Initial proposal for a set of customizable sender algorithms. 2019-10-06
P1898 - Forward progress delegation for executors Proposes a model of forward progress for executors and asynchronous graphs of work. 2019-10-06
P2006 - Splitting submit() into connect()/start() Proposes refactoring submit into more fundamental connect and start sender operations. 2020-01-13
P2033 - History of Executor Properties Documents the evolution of P1393’s property system, especially as it relates to executors. 2020-01-13

2.8 Appendix: A note on coroutines

P1341 leverages the structural similarities between coroutines and the sender/receiver abstraction to give a class of senders a standard-provided operator co_await. The end result is that a sender, simply by dint of being a sender, can be co_await-ed in a coroutine. With the refinement of sender/receiver being proposed in P2006 — namely, the splitting of submit into connect/start — that automatic adaptation from sender-to-awaitable is allocation- and synchronization-free.

2.9 Appendix: The retry Algorithm

Below is an implementation of a simple retry algorithm in terms of sender/receiver. This algorithm is Generic in the sense that it will retry any multi-shot asynchronous operation that satisfies the sender concept. More accurately, it takes any deferred async operation and wraps it so that when it is executed, it will retry the wrapped operation until it either succeeds or is cancelled.

Full working code can be found here: https://godbolt.org/z/nm6GmH

// _conv needed so we can emplace construct non-movable types into
// a std::optional.
template<invocable F>
    requires std::is_nothrow_move_constructible_v<F>
struct _conv {
    F f_;
    explicit _conv(F f) noexcept : f_((F&&) f) {}
    operator invoke_result_t<F>() && {
        return ((F&&) f_)();
    }
};

// pass through set_value and set_error, but retry the operation
// from set_error.
template<class O, class R>
struct _retry_receiver {
    O* o_;
    explicit _retry_receiver(O* o): o_(o) {}
    template<class... As>
        requires receiver_of<R, As...>
    void set_value(As&&... as) &&
        noexcept(is_nothrow_receiver_of_v<R, As...>) {
        ::set_value(std::move(o_->r_), (As&&) as...);
    }
    void set_error(auto&&) && noexcept {
        o_->_retry(); // This causes the op to be retried
    }
    void set_done() && noexcept {
        ::set_done(std::move(o_->r_));
    }
};

template<sender S>
struct _retry_sender : sender_base {
    S s_;
    explicit _retry_sender(S s): s_((S&&) s) {}

    // Hold the nested operation state in an optional so we can
    // re-construct and re-start it when the operation fails.
    template<receiver R>
    struct _op {
        S s_;
        R r_;
        std::optional<state_t<S&, _retry_receiver<_op, R>>> o_;

        _op(S s, R r): s_((S&&)s), r_((R&&)r), o_{_connect()} {}
        _op(_op&&) = delete;

        auto _connect() noexcept {
            return _conv{[this] {
                return ::connect(s_, _retry_receiver<_op, R>{this});
            }};
        }
        void _retry() noexcept try {
            o_.emplace(_connect()); // potentially throwing
            ::start(std::move(*o_));
        } catch(...) {
            ::set_error((R&&) r_, std::current_exception());
        }
        void start() && noexcept {
            ::start(std::move(*o_));
        }
    };

    template<receiver R>
        requires sender_to<S&, _retry_receiver<_op<R>, R>>
    auto connect(R r) && -> _op<R> {
        return _op<R>{(S&&) s_, (R&&) r};
    }
};

template<sender S>
sender auto retry(S s) {
    return _retry_sender{(S&&)s};
}