xref: /dports/devel/transwarp/transwarp-2.2.2/include/transwarp.h

/// @mainpage transwarp is a header-only C++ library for task concurrency
/// @details https://github.com/bloomen/transwarp
/// @version 2.2.2
/// @author Christian Blume, Guan Wang
/// @date 2019
/// @copyright MIT http://www.opensource.org/licenses/mit-license.php
#pragma once
#include <algorithm>
#ifndef TRANSWARP_CPP11
#include <any>
#endif
#include <array>
#include <atomic>
#include <chrono>
#include <cstddef>
#include <cstdint>
#include <functional>
#include <future>
#include <map>
#include <memory>
#include <mutex>
#ifndef TRANSWARP_CPP11
#include <optional>
#endif
#include <queue>
#include <stdexcept>
#include <string>
#include <thread>
#include <tuple>
#include <type_traits>
#include <unordered_map>
#include <utility>
#include <vector>


#ifdef TRANSWARP_MINIMUM_TASK_SIZE

#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
#define TRANSWARP_DISABLE_TASK_CUSTOM_DATA
#endif

#ifndef TRANSWARP_DISABLE_TASK_NAME
#define TRANSWARP_DISABLE_TASK_NAME
#endif

#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
#define TRANSWARP_DISABLE_TASK_PRIORITY
#endif

#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
#define TRANSWARP_DISABLE_TASK_REFCOUNT
#endif

#ifndef TRANSWARP_DISABLE_TASK_TIME
#define TRANSWARP_DISABLE_TASK_TIME
#endif

#endif


/// The transwarp namespace
namespace transwarp {


#ifdef TRANSWARP_CPP11
/// A simple value class that optionally holds a string
class option_str {
public:
    option_str() {}

    option_str(std::string str)
    : str_(std::move(str)),
      valid_(true)
    {}

    // default copy/move semantics
    option_str(const option_str&) = default;
    option_str& operator=(const option_str&) = default;
    option_str(option_str&&) = default;
    option_str& operator=(option_str&&) = default;

    operator bool() const noexcept {
        return valid_;
    }

    const std::string& operator*() const noexcept {
        return str_;
    }

private:
    std::string str_;
    bool valid_ = false;
};

/// Detail namespace for internal functionality only
namespace detail {

/// Used to handle data storage for a type-erased object
class storage {
public:
    virtual ~storage() = default;
    virtual std::unique_ptr<storage> clone() const = 0;
    virtual void destroy(void* data) const noexcept = 0;
    virtual void copy(const void* src, void*& dest) const = 0;
    virtual void move(void*& src, void*& dest) const noexcept = 0;
};

/// Used to handle data storage for a type-erased object
template<typename T>
class storage_impl : public transwarp::detail::storage {
public:
    std::unique_ptr<transwarp::detail::storage> clone() const override {
        return std::unique_ptr<transwarp::detail::storage>(new storage_impl);
    }
    void destroy(void* data) const noexcept override {
        delete reinterpret_cast<T*>(data);
    }
    void copy(const void* src, void*& dest) const override {
        dest = new T(*reinterpret_cast<const T*>(src));
    }
    void move(void*& src, void*& dest) const noexcept override {
        dest = src;
        src = nullptr;
    }
};

} // detail

/// A simple value class that can hold any value
class any_data {
public:
    any_data()
    : data_(nullptr)
    {}

    template<typename T>
    any_data(T&& value)
    : storage_(new transwarp::detail::storage_impl<typename std::decay<T>::type>),
      data_(new typename std::decay<T>::type(std::forward<T>(value)))
    {}

    any_data(const any_data& other)
    : storage_(other.storage_ ? other.storage_->clone() : nullptr)
    {
        if (other.data_) {
            storage_->copy(other.data_, data_);
        } else {
            data_ = nullptr;
        }
    }

    any_data& operator=(const any_data& other) {
        if (this != &other) {
            if (storage_) {
                storage_->destroy(data_);
            }
            storage_ = other.storage_ ? other.storage_->clone() : nullptr;
            if (other.data_) {
                storage_->copy(other.data_, data_);
            } else {
                data_ = nullptr;
            }
        }
        return *this;
    }

    any_data(any_data&& other)
    : storage_(std::move(other.storage_))
    {
        if (other.data_) {
            storage_->move(other.data_, data_);
        } else {
            data_ = nullptr;
        }
    }

    any_data& operator=(any_data&& other) {
        if (this != &other) {
            if (storage_) {
                storage_->destroy(data_);
            }
            storage_ = std::move(other.storage_);
            if (other.data_) {
                storage_->move(other.data_, data_);
            } else {
                data_ = nullptr;
            }
        }
        return *this;
    }

    ~any_data() {
        if (data_) {
            storage_->destroy(data_);
        }
    }

    bool has_value() const noexcept {
        return data_ != nullptr;
    }

    template<typename T>
    const T& get() const {
        return *reinterpret_cast<const T*>(data_);
    }

private:
    std::unique_ptr<transwarp::detail::storage> storage_;
    void* data_;
};

using str_view = const std::string&;
#else
using any_data = std::any;
using option_str = std::optional<std::string>;
using str_view = std::string_view;
#endif


/// The possible task types
enum class task_type {
    root, ///< The task has no parents
    accept, ///< The task's functor accepts all parent futures
    accept_any, ///< The task's functor accepts the first parent future that becomes ready
    consume, ///< The task's functor consumes all parent results
    consume_any, ///< The task's functor consumes the first parent result that becomes ready
    wait, ///< The task's functor takes no arguments but waits for all parents to finish
    wait_any, ///< The task's functor takes no arguments but waits for the first parent to finish
};


/// Base class for exceptions
class transwarp_error : public std::runtime_error {
public:
    explicit transwarp_error(const std::string& message)
    : std::runtime_error{message}
    {}
};


/// Exception thrown when a task is canceled
class task_canceled : public transwarp::transwarp_error {
public:
    explicit task_canceled(const std::string& task_repr)
    : transwarp::transwarp_error{"Task canceled: " + task_repr}
    {}
};


/// Exception thrown when a task was destroyed prematurely
class task_destroyed : public transwarp::transwarp_error {
public:
    explicit task_destroyed(const std::string& task_repr)
    : transwarp::transwarp_error{"Task destroyed: " + task_repr}
    {}
};


/// Exception thrown when an invalid parameter was passed to a function
class invalid_parameter : public transwarp::transwarp_error {
public:
    explicit invalid_parameter(const std::string& parameter)
    : transwarp::transwarp_error{"Invalid parameter: " + parameter}
    {}
};


/// Exception thrown when a task is used in unintended ways
class control_error : public transwarp::transwarp_error {
public:
    explicit control_error(const std::string& message)
    : transwarp::transwarp_error{"Control error: " + message}
    {}
};


/// The root type. Used for tag dispatch
struct root_type : std::integral_constant<transwarp::task_type, transwarp::task_type::root> {};
constexpr transwarp::root_type root{}; ///< The root task tag

/// The accept type. Used for tag dispatch
struct accept_type : std::integral_constant<transwarp::task_type, transwarp::task_type::accept> {};
constexpr transwarp::accept_type accept{}; ///< The accept task tag

/// The accept_any type. Used for tag dispatch
struct accept_any_type : std::integral_constant<transwarp::task_type, transwarp::task_type::accept_any> {};
constexpr transwarp::accept_any_type accept_any{}; ///< The accept_any task tag

/// The consume type. Used for tag dispatch
struct consume_type : std::integral_constant<transwarp::task_type, transwarp::task_type::consume> {};
constexpr transwarp::consume_type consume{}; ///< The consume task tag

/// The consume_any type. Used for tag dispatch
struct consume_any_type : std::integral_constant<transwarp::task_type, transwarp::task_type::consume_any> {};
constexpr transwarp::consume_any_type consume_any{}; ///< The consume_any task tag

/// The wait type. Used for tag dispatch
struct wait_type : std::integral_constant<transwarp::task_type, transwarp::task_type::wait> {};
constexpr transwarp::wait_type wait{}; ///< The wait task tag

/// The wait_any type. Used for tag dispatch
struct wait_any_type : std::integral_constant<transwarp::task_type, transwarp::task_type::wait_any> {};
constexpr transwarp::wait_any_type wait_any{}; ///< The wait_any task tag


class itask;


/// Detail namespace for internal functionality only
namespace detail {

struct visit_visitor;
struct unvisit_visitor;
struct final_visitor;
struct schedule_visitor;
struct parent_visitor;
struct decrement_refcount_functor;

} // detail


/// The executor interface used to perform custom task execution
class executor {
public:
    virtual ~executor() = default;

    /// Returns the name of the executor
    virtual std::string name() const = 0;

    /// Runs a task which is wrapped by the given functor. The functor only
    /// captures one shared pointer and can hence be copied at low cost.
    /// task represents the task that the functor belongs to.
    /// This function is only ever called on the thread of the caller to schedule().
    /// The implementer needs to ensure that this never throws exceptions
    virtual void execute(const std::function<void()>& functor, transwarp::itask& task) = 0;
};


/// The task events that can be subscribed to using the listener interface
enum class event_type {
    before_scheduled, ///< Just before a task is scheduled (handle_event called on thread of caller to schedule())
    after_future_changed, ///< Just after the task's future was changed (handle_event called on thread that changed the task's future)
    before_started, ///< Just before a task starts running (handle_event called on thread that task is run on)
    before_invoked, ///< Just before a task's functor is invoked (handle_event called on thread that task is run on)
    after_finished, ///< Just after a task has finished running (handle_event called on thread that task is run on)
    after_canceled, ///< Just after a task was canceled (handle_event called on thread that task is run on)
    after_satisfied, ///< Just after a task has satisfied all its children with results (handle_event called on thread where the last child is satisfied)
    after_custom_data_set, ///< Just after custom data was assigned (handle_event called on thread that custom data was set on)
    count,
};


/// The listener interface to listen to events raised by tasks
class listener {
public:
    virtual ~listener() = default;

    /// This may be called from arbitrary threads depending on the event type (see transwarp::event_type).
    /// The implementer needs to ensure that this never throws exceptions.
    virtual void handle_event(transwarp::event_type event, transwarp::itask& task) = 0;
};


/// An edge between two tasks
class edge {
public:
    edge(transwarp::itask& parent, transwarp::itask& child) noexcept
    : parent_(&parent), child_(&child)
    {}

    // default copy/move semantics
    edge(const edge&) = default;
    edge& operator=(const edge&) = default;
    edge(edge&&) = default;
    edge& operator=(edge&&) = default;

    /// Returns the parent task
    const transwarp::itask& parent() const noexcept {
        return *parent_;
    }

    /// Returns the parent task
    transwarp::itask& parent() noexcept {
        return *parent_;
    }

    /// Returns the child task
    const transwarp::itask& child() const noexcept {
        return *child_;
    }

    /// Returns the child task
    transwarp::itask& child() noexcept {
        return *child_;
    }

private:
    transwarp::itask* parent_;
    transwarp::itask* child_;
};


class timer;
class releaser;

/// An interface for the task class
class itask : public std::enable_shared_from_this<itask> {
public:
    virtual ~itask() = default;

    virtual void finalize() = 0;
    virtual std::size_t id() const noexcept = 0;
    virtual std::size_t level() const noexcept = 0;
    virtual transwarp::task_type type() const noexcept = 0;
    virtual const transwarp::option_str& name() const noexcept = 0;
    virtual std::shared_ptr<transwarp::executor> executor() const noexcept = 0;
    virtual std::int64_t priority() const noexcept = 0;
    virtual const transwarp::any_data& custom_data() const noexcept = 0;
    virtual bool canceled() const noexcept = 0;
    virtual std::int64_t avg_idletime_us() const noexcept = 0;
    virtual std::int64_t avg_waittime_us() const noexcept = 0;
    virtual std::int64_t avg_runtime_us() const noexcept = 0;
    virtual void set_executor(std::shared_ptr<transwarp::executor> executor) = 0;
    virtual void set_executor_all(std::shared_ptr<transwarp::executor> executor) = 0;
    virtual void remove_executor() = 0;
    virtual void remove_executor_all() = 0;
    virtual void set_priority(std::int64_t priority) = 0;
    virtual void set_priority_all(std::int64_t priority) = 0;
    virtual void reset_priority() = 0;
    virtual void reset_priority_all() = 0;
    virtual void set_custom_data(transwarp::any_data custom_data) = 0;
    virtual void set_custom_data_all(transwarp::any_data custom_data) = 0;
    virtual void remove_custom_data() = 0;
    virtual void remove_custom_data_all() = 0;
    virtual void add_listener(std::shared_ptr<transwarp::listener> listener) = 0;
    virtual void add_listener(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener) = 0;
    virtual void add_listener_all(std::shared_ptr<transwarp::listener> listener) = 0;
    virtual void add_listener_all(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener) = 0;
    virtual void remove_listener(const std::shared_ptr<transwarp::listener>& listener) = 0;
    virtual void remove_listener(transwarp::event_type event, const std::shared_ptr<transwarp::listener>& listener) = 0;
    virtual void remove_listener_all(const std::shared_ptr<transwarp::listener>& listener) = 0;
    virtual void remove_listener_all(transwarp::event_type event, const std::shared_ptr<transwarp::listener>& listener) = 0;
    virtual void remove_listeners() = 0;
    virtual void remove_listeners(transwarp::event_type event) = 0;
    virtual void remove_listeners_all() = 0;
    virtual void remove_listeners_all(transwarp::event_type event) = 0;
    virtual void schedule() = 0;
    virtual void schedule(transwarp::executor& executor) = 0;
    virtual void schedule(bool reset) = 0;
    virtual void schedule(transwarp::executor& executor, bool reset) = 0;
    virtual void schedule_all() = 0;
    virtual void schedule_all(transwarp::executor& executor) = 0;
    virtual void schedule_all(bool reset_all) = 0;
    virtual void schedule_all(transwarp::executor& executor, bool reset_all) = 0;
    virtual void set_exception(std::exception_ptr exception) = 0;
    virtual bool was_scheduled() const noexcept = 0;
    virtual void wait() const = 0;
    virtual bool is_ready() const = 0;
    virtual bool has_result() const = 0;
    virtual void reset() = 0;
    virtual void reset_all() = 0;
    virtual void cancel(bool enabled) noexcept = 0;
    virtual void cancel_all(bool enabled) noexcept = 0;
    virtual std::vector<itask*> parents() const = 0;
    virtual const std::vector<itask*>& tasks() = 0;
    virtual std::vector<transwarp::edge> edges() = 0;

protected:
    virtual void schedule_impl(bool reset, transwarp::executor* executor=nullptr) = 0;

private:
    friend struct transwarp::detail::visit_visitor;
    friend struct transwarp::detail::unvisit_visitor;
    friend struct transwarp::detail::final_visitor;
    friend struct transwarp::detail::schedule_visitor;
    friend struct transwarp::detail::parent_visitor;
    friend class transwarp::timer;
    friend class transwarp::releaser;
    friend struct transwarp::detail::decrement_refcount_functor;

    virtual void visit(const std::function<void(itask&)>& visitor) = 0;
    virtual void unvisit() noexcept = 0;
    virtual void set_id(std::size_t id) noexcept = 0;
    virtual void set_level(std::size_t level) noexcept = 0;
    virtual void set_type(transwarp::task_type type) noexcept = 0;
    virtual void set_name(transwarp::option_str name) noexcept = 0;
    virtual void set_avg_idletime_us(std::int64_t idletime) noexcept = 0;
    virtual void set_avg_waittime_us(std::int64_t waittime) noexcept = 0;
    virtual void set_avg_runtime_us(std::int64_t runtime) noexcept = 0;
    virtual void increment_childcount() noexcept = 0;
    virtual void decrement_refcount() = 0;
    virtual void reset_future() = 0;
};


/// String conversion for the task_type enumeration
inline
std::string to_string(const transwarp::task_type& type) {
    switch (type) {
    case transwarp::task_type::root: return "root";
    case transwarp::task_type::accept: return "accept";
    case transwarp::task_type::accept_any: return "accept_any";
    case transwarp::task_type::consume: return "consume";
    case transwarp::task_type::consume_any: return "consume_any";
    case transwarp::task_type::wait: return "wait";
    case transwarp::task_type::wait_any: return "wait_any";
    }
    throw transwarp::invalid_parameter{"task type"};
}


/// String conversion for the itask class
inline
std::string to_string(const transwarp::itask& task, transwarp::str_view separator="\n") {
    std::string s;
    s += '"';
    const transwarp::option_str& name = task.name();
    if (name) {
        s += std::string{"<"} + *name + std::string{">"} + separator.data();
    }
    s += transwarp::to_string(task.type());
    s += std::string{" id="} + std::to_string(task.id());
    s += std::string{" lev="} + std::to_string(task.level());
    const std::shared_ptr<transwarp::executor> exec = task.executor();
    if (exec) {
        s += separator.data() + std::string{"<"} + exec->name() + std::string{">"};
    }
    const std::int64_t avg_idletime_us = task.avg_idletime_us();
    if (avg_idletime_us >= 0) {
        s += separator.data() + std::string{"avg-idle-us="} + std::to_string(avg_idletime_us);
    }
    const std::int64_t avg_waittime_us = task.avg_waittime_us();
    if (avg_waittime_us >= 0) {
        s += separator.data() + std::string{"avg-wait-us="} + std::to_string(avg_waittime_us);
    }
    const std::int64_t avg_runtime_us = task.avg_runtime_us();
    if (avg_runtime_us >= 0) {
        s += separator.data() + std::string{"avg-run-us="} + std::to_string(avg_runtime_us);
    }
    return s + '"';
}


/// String conversion for the edge class
inline
std::string to_string(const transwarp::edge& edge, transwarp::str_view separator="\n") {
    return transwarp::to_string(edge.parent(), separator) + std::string{" -> "} + transwarp::to_string(edge.child(), separator);
}


/// Creates a dot-style string from the given edges
inline
std::string to_string(const std::vector<transwarp::edge>& edges, transwarp::str_view separator="\n") {
    std::string dot = std::string{"digraph {"} + separator.data();
    for (const transwarp::edge& edge : edges) {
        dot += transwarp::to_string(edge, separator) + separator.data();
    }
    dot += std::string{"}"};
    return dot;
}


/// Removes reference and const from a type
template<typename T>
struct decay {
    using type = typename std::remove_const<typename std::remove_reference<T>::type>::type;
};


/// Returns the result type of a std::shared_future<T>
template<typename T>
struct result {
    using type = decltype(std::declval<std::shared_future<T>>().get());
};


/// Detail namespace for internal functionality only
namespace detail {

/// Clones a task
template<typename TaskType>
std::shared_ptr<TaskType> clone_task(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const std::shared_ptr<TaskType>& t) {
    const auto original_task = std::static_pointer_cast<transwarp::itask>(t);
    const auto task_cache_it = task_cache.find(original_task);
    if (task_cache_it != task_cache.cend()) {
        return std::static_pointer_cast<TaskType>(task_cache_it->second);
    } else {
        auto cloned_task = t->clone_impl(task_cache);
        task_cache[original_task] = cloned_task;
        return cloned_task;
    }
}

} // detail


/// The task class
template<typename ResultType>
class task : public transwarp::itask {
public:
    using result_type = ResultType;

    virtual ~task() = default;

    std::shared_ptr<task> clone() const {
        std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>> task_cache;
        return clone_impl(task_cache);
    }

    virtual void set_value(const typename transwarp::decay<result_type>::type& value) = 0;
    virtual void set_value(typename transwarp::decay<result_type>::type&& value) = 0;
    virtual std::shared_future<result_type> future() const noexcept = 0;
    virtual typename transwarp::result<result_type>::type get() const = 0;

private:
    template<typename T>
    friend std::shared_ptr<T> transwarp::detail::clone_task(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const std::shared_ptr<T>& t);

    virtual std::shared_ptr<transwarp::task<result_type>> clone_impl(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache) const = 0;
};

/// The task class (reference result type)
template<typename ResultType>
class task<ResultType&> : public transwarp::itask {
public:
    using result_type = ResultType&;

    virtual ~task() = default;

    std::shared_ptr<task> clone() const {
        std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>> task_cache;
        return clone_impl(task_cache);
    }

    virtual void set_value(typename transwarp::decay<result_type>::type& value) = 0;
    virtual std::shared_future<result_type> future() const noexcept = 0;
    virtual typename transwarp::result<result_type>::type get() const = 0;

private:
    template<typename T>
    friend std::shared_ptr<T> transwarp::detail::clone_task(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const std::shared_ptr<T>& t);

    virtual std::shared_ptr<transwarp::task<result_type>> clone_impl(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache) const = 0;
};

/// The task class (void result type)
template<>
class task<void> : public transwarp::itask {
public:
    using result_type = void;

    virtual ~task() = default;

    std::shared_ptr<task> clone() const {
        std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>> task_cache;
        return clone_impl(task_cache);
    }

    virtual void set_value() = 0;
    virtual std::shared_future<result_type> future() const noexcept = 0;
    virtual result_type get() const = 0;

private:
    template<typename T>
    friend std::shared_ptr<T> transwarp::detail::clone_task(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const std::shared_ptr<T>& t);

    virtual std::shared_ptr<transwarp::task<result_type>> clone_impl(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache) const = 0;
};


/// Detail namespace for internal functionality only
namespace detail {

template<bool>
struct assign_task_if_impl;

} // detail


/// A base class for a user-defined functor that needs access to the associated
/// task or a cancel point to stop a task while it's running
class functor {
public:

    virtual ~functor() = default;

protected:

    /// The associated task (only to be called after the task was constructed)
    const transwarp::itask& transwarp_task() const noexcept {
        return *transwarp_task_;
    }

    /// The associated task (only to be called after the task was constructed)
    transwarp::itask& transwarp_task() noexcept {
        return *transwarp_task_;
    }

    /// If the associated task is canceled then this will throw transwarp::task_canceled
    /// which will stop the task while it's running (only to be called after the task was constructed)
    void transwarp_cancel_point() const {
        if (transwarp_task_->canceled()) {
            throw transwarp::task_canceled{std::to_string(transwarp_task_->id())};
        }
    }

private:
    template<bool>
    friend struct transwarp::detail::assign_task_if_impl;

    transwarp::itask* transwarp_task_{};
};


/// Detail namespace for internal functionality only
namespace detail {


/// A simple thread pool used to execute tasks in parallel
class thread_pool {
public:

    explicit thread_pool(const std::size_t n_threads,
                         std::function<void(std::size_t thread_index)> on_thread_started = nullptr)
    : on_thread_started_{std::move(on_thread_started)}
    {
        if (n_threads == 0) {
            throw transwarp::invalid_parameter{"number of threads"};
        }
        for (std::size_t i = 0; i < n_threads; ++i) {
            std::thread thread;
            try {
                thread = std::thread(&thread_pool::worker, this, i);
            } catch (...) {
                shutdown();
                throw;
            }
            try {
                threads_.push_back(std::move(thread));
            } catch (...) {
                shutdown();
                thread.join();
                throw;
            }
        }
    }

    // delete copy/move semantics
    thread_pool(const thread_pool&) = delete;
    thread_pool& operator=(const thread_pool&) = delete;
    thread_pool(thread_pool&&) = delete;
    thread_pool& operator=(thread_pool&&) = delete;

    ~thread_pool() {
        shutdown();
    }

    void push(const std::function<void()>& functor) {
        {
            std::lock_guard<std::mutex> lock{mutex_};
            functors_.push(functor);
        }
        cond_var_.notify_one();
    }

private:

    void worker(const std::size_t index) {
        if (on_thread_started_) {
            on_thread_started_(index);
        }
        for (;;) {
            std::function<void()> functor;
            {
                std::unique_lock<std::mutex> lock{mutex_};
                cond_var_.wait(lock, [this]{
                    return done_ || !functors_.empty();
                });
                if (done_ && functors_.empty()) {
                    break;
                }
                functor = std::move(functors_.front());
                functors_.pop();
            }
            functor();
        }
    }

    void shutdown() {
        {
            std::lock_guard<std::mutex> lock{mutex_};
            done_ = true;
        }
        cond_var_.notify_all();
        for (std::thread& thread : threads_) {
            thread.join();
        }
        threads_.clear();
    }

    bool done_ = false;
    std::function<void(std::size_t)> on_thread_started_;
    std::vector<std::thread> threads_;
    std::queue<std::function<void()>> functors_;
    std::condition_variable cond_var_;
    std::mutex mutex_;
};


#ifdef TRANSWARP_CPP11
template<std::size_t...> struct indices {};

template<std::size_t...> struct construct_range;

template<std::size_t end, std::size_t idx, std::size_t... i>
struct construct_range<end, idx, i...> : construct_range<end, idx + 1, i..., idx> {};

template<std::size_t end, std::size_t... i>
struct construct_range<end, end, i...> {
    using type = transwarp::detail::indices<i...>;
};

template<std::size_t b, std::size_t e>
struct index_range {
    using type = typename transwarp::detail::construct_range<e, b>::type;
};

template<typename Functor, typename Tuple>
void call_with_each_index(transwarp::detail::indices<>, Functor&&, Tuple&&) {}

template<std::size_t i, std::size_t... j, typename Functor, typename Tuple>
void call_with_each_index(transwarp::detail::indices<i, j...>, Functor&& f, Tuple&& t) {
    f(std::get<i>(t));
    transwarp::detail::call_with_each_index(transwarp::detail::indices<j...>{}, std::forward<Functor>(f), std::forward<Tuple>(t));
}
#endif

template<typename Functor, typename Tuple>
void apply_to_each(Functor&& f, Tuple&& t) {
#ifdef TRANSWARP_CPP11
    constexpr std::size_t n = std::tuple_size<typename std::decay<Tuple>::type>::value;
    using index_t = typename transwarp::detail::index_range<0, n>::type;
    transwarp::detail::call_with_each_index(index_t{}, std::forward<Functor>(f), std::forward<Tuple>(t));
#else
    std::apply([&f](auto&&... arg){(..., std::forward<Functor>(f)(std::forward<decltype(arg)>(arg)));}, std::forward<Tuple>(t));
#endif
}

template<typename Functor, typename ElementType>
void apply_to_each(Functor&& f, const std::vector<ElementType>& v) {
    std::for_each(v.begin(), v.end(), std::forward<Functor>(f));
}

template<typename Functor, typename ElementType>
void apply_to_each(Functor&& f, std::vector<ElementType>& v) {
    std::for_each(v.begin(), v.end(), std::forward<Functor>(f));
}


template<int offset, typename... ParentResults>
struct assign_futures_impl {
    static void work(const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& source, std::tuple<std::shared_future<ParentResults>...>& target) {
        std::get<offset>(target) = std::get<offset>(source)->future();
        assign_futures_impl<offset - 1, ParentResults...>::work(source, target);
    }
};

template<typename... ParentResults>
struct assign_futures_impl<-1, ParentResults...> {
    static void work(const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>&, std::tuple<std::shared_future<ParentResults>...>&) {}
};

/// Returns the futures from the given tuple of tasks
template<typename... ParentResults>
std::tuple<std::shared_future<ParentResults>...> get_futures(const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& input) {
    std::tuple<std::shared_future<ParentResults>...> result;
    assign_futures_impl<static_cast<int>(sizeof...(ParentResults)) - 1, ParentResults...>::work(input, result);
    return result;
}

/// Returns the futures from the given vector of tasks
template<typename ParentResultType>
std::vector<std::shared_future<ParentResultType>> get_futures(const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& input) {
    std::vector<std::shared_future<ParentResultType>> result;
    result.reserve(input.size());
    for (const std::shared_ptr<transwarp::task<ParentResultType>>& task : input) {
        result.emplace_back(task->future());
    }
    return result;
}


/// Runs the task with the given arguments, hence, invoking the task's functor
template<typename Result, typename Task, typename... Args>
Result run_task(std::size_t task_id, const std::weak_ptr<Task>& task, Args&&... args) {
    const std::shared_ptr<Task> t = task.lock();
    if (!t) {
        throw transwarp::task_destroyed{std::to_string(task_id)};
    }
    if (t->canceled()) {
        throw transwarp::task_canceled{std::to_string(task_id)};
    }
    t->raise_event(transwarp::event_type::before_invoked);
    return (*t->functor_)(std::forward<Args>(args)...);
}


struct wait_for_all_functor {
    template<typename T>
    void operator()(const T& p) const {
        p->future().wait();
    }
};

/// Waits for all parents to finish
template<typename... ParentResults>
void wait_for_all(const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::wait_for_all_functor{}, parents);
}

/// Waits for all parents to finish
template<typename ParentResultType>
void wait_for_all(const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::wait_for_all_functor{}, parents);
}


template<typename Parent>
Parent wait_for_any_impl() {
    return {};
}

template<typename Parent, typename ParentResult, typename... ParentResults>
Parent wait_for_any_impl(const std::shared_ptr<transwarp::task<ParentResult>>& parent, const std::shared_ptr<transwarp::task<ParentResults>>& ...parents) {
    const std::future_status status = parent->future().wait_for(std::chrono::microseconds(1));
    if (status == std::future_status::ready) {
        return parent;
    }
    return transwarp::detail::wait_for_any_impl<Parent>(parents...);
}

/// Waits for the first parent to finish
template<typename Parent, typename... ParentResults>
Parent wait_for_any(const std::shared_ptr<transwarp::task<ParentResults>>& ...parents) {
    for (;;) {
        Parent parent = transwarp::detail::wait_for_any_impl<Parent>(parents...);
        if (parent) {
            return parent;
        }
    }
}


/// Waits for the first parent to finish
template<typename ParentResultType>
std::shared_ptr<transwarp::task<ParentResultType>> wait_for_any(const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
    for (;;) {
        for (const std::shared_ptr<transwarp::task<ParentResultType>>& parent : parents) {
            const std::future_status status = parent->future().wait_for(std::chrono::microseconds(1));
            if (status == std::future_status::ready) {
                return parent;
            }
        }
    }
}


template<typename OneResult>
struct cancel_all_but_one_functor {
    explicit cancel_all_but_one_functor(const std::shared_ptr<transwarp::task<OneResult>>& one) noexcept
    : one_(one) {}

    template<typename T>
    void operator()(const T& parent) const {
        if (one_ != parent) {
            parent->cancel(true);
        }
    }

    const std::shared_ptr<transwarp::task<OneResult>>& one_;
};

/// Cancels all tasks but one
template<typename OneResult, typename... ParentResults>
void cancel_all_but_one(const std::shared_ptr<transwarp::task<OneResult>>& one, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::cancel_all_but_one_functor<OneResult>{one}, parents);
}

/// Cancels all tasks but one
template<typename OneResult, typename ParentResultType>
void cancel_all_but_one(const std::shared_ptr<transwarp::task<OneResult>>& one, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::cancel_all_but_one_functor<OneResult>{one}, parents);
}


struct decrement_refcount_functor {
    template<typename T>
    void operator()(const T& task) const {
        task->decrement_refcount();
    }
};

/// Decrements the refcount of all parents
template<typename... ParentResults>
void decrement_refcount(const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::decrement_refcount_functor{}, parents);
}

/// Decrements the refcount of all parents
template<typename ParentResultType>
void decrement_refcount(const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
    transwarp::detail::apply_to_each(transwarp::detail::decrement_refcount_functor{}, parents);
}


template<typename TaskType, bool done, int total, int... n>
struct call_impl {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        return call_impl<TaskType, total == 1 + static_cast<int>(sizeof...(n)), total, n..., static_cast<int>(sizeof...(n))>::template
                work<Result>(task_id, task, parents);
    }
};

template<typename TaskType>
struct call_impl_vector;

template<int total, int... n>
struct call_impl<transwarp::root_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>&) {
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

template<>
struct call_impl_vector<transwarp::root_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>&) {
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

template<int total, int... n>
struct call_impl<transwarp::accept_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        transwarp::detail::wait_for_all(parents);
        const std::tuple<std::shared_future<ParentResults>...> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, std::get<n>(futures)...);
    }
};

template<>
struct call_impl_vector<transwarp::accept_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        transwarp::detail::wait_for_all(parents);
        std::vector<std::shared_future<ParentResultType>> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, std::move(futures));
    }
};

template<int total, int... n>
struct call_impl<transwarp::accept_any_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        using parent_t = typename std::remove_reference<decltype(std::get<0>(parents))>::type; // Use first type as reference
        parent_t parent = transwarp::detail::wait_for_any<parent_t>(std::get<n>(parents)...);
        transwarp::detail::cancel_all_but_one(parent, parents);
        auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, std::move(future));
    }
};

template<>
struct call_impl_vector<transwarp::accept_any_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        std::shared_ptr<transwarp::task<ParentResultType>> parent = transwarp::detail::wait_for_any(parents);
        transwarp::detail::cancel_all_but_one(parent, parents);
        auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, std::move(future));
    }
};

template<int total, int... n>
struct call_impl<transwarp::consume_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        transwarp::detail::wait_for_all(parents);
        const std::tuple<std::shared_future<ParentResults>...> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, std::get<n>(futures).get()...);
    }
};

template<>
struct call_impl_vector<transwarp::consume_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        transwarp::detail::wait_for_all(parents);
        const std::vector<std::shared_future<ParentResultType>> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        std::vector<ParentResultType> results;
        results.reserve(futures.size());
        for (const std::shared_future<ParentResultType>& future : futures) {
            results.emplace_back(future.get());
        }
        return transwarp::detail::run_task<Result>(task_id, task, std::move(results));
    }
};

template<int total, int... n>
struct call_impl<transwarp::consume_any_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        using parent_t = typename std::remove_reference<decltype(std::get<0>(parents))>::type; /// Use first type as reference
        parent_t parent = transwarp::detail::wait_for_any<parent_t>(std::get<n>(parents)...);
        transwarp::detail::cancel_all_but_one(parent, parents);
        const auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, future.get());
    }
};

template<>
struct call_impl_vector<transwarp::consume_any_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        std::shared_ptr<transwarp::task<ParentResultType>> parent = transwarp::detail::wait_for_any(parents);
        transwarp::detail::cancel_all_but_one(parent, parents);
        const auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        return transwarp::detail::run_task<Result>(task_id, task, future.get());
    }
};

struct future_get_functor {
    template<typename T>
    void operator()(const std::shared_future<T>& f) const {
        f.get();
    }
};

template<int total, int... n>
struct call_impl<transwarp::wait_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        transwarp::detail::wait_for_all(parents);
        const std::tuple<std::shared_future<ParentResults>...> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        transwarp::detail::apply_to_each(transwarp::detail::future_get_functor{}, futures); // Ensures that exceptions are propagated
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

template<>
struct call_impl_vector<transwarp::wait_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        transwarp::detail::wait_for_all(parents);
        const std::vector<std::shared_future<ParentResultType>> futures = transwarp::detail::get_futures(parents);
        transwarp::detail::decrement_refcount(parents);
        transwarp::detail::apply_to_each(transwarp::detail::future_get_functor{}, futures); // Ensures that exceptions are propagated
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

template<int total, int... n>
struct call_impl<transwarp::wait_any_type, true, total, n...> {
    template<typename Result, typename Task, typename... ParentResults>
    static Result work(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
        using parent_t = typename std::remove_reference<decltype(std::get<0>(parents))>::type; // Use first type as reference
        parent_t parent = transwarp::detail::wait_for_any<parent_t>(std::get<n>(parents)...);
        transwarp::detail::cancel_all_but_one(parent, parents);
        const auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        future.get(); // Ensures that exceptions are propagated
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

template<>
struct call_impl_vector<transwarp::wait_any_type> {
    template<typename Result, typename Task, typename ParentResultType>
    static Result work(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
        std::shared_ptr<transwarp::task<ParentResultType>> parent = transwarp::detail::wait_for_any(parents);
        transwarp::detail::cancel_all_but_one(parent, parents);
        const auto future = parent->future();
        transwarp::detail::decrement_refcount(parents);
        future.get(); // Ensures that exceptions are propagated
        return transwarp::detail::run_task<Result>(task_id, task);
    }
};

/// Calls the functor of the given task with the results from the tuple of parents.
/// Throws transwarp::task_canceled if the task is canceled.
/// Throws transwarp::task_destroyed in case the task was destroyed prematurely.
template<typename TaskType, typename Result, typename Task, typename... ParentResults>
Result call(std::size_t task_id, const Task& task, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& parents) {
    constexpr std::size_t n = std::tuple_size<std::tuple<std::shared_future<ParentResults>...>>::value;
    return transwarp::detail::call_impl<TaskType, 0 == n, static_cast<int>(n)>::template
            work<Result>(task_id, task, parents);
}

/// Calls the functor of the given task with the results from the vector of parents.
/// Throws transwarp::task_canceled if the task is canceled.
/// Throws transwarp::task_destroyed in case the task was destroyed prematurely.
template<typename TaskType, typename Result, typename Task, typename ParentResultType>
Result call(std::size_t task_id, const Task& task, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& parents) {
    return transwarp::detail::call_impl_vector<TaskType>::template
            work<Result>(task_id, task, parents);
}


template<typename Functor>
struct call_with_each_functor {
    explicit call_with_each_functor(const Functor& f) noexcept
    : f_(f) {}

    template<typename T>
    void operator()(const T& task) const {
        if (!task) {
            throw transwarp::invalid_parameter{"task pointer"};
        }
        f_(*task);
    }

    const Functor& f_;
};

/// Calls the functor with every element in the tuple
template<typename Functor, typename... ParentResults>
void call_with_each(const Functor& f, const std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>& t) {
    transwarp::detail::apply_to_each(transwarp::detail::call_with_each_functor<Functor>{f}, t);
}

/// Calls the functor with every element in the vector
template<typename Functor, typename ParentResultType>
void call_with_each(const Functor& f, const std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>& v) {
    transwarp::detail::apply_to_each(transwarp::detail::call_with_each_functor<Functor>{f}, v);
}


/// Sets level of a task and increments the child count
struct parent_visitor {
    explicit parent_visitor(transwarp::itask& task) noexcept
    : task_(task) {}

    void operator()(transwarp::itask& task) const {
        if (task_.level() <= task.level()) {
            // A child's level is always larger than any of its parents' levels
            task_.set_level(task.level() + 1);
        }
        task.increment_childcount();
    }

    transwarp::itask& task_;
};

/// Applies final bookkeeping to the task and collects the task
struct final_visitor {
    explicit final_visitor(std::vector<transwarp::itask*>& tasks) noexcept
    : tasks_(tasks) {}

    void operator()(transwarp::itask& task) noexcept {
        tasks_.push_back(&task);
        task.set_id(id_++);
    }

    std::vector<transwarp::itask*>& tasks_;
    std::size_t id_ = 0;
};

/// Generates edges
struct edges_visitor {
    explicit edges_visitor(std::vector<transwarp::edge>& edges) noexcept
    : edges_(edges) {}

    void operator()(transwarp::itask& task) {
        for (transwarp::itask* parent : task.parents()) {
            edges_.emplace_back(*parent, task);
        }
    }

    std::vector<transwarp::edge>& edges_;
};

/// Schedules using the given executor
struct schedule_visitor {
    schedule_visitor(bool reset, transwarp::executor* executor) noexcept
    : reset_(reset), executor_(executor) {}

    void operator()(transwarp::itask& task) {
        task.schedule_impl(reset_, executor_);
    }

    bool reset_;
    transwarp::executor* executor_;
};

/// Resets the given task
struct reset_visitor {

    void operator()(transwarp::itask& task) const {
        task.reset();
    }
};

/// Cancels or resumes the given task
struct cancel_visitor {
    explicit cancel_visitor(bool enabled) noexcept
    : enabled_{enabled} {}

    void operator()(transwarp::itask& task) const noexcept {
        task.cancel(enabled_);
    }

    bool enabled_;
};

/// Assigns an executor to the given task
struct set_executor_visitor {
    explicit set_executor_visitor(std::shared_ptr<transwarp::executor> executor) noexcept
    : executor_{std::move(executor)} {}

    void operator()(transwarp::itask& task) const noexcept {
        task.set_executor(executor_);
    }

    std::shared_ptr<transwarp::executor> executor_;
};

/// Removes the executor from the given task
struct remove_executor_visitor {

    void operator()(transwarp::itask& task) const noexcept {
        task.remove_executor();
    }
};

/// Assigns a priority to the given task
struct set_priority_visitor {
    explicit set_priority_visitor(std::int64_t priority) noexcept
    : priority_{priority} {}

    void operator()(transwarp::itask& task) const noexcept {
        task.set_priority(priority_);
    }

    std::int64_t priority_;
};

/// Resets the priority of the given task
struct reset_priority_visitor {

    void operator()(transwarp::itask& task) const noexcept {
        task.reset_priority();
    }
};

/// Assigns custom data to the given task
struct set_custom_data_visitor {
    explicit set_custom_data_visitor(transwarp::any_data custom_data) noexcept
    : custom_data_{std::move(custom_data)} {}

    void operator()(transwarp::itask& task) const noexcept {
        task.set_custom_data(custom_data_);
    }

    transwarp::any_data custom_data_;
};

/// Removes custom data from the given task
struct remove_custom_data_visitor {

    void operator()(transwarp::itask& task) const noexcept {
        task.remove_custom_data();
    }
};

/// Pushes the given task into the vector of tasks
struct push_task_visitor {
    explicit push_task_visitor(std::vector<transwarp::itask*>& tasks)
    : tasks_(tasks) {}

    void operator()(transwarp::itask& task) {
        tasks_.push_back(&task);
    }

    std::vector<transwarp::itask*>& tasks_;
};

/// Adds a new listener to the given task
struct add_listener_visitor {
    explicit add_listener_visitor(std::shared_ptr<transwarp::listener> listener)
    : listener_(std::move(listener))
    {}

    void operator()(transwarp::itask& task) {
        task.add_listener(listener_);
    }

    std::shared_ptr<transwarp::listener> listener_;
};

/// Adds a new listener per event type to the given task
struct add_listener_per_event_visitor {
    add_listener_per_event_visitor(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener)
    : event_(event), listener_(std::move(listener))
    {}

    void operator()(transwarp::itask& task) {
        task.add_listener(event_, listener_);
    }

    transwarp::event_type event_;
    std::shared_ptr<transwarp::listener> listener_;
};

/// Removes a listener from the given task
struct remove_listener_visitor {
    explicit remove_listener_visitor(std::shared_ptr<transwarp::listener> listener)
    : listener_(std::move(listener))
    {}

    void operator()(transwarp::itask& task) {
        task.remove_listener(listener_);
    }

    std::shared_ptr<transwarp::listener> listener_;
};

/// Removes a listener per event type from the given task
struct remove_listener_per_event_visitor {
    remove_listener_per_event_visitor(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener)
    : event_(event), listener_(std::move(listener))
    {}

    void operator()(transwarp::itask& task) {
        task.remove_listener(event_, listener_);
    }

    transwarp::event_type event_;
    std::shared_ptr<transwarp::listener> listener_;
};

/// Removes all listeners from the given task
struct remove_listeners_visitor {

    void operator()(transwarp::itask& task) {
        task.remove_listeners();
    }

};

/// Removes all listeners per event type from the given task
struct remove_listeners_per_event_visitor {
    explicit remove_listeners_per_event_visitor(transwarp::event_type event)
    : event_(event)
    {}

    void operator()(transwarp::itask& task) {
        task.remove_listeners(event_);
    }

    transwarp::event_type event_;
};

/// Visits the given task using the visitor given in the constructor
struct visit_visitor {
    explicit visit_visitor(const std::function<void(transwarp::itask&)>& visitor) noexcept
    : visitor_(visitor) {}

    void operator()(transwarp::itask& task) const {
        task.visit(visitor_);
    }

    const std::function<void(transwarp::itask&)>& visitor_;
};

/// Unvisits the given task
struct unvisit_visitor {

    void operator()(transwarp::itask& task) const noexcept {
        task.unvisit();
    }
};

/// Determines the result type of the Functor dispatching on the task type
template<typename TaskType, typename Functor, typename... ParentResults>
struct functor_result {
    static_assert(std::is_same<TaskType, transwarp::root_type>::value ||
                  std::is_same<TaskType, transwarp::accept_type>::value ||
                  std::is_same<TaskType, transwarp::accept_any_type>::value ||
                  std::is_same<TaskType, transwarp::consume_type>::value ||
                  std::is_same<TaskType, transwarp::consume_any_type>::value ||
                  std::is_same<TaskType, transwarp::wait_type>::value ||
                  std::is_same<TaskType, transwarp::wait_any_type>::value,
                  "Invalid task type, must be one of: root, accept, accept_any, consume, consume_any, wait, wait_any");
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::root_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) == 0, "A root task cannot have parent tasks");
    using type = decltype(std::declval<Functor>()());
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::accept_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "An accept task must have at least one parent");
    using type = decltype(std::declval<Functor>()(std::declval<std::shared_future<ParentResults>>()...));
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::accept_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()(std::declval<std::vector<std::shared_future<ParentResultType>>>()));
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::accept_any_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "An accept_any task must have at least one parent");
    using arg_t = typename std::tuple_element<0, std::tuple<ParentResults...>>::type; // Using first type as reference
    using type = decltype(std::declval<Functor>()(std::declval<std::shared_future<arg_t>>()));
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::accept_any_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()(std::declval<std::shared_future<ParentResultType>>()));
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::consume_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "A consume task must have at least one parent");
    using type = decltype(std::declval<Functor>()(std::declval<ParentResults>()...));
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::consume_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()(std::declval<std::vector<ParentResultType>>()));
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::consume_any_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "A consume_any task must have at least one parent");
    using arg_t = typename std::tuple_element<0, std::tuple<ParentResults...>>::type; // Using first type as reference
    using type = decltype(std::declval<Functor>()(std::declval<arg_t>()));
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::consume_any_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()(std::declval<ParentResultType>()));
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::wait_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "A wait task must have at least one parent");
    using type = decltype(std::declval<Functor>()());
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::wait_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()());
};

template<typename Functor, typename... ParentResults>
struct functor_result<transwarp::wait_any_type, Functor, ParentResults...> {
    static_assert(sizeof...(ParentResults) > 0, "A wait_any task must have at least one parent");
    using type = decltype(std::declval<Functor>()());
};

template<typename Functor, typename ParentResultType>
struct functor_result<transwarp::wait_any_type, Functor, std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = decltype(std::declval<Functor>()());
};


template<bool is_transwarp_functor>
struct assign_task_if_impl;

template<>
struct assign_task_if_impl<false> {
    template<typename Functor>
    void operator()(Functor&, transwarp::itask&) const noexcept {}
};

template<>
struct assign_task_if_impl<true> {
    template<typename Functor>
    void operator()(Functor& functor, transwarp::itask& task) const noexcept {
        functor.transwarp_task_ = &task;
    }
};

/// Assigns the task to the given functor if the functor is a subclass of transwarp::functor
template<typename Functor>
void assign_task_if(Functor& functor, transwarp::itask& task) noexcept {
    transwarp::detail::assign_task_if_impl<std::is_base_of<transwarp::functor, Functor>::value>{}(functor, task);
}


/// Returns a ready future with the given value as its state
template<typename ResultType, typename Value>
std::shared_future<ResultType> make_future_with_value(Value&& value) {
    std::promise<ResultType> promise;
    promise.set_value(std::forward<Value>(value));
    return promise.get_future();
}

/// Returns a ready future
inline
std::shared_future<void> make_ready_future() {
    std::promise<void> promise;
    promise.set_value();
    return promise.get_future();
}

/// Returns a ready future with the given exception as its state
template<typename ResultType>
std::shared_future<ResultType> make_future_with_exception(std::exception_ptr exception) {
    if (!exception) {
        throw transwarp::invalid_parameter{"exception pointer"};
    }
    std::promise<ResultType> promise;
    promise.set_exception(exception);
    return promise.get_future();
}


struct clone_task_functor {
    explicit clone_task_functor(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache) noexcept
    : task_cache_(task_cache) {}

    template<typename T>
    void operator()(T& t) {
        t = transwarp::detail::clone_task(task_cache_, t);
    }

    std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache_;
};


struct push_task_functor {
    explicit push_task_functor(std::vector<transwarp::itask*>& tasks) noexcept
    : tasks_(tasks) {}

    template<typename T>
    void operator()(T& t) {
        tasks_.push_back(t.get());
    }

    std::vector<transwarp::itask*>& tasks_;
};


/// Determines the type of the parents
template<typename... ParentResults>
struct parents {
    using type = std::tuple<std::shared_ptr<transwarp::task<ParentResults>>...>;
    static std::size_t size(const type&) {
        return std::tuple_size<type>::value;
    }
    static type clone(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const type& obj) {
        type cloned = obj;
        transwarp::detail::apply_to_each(transwarp::detail::clone_task_functor{task_cache}, cloned);
        return cloned;
    }
    static std::vector<transwarp::itask*> tasks(const type& parents) {
        std::vector<transwarp::itask*> tasks;
        transwarp::detail::apply_to_each(transwarp::detail::push_task_functor{tasks}, parents);
        return tasks;
    }
};

/// Determines the type of the parents. Specialization for vector parents
template<typename ParentResultType>
struct parents<std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>> {
    using type = std::vector<std::shared_ptr<transwarp::task<ParentResultType>>>;
    static std::size_t size(const type& obj) {
        return obj.size();
    }
    static type clone(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache, const type& obj) {
        type cloned = obj;
        transwarp::detail::apply_to_each(transwarp::detail::clone_task_functor{task_cache}, cloned);
        return cloned;
    }
    static std::vector<transwarp::itask*> tasks(const type& parents) {
        std::vector<transwarp::itask*> tasks;
        transwarp::detail::apply_to_each(transwarp::detail::push_task_functor{tasks}, parents);
        return tasks;
    }
};


template<typename ResultType, typename TaskType>
class base_runner {
protected:

    template<typename Task, typename Parents>
    void call(std::size_t task_id,
              const std::weak_ptr<Task>& task,
              const Parents& parents) {
        promise_.set_value(transwarp::detail::call<TaskType, ResultType>(task_id, task, parents));
    }

    std::promise<ResultType> promise_;
};

template<typename TaskType>
class base_runner<void, TaskType> {
protected:

    template<typename Task, typename Parents>
    void call(std::size_t task_id,
              const std::weak_ptr<Task>& task,
              const Parents& parents) {
        transwarp::detail::call<TaskType, void>(task_id, task, parents);
        promise_.set_value();
    }

    std::promise<void> promise_;
};

/// A callable to run a task given its parents
template<typename ResultType, typename TaskType, typename Task, typename Parents>
class runner : public transwarp::detail::base_runner<ResultType, TaskType> {
public:

    runner(std::size_t task_id,
           const std::weak_ptr<Task>& task,
           const typename transwarp::decay<Parents>::type& parents)
    : task_id_(task_id),
      task_(task),
      parents_(parents)
    {}

    std::future<ResultType> future() {
        return this->promise_.get_future();
    }

    void operator()() {
        if (const std::shared_ptr<Task> t = task_.lock()) {
            t->raise_event(transwarp::event_type::before_started);
        }
        try {
            this->call(task_id_, task_, parents_);
        } catch (const transwarp::task_canceled&) {
            this->promise_.set_exception(std::current_exception());
            if (const std::shared_ptr<Task> t = task_.lock()) {
                t->raise_event(transwarp::event_type::after_canceled);
            }
        } catch (...) {
            this->promise_.set_exception(std::current_exception());
        }
        if (const std::shared_ptr<Task> t = task_.lock()) {
            t->raise_event(transwarp::event_type::after_finished);
        }
    }

private:
    const std::size_t task_id_;
    const std::weak_ptr<Task> task_;
    const typename transwarp::decay<Parents>::type parents_;
};


/// A simple circular buffer (FIFO).
/// ValueType must support default construction. The buffer lets you push
/// new values onto the back and pop old values off the front.
template<typename ValueType>
class circular_buffer {
public:

    static_assert(std::is_default_constructible<ValueType>::value, "ValueType must be default constructible");

    using value_type = ValueType;

    /// Constructs a circular buffer with a given fixed capacity
    explicit
    circular_buffer(std::size_t capacity)
    : data_(capacity)
    {
        if (capacity < 1) {
            throw transwarp::invalid_parameter{"capacity"};
        }
    }

    // delete copy/move semantics
    circular_buffer(const circular_buffer&) = delete;
    circular_buffer& operator=(const circular_buffer&) = delete;
    circular_buffer(circular_buffer&& other) = delete;
    circular_buffer& operator=(circular_buffer&&) = delete;

    /// Pushes a new value onto the end of the buffer. If that exceeds the capacity
    /// of the buffer then the oldest value gets dropped (the one at the front).
    template<typename T, typename = typename std::enable_if<std::is_same<typename std::decay<T>::type, value_type>::value>::type>
    void push(T&& value) {
        data_[end_] = std::forward<T>(value);
        increment();
    }

    /// Returns the value at the front of the buffer (the oldest value).
    /// This is undefined if the buffer is empty
    const value_type& front() const {
        return data_[front_];
    }

    /// Removes the value at the front of the buffer (the oldest value)
    void pop() {
        if (!empty()) {
            data_[front_] = ValueType{};
            decrement();
        }
    }

    /// Returns the capacity of the buffer
    std::size_t capacity() const {
        return data_.size();
    }

    /// Returns the number of populated values of the buffer. Its maximum value
    /// equals the capacity of the buffer
    std::size_t size() const {
        return size_;
    }

    /// Returns whether the buffer is empty
    bool empty() const {
        return size_ == 0;
    }

    /// Returns whether the buffer is full
    bool full() const {
        return size_ == data_.size();
    }

    /// Swaps this buffer with the given buffer
    void swap(circular_buffer& buffer) {
        std::swap(end_, buffer.end_);
        std::swap(front_, buffer.front_);
        std::swap(size_, buffer.size_);
        std::swap(data_, buffer.data_);
    }

private:

    void increment_or_wrap(std::size_t& value) const {
        if (value == data_.size() - 1) {
            value = 0;
        } else {
            ++value;
        }
    }

    void increment() {
        increment_or_wrap(end_);
        if (full()) {
            increment_or_wrap(front_);
        } else {
            ++size_;
        }
    }

    void decrement() {
        increment_or_wrap(front_);
        --size_;
    }

    std::size_t end_{};
    std::size_t front_{};
    std::size_t size_{};
    std::vector<value_type> data_;
};


class spinlock {
public:

    void lock() noexcept {
        while (locked_.test_and_set(std::memory_order_acquire));
    }

    void unlock() noexcept {
        locked_.clear(std::memory_order_release);
    }

private:
    std::atomic_flag locked_ = ATOMIC_FLAG_INIT;
};


} // detail


/// A functor not doing nothing
struct no_op_functor {
    void operator()() const noexcept {}
};

/// An object to use in places where a no-op functor is required
constexpr no_op_functor no_op{};


/// Executor for sequential execution. Runs functors sequentially on the same thread
class sequential : public transwarp::executor {
public:

    sequential() = default;

    // delete copy/move semantics
    sequential(const sequential&) = delete;
    sequential& operator=(const sequential&) = delete;
    sequential(sequential&&) = delete;
    sequential& operator=(sequential&&) = delete;

    /// Returns the name of the executor
    std::string name() const override {
        return "transwarp::sequential";
    }

    /// Runs the functor on the current thread
    void execute(const std::function<void()>& functor, transwarp::itask&) override {
        functor();
    }
};


/// Executor for parallel execution. Uses a simple thread pool
class parallel : public transwarp::executor {
public:

    explicit parallel(const std::size_t n_threads,
                      std::function<void(std::size_t thread_index)> on_thread_started = nullptr)
    : pool_{n_threads, std::move(on_thread_started)}
    {}

    // delete copy/move semantics
    parallel(const parallel&) = delete;
    parallel& operator=(const parallel&) = delete;
    parallel(parallel&&) = delete;
    parallel& operator=(parallel&&) = delete;

    /// Returns the name of the executor
    std::string name() const override {
        return "transwarp::parallel";
    }

    /// Pushes the functor into the thread pool for asynchronous execution
    void execute(const std::function<void()>& functor, transwarp::itask&) override {
        pool_.push(functor);
    }

private:
    transwarp::detail::thread_pool pool_;
};


/// Detail namespace for internal functionality only
namespace detail {

const transwarp::option_str nullopt_string;
const transwarp::any_data any_empty;


template<typename ResultType, bool is_void>
struct make_future_functor;

template<typename ResultType>
struct make_future_functor<ResultType, true> {
    template<typename Future, typename OtherFuture>
    void operator()(Future& future, const OtherFuture& other) const {
        other.get();
        future = transwarp::detail::make_ready_future();
    }
};

template<typename ResultType>
struct make_future_functor<ResultType, false> {
    template<typename Future, typename OtherFuture>
    void operator()(Future& future, const OtherFuture& other) const {
        future = transwarp::detail::make_future_with_value<ResultType>(other.get());
    }
};


/// Common task functionality shared across `task_impl` and `value_task`
template<typename ResultType>
class task_common : public transwarp::task<ResultType> {
public:
    /// The result type of this task
    using result_type = ResultType;

    /// The task's id
    std::size_t id() const noexcept override {
        return id_;
    }

    /// The optional task name
    const transwarp::option_str& name() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_NAME
        return name_;
#else
        return transwarp::detail::nullopt_string;
#endif
    }

    /// The task priority (defaults to 0)
    std::int64_t priority() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        return priority_;
#else
        return 0;
#endif
    }

    /// The custom task data (may not hold a value)
    const transwarp::any_data& custom_data() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        return custom_data_;
#else
        return transwarp::detail::any_empty;
#endif
    }

    /// Sets a task priority (defaults to 0). transwarp will not directly use this.
    /// This is only useful if something else is using the priority (e.g. a custom executor)
    void set_priority(std::int64_t priority) override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        ensure_task_not_running();
        priority_ = priority;
#else
        (void)priority;
#endif
    }

    /// Resets the task priority to 0
    void reset_priority() override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        ensure_task_not_running();
        priority_ = 0;
#endif
    }

    /// Assigns custom data to this task. transwarp will not directly use this.
    /// This is only useful if something else is using this custom data (e.g. a custom executor)
    void set_custom_data(transwarp::any_data custom_data) override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        ensure_task_not_running();
        if (!custom_data.has_value()) {
            throw transwarp::invalid_parameter{"custom data"};
        }
        custom_data_ = std::move(custom_data);
        raise_event(transwarp::event_type::after_custom_data_set);
#else
        (void)custom_data;
#endif
    }

    /// Removes custom data from this task
    void remove_custom_data() override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        ensure_task_not_running();
        custom_data_ = {};
        raise_event(transwarp::event_type::after_custom_data_set);
#endif
    }

    /// Returns the future associated to the underlying execution
    std::shared_future<result_type> future() const noexcept override {
        return future_;
    }

    /// Adds a new listener for all event types
    void add_listener(std::shared_ptr<transwarp::listener> listener) override {
        ensure_task_not_running();
        check_listener(listener);
        ensure_listeners_object();
        for (int i=0; i<static_cast<int>(transwarp::event_type::count); ++i) {
            (*listeners_)[static_cast<transwarp::event_type>(i)].push_back(listener);
        }
    }

    /// Adds a new listener for the given event type only
    void add_listener(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener) override {
        ensure_task_not_running();
        check_listener(listener);
        ensure_listeners_object();
        (*listeners_)[event].push_back(std::move(listener));
    }

    /// Removes the listener for all event types
    void remove_listener(const std::shared_ptr<transwarp::listener>& listener) override {
        ensure_task_not_running();
        check_listener(listener);
        if (!listeners_) {
            return;
        }
        for (int i=0; i<static_cast<int>(transwarp::event_type::count); ++i) {
            auto listeners_pair = listeners_->find(static_cast<transwarp::event_type>(i));
            if (listeners_pair != listeners_->end()) {
                std::vector<std::shared_ptr<transwarp::listener>>& l = listeners_pair->second;
                l.erase(std::remove(l.begin(), l.end(), listener), l.end());
            }
        }
    }

    /// Removes the listener for the given event type only
    void remove_listener(transwarp::event_type event, const std::shared_ptr<transwarp::listener>& listener) override {
        ensure_task_not_running();
        check_listener(listener);
        if (!listeners_) {
            return;
        }
        auto listeners_pair = listeners_->find(event);
        if (listeners_pair != listeners_->end()) {
            std::vector<std::shared_ptr<transwarp::listener>>& l = listeners_pair->second;
            l.erase(std::remove(l.begin(), l.end(), listener), l.end());
        }
    }

    /// Removes all listeners
    void remove_listeners() override {
        ensure_task_not_running();
        if (!listeners_) {
            return;
        }
        listeners_->clear();
    }

    /// Removes all listeners for the given event type
    void remove_listeners(transwarp::event_type event) override {
        ensure_task_not_running();
        if (!listeners_) {
            return;
        }
        auto listeners_pair = listeners_->find(event);
        if (listeners_pair != listeners_->end()) {
            listeners_pair->second.clear();
        }
    }

protected:

    using listeners_t = std::map<transwarp::event_type, std::vector<std::shared_ptr<transwarp::listener>>>;
    using tasks_t = std::vector<transwarp::itask*>;

    /// Checks if the task is currently running and throws transwarp::control_error if it is
    void ensure_task_not_running() const {
        if (future_.valid() && future_.wait_for(std::chrono::seconds{0}) != std::future_status::ready) {
            throw transwarp::control_error{"task currently running: " + transwarp::to_string(*this, " ")};
        }
    }

    /// Raises the given event to all listeners
    void raise_event(transwarp::event_type event) {
        if (!listeners_) {
            return;
        }
        auto listeners_pair = listeners_->find(event);
        if (listeners_pair != listeners_->end()) {
            for (const std::shared_ptr<transwarp::listener>& listener : listeners_pair->second) {
                listener->handle_event(event, *this);
            }
        }
    }

    /// Check for non-null listener pointer
    void check_listener(const std::shared_ptr<transwarp::listener>& listener) const {
        if (!listener) {
            throw transwarp::invalid_parameter{"listener pointer"};
        }
    }

    void ensure_listeners_object() {
        if (!listeners_) {
            listeners_.reset(new listeners_t);
        }
    }

    /// Assigns the given id
    void set_id(std::size_t id) noexcept override {
        id_ = id;
    }

    /// Assigns the given name
    void set_name(transwarp::option_str name) noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_NAME
        name_ = std::move(name);
#else
        (void)name;
#endif
    }

    void copy_from(const task_common& task) {
        id_ = task.id_;
#ifndef TRANSWARP_DISABLE_TASK_NAME
        name_ = task.name_;
#endif
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        priority_ = task.priority_;
#endif
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        custom_data_ = task.custom_data_;
#endif
        if (task.has_result()) {
            try {
                make_future_functor<result_type, std::is_void<result_type>::value>{}(future_, task.future_);
            } catch (...) {
                future_ = transwarp::detail::make_future_with_exception<result_type>(std::current_exception());
            }
        }
        visited_ = task.visited_;
        if (task.listeners_) {
            listeners_.reset(new listeners_t(*task.listeners_));
        }
    }

    std::size_t id_ = 0;
#ifndef TRANSWARP_DISABLE_TASK_NAME
    transwarp::option_str name_;
#endif
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
    std::int64_t priority_ = 0;
#endif
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
    transwarp::any_data custom_data_;
#endif
    std::shared_future<result_type> future_;
    bool visited_ = false;
    std::unique_ptr<listeners_t> listeners_;
    std::unique_ptr<tasks_t> tasks_;
};


/// The base task class that contains the functionality that can be used
/// with all result types (void and non-void).
template<typename ResultType, typename TaskType, typename Functor, typename... ParentResults>
class task_impl_base : public transwarp::detail::task_common<ResultType> {
public:
    /// The task type
    using task_type = TaskType;

    /// The result type of this task
    using result_type = ResultType;

    /// Can be called to explicitly finalize this task making this task
    /// the terminal task in the graph. This is also done implicitly when
    /// calling, e.g., any of the *_all methods. It should normally not be
    /// necessary to call this method directly
    void finalize() override {
        if (!this->tasks_) {
            this->tasks_.reset(new typename transwarp::detail::task_common<result_type>::tasks_t);
            visit(transwarp::detail::final_visitor{*this->tasks_});
            unvisit();
            auto compare = [](const transwarp::itask* const l, const transwarp::itask* const r) {
                const std::size_t l_level = l->level();
                const std::size_t l_id = l->id();
                const std::size_t r_level = r->level();
                const std::size_t r_id = r->id();
                return std::tie(l_level, l_id) < std::tie(r_level, r_id);
            };
            std::sort(this->tasks_->begin(), this->tasks_->end(), compare);
        }
    }

    /// The task's level
    std::size_t level() const noexcept override {
        return level_;
    }

    /// The task's type
    transwarp::task_type type() const noexcept override {
        return type_;
    }

    /// The task's executor (may be null)
    std::shared_ptr<transwarp::executor> executor() const noexcept override {
        return executor_;
    }

    /// Returns whether the associated task is canceled
    bool canceled() const noexcept override {
        return canceled_.load();
    }

    /// Returns the average idletime in microseconds (-1 if never set)
    std::int64_t avg_idletime_us() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        return avg_idletime_us_.load();
#else
        return -1;
#endif
    }

    /// Returns the average waittime in microseconds (-1 if never set)
    std::int64_t avg_waittime_us() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        return avg_waittime_us_.load();
#else
        return -1;
#endif
    }

    /// Returns the average runtime in microseconds (-1 if never set)
    std::int64_t avg_runtime_us() const noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        return avg_runtime_us_.load();
#else
        return -1;
#endif
    }

    /// Assigns an executor to this task which takes precedence over
    /// the executor provided in schedule() or schedule_all()
    void set_executor(std::shared_ptr<transwarp::executor> executor) override {
        this->ensure_task_not_running();
        if (!executor) {
            throw transwarp::invalid_parameter{"executor pointer"};
        }
        executor_ = std::move(executor);
    }

    /// Assigns an executor to all tasks which takes precedence over
    /// the executor provided in schedule() or schedule_all()
    void set_executor_all(std::shared_ptr<transwarp::executor> executor) override {
        this->ensure_task_not_running();
        transwarp::detail::set_executor_visitor visitor{std::move(executor)};
        visit_all(visitor);
    }

    /// Removes the executor from this task
    void remove_executor() override {
        this->ensure_task_not_running();
        executor_.reset();
    }

    /// Removes the executor from all tasks
    void remove_executor_all() override {
        this->ensure_task_not_running();
        transwarp::detail::remove_executor_visitor visitor;
        visit_all(visitor);
    }

    /// Schedules this task for execution on the caller thread.
    /// The task-specific executor gets precedence if it exists.
    /// This overload will reset the underlying future.
    void schedule() override {
        this->ensure_task_not_running();
        this->schedule_impl(true);
    }

    /// Schedules this task for execution on the caller thread.
    /// The task-specific executor gets precedence if it exists.
    /// reset denotes whether schedule should reset the underlying
    /// future and schedule even if the future is already valid.
    void schedule(bool reset) override {
        this->ensure_task_not_running();
        this->schedule_impl(reset);
    }

    /// Schedules this task for execution using the provided executor.
    /// The task-specific executor gets precedence if it exists.
    /// This overload will reset the underlying future.
    void schedule(transwarp::executor& executor) override {
        this->ensure_task_not_running();
        this->schedule_impl(true, &executor);
    }

    /// Schedules this task for execution using the provided executor.
    /// The task-specific executor gets precedence if it exists.
    /// reset denotes whether schedule should reset the underlying
    /// future and schedule even if the future is already valid.
    void schedule(transwarp::executor& executor, bool reset) override {
        this->ensure_task_not_running();
        this->schedule_impl(reset, &executor);
    }

    /// Schedules all tasks in the graph for execution on the caller thread.
    /// The task-specific executors get precedence if they exist.
    /// This overload will reset the underlying futures.
    void schedule_all() override {
        this->ensure_task_not_running();
        schedule_all_impl(true);
    }

    /// Schedules all tasks in the graph for execution using the provided executor.
    /// The task-specific executors get precedence if they exist.
    /// This overload will reset the underlying futures.
    void schedule_all(transwarp::executor& executor) override {
        this->ensure_task_not_running();
        schedule_all_impl(true, &executor);
    }

    /// Schedules all tasks in the graph for execution on the caller thread.
    /// The task-specific executors get precedence if they exist.
    /// reset_all denotes whether schedule_all should reset the underlying
    /// futures and schedule even if the futures are already present.
    void schedule_all(bool reset_all) override {
        this->ensure_task_not_running();
        schedule_all_impl(reset_all);
    }

    /// Schedules all tasks in the graph for execution using the provided executor.
    /// The task-specific executors get precedence if they exist.
    /// reset_all denotes whether schedule_all should reset the underlying
    /// futures and schedule even if the futures are already present.
    void schedule_all(transwarp::executor& executor, bool reset_all) override {
        this->ensure_task_not_running();
        schedule_all_impl(reset_all, &executor);
    }

    /// Assigns an exception to this task. Scheduling will have no effect after an exception
    /// has been set. Calling reset() will remove the exception and re-enable scheduling.
    void set_exception(std::exception_ptr exception) override {
        this->ensure_task_not_running();
        this->future_ = transwarp::detail::make_future_with_exception<result_type>(exception);
        schedule_mode_ = false;
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Returns whether the task was scheduled and not reset afterwards.
    /// This means that the underlying future is valid
    bool was_scheduled() const noexcept override {
        return this->future_.valid();
    }

    /// Waits for the task to complete. Should only be called if was_scheduled()
    /// is true, throws transwarp::control_error otherwise
    void wait() const override {
        ensure_task_was_scheduled();
        this->future_.wait();
    }

    /// Returns whether the task has finished processing. Should only be called
    /// if was_scheduled() is true, throws transwarp::control_error otherwise
    bool is_ready() const override {
        ensure_task_was_scheduled();
        return this->future_.wait_for(std::chrono::seconds(0)) == std::future_status::ready;
    }

    /// Returns whether this task contains a result
    bool has_result() const noexcept override {
        return was_scheduled() && this->future_.wait_for(std::chrono::seconds(0)) == std::future_status::ready;
    }

    /// Resets this task
    void reset() override {
        this->ensure_task_not_running();
        this->future_ = std::shared_future<result_type>{};
        cancel(false);
        schedule_mode_ = true;
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
        refcount_ = childcount_;
#endif
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Resets all tasks in the graph
    void reset_all() override {
        this->ensure_task_not_running();
        transwarp::detail::reset_visitor visitor;
        visit_all(visitor);
    }

    /// If enabled then this task is canceled which will
    /// throw transwarp::task_canceled when retrieving the task result.
    /// Passing false is equivalent to resume.
    void cancel(bool enabled) noexcept override {
        canceled_ = enabled;
    }

    /// If enabled then all pending tasks in the graph are canceled which will
    /// throw transwarp::task_canceled when retrieving the task result.
    /// Passing false is equivalent to resume.
    void cancel_all(bool enabled) noexcept override {
        transwarp::detail::cancel_visitor visitor{enabled};
        visit_all(visitor);
    }

    /// Sets a priority to all tasks (defaults to 0). transwarp will not directly use this.
    /// This is only useful if something else is using the priority (e.g. a custom executor)
    void set_priority_all(std::int64_t priority) override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        this->ensure_task_not_running();
        transwarp::detail::set_priority_visitor visitor{priority};
        visit_all(visitor);
#else
        (void)priority;
#endif
    }

    /// Resets the priority of all tasks to 0
    void reset_priority_all() override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        this->ensure_task_not_running();
        transwarp::detail::reset_priority_visitor visitor;
        visit_all(visitor);
#endif
    }

    /// Assigns custom data to all tasks. transwarp will not directly use this.
    /// This is only useful if something else is using this custom data (e.g. a custom executor)
    void set_custom_data_all(transwarp::any_data custom_data) override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        this->ensure_task_not_running();
        transwarp::detail::set_custom_data_visitor visitor{std::move(custom_data)};
        visit_all(visitor);
#else
        (void)custom_data;
#endif
    }

    /// Removes custom data from all tasks
    void remove_custom_data_all() override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        this->ensure_task_not_running();
        transwarp::detail::remove_custom_data_visitor visitor;
        visit_all(visitor);
#endif
    }

    /// Adds a new listener for all event types and for all parents
    void add_listener_all(std::shared_ptr<transwarp::listener> listener) override {
        this->ensure_task_not_running();
        transwarp::detail::add_listener_visitor visitor{std::move(listener)};
        visit_all(visitor);
    }

    /// Adds a new listener for the given event type only and for all parents
    void add_listener_all(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener) override {
        this->ensure_task_not_running();
        transwarp::detail::add_listener_per_event_visitor visitor{event, std::move(listener)};
        visit_all(visitor);
    }

    /// Removes the listener for all event types and for all parents
    void remove_listener_all(const std::shared_ptr<transwarp::listener>& listener) override {
        this->ensure_task_not_running();
        transwarp::detail::remove_listener_visitor visitor{std::move(listener)};
        visit_all(visitor);
    }

    /// Removes the listener for the given event type only and for all parents
    void remove_listener_all(transwarp::event_type event, const std::shared_ptr<transwarp::listener>& listener) override {
        this->ensure_task_not_running();
        transwarp::detail::remove_listener_per_event_visitor visitor{event, std::move(listener)};
        visit_all(visitor);
    }

    /// Removes all listeners and for all parents
    void remove_listeners_all() override {
        this->ensure_task_not_running();
        transwarp::detail::remove_listeners_visitor visitor;
        visit_all(visitor);
    }

    /// Removes all listeners for the given event type and for all parents
    void remove_listeners_all(transwarp::event_type event) override {
        this->ensure_task_not_running();
        transwarp::detail::remove_listeners_per_event_visitor visitor{event};
        visit_all(visitor);
    }

    /// Returns the task's parents (may be empty)
    std::vector<transwarp::itask*> parents() const override {
        return transwarp::detail::parents<ParentResults...>::tasks(parents_);
    }

    /// Returns all tasks in the graph in breadth order
    const std::vector<transwarp::itask*>& tasks() override {
        finalize();
        return *this->tasks_;
    }

    /// Returns all edges in the graph. This is mainly for visualizing
    /// the tasks and their interdependencies. Pass the result into transwarp::to_string
    /// to retrieve a dot-style graph representation for easy viewing.
    std::vector<transwarp::edge> edges() override {
        std::vector<transwarp::edge> edges;
        transwarp::detail::edges_visitor visitor{edges};
        visit_all(visitor);
        return edges;
    }

protected:

    task_impl_base() {}

    template<typename F>
    task_impl_base(F&& functor, std::shared_ptr<transwarp::task<ParentResults>>... parents)
    : functor_(new Functor(std::forward<F>(functor))),
      parents_(std::move(parents)...)
    {
        init();
    }

    template<typename F, typename P>
    task_impl_base(F&& functor, std::vector<std::shared_ptr<transwarp::task<P>>> parents)
    : functor_(new Functor(std::forward<F>(functor))),
      parents_(std::move(parents))
    {
        init();
        if (parents_.empty()) {
            set_type(transwarp::task_type::root);
        }
    }

    void init() {
        set_type(task_type::value);
        transwarp::detail::assign_task_if(*functor_, *this);
        transwarp::detail::call_with_each(transwarp::detail::parent_visitor{*this}, parents_);
    }

    template<typename R, typename Y, typename T, typename P>
    friend class transwarp::detail::runner;

    template<typename R, typename T, typename... A>
    friend R transwarp::detail::run_task(std::size_t, const std::weak_ptr<T>&, A&&...);

    /// Assigns the given level
    void set_level(std::size_t level) noexcept override {
        level_ = level;
    }

    /// Assigns the given type
    void set_type(transwarp::task_type type) noexcept override {
        type_ = type;
    }

    /// Assigns the given idletime
    void set_avg_idletime_us(std::int64_t idletime) noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        avg_idletime_us_ = idletime;
#else
        (void)idletime;
#endif
    }

    /// Assigns the given waittime
    void set_avg_waittime_us(std::int64_t waittime) noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        avg_waittime_us_ = waittime;
#else
        (void)waittime;
#endif
    }

    /// Assigns the given runtime
    void set_avg_runtime_us(std::int64_t runtime) noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_TIME
        avg_runtime_us_ = runtime;
#else
        (void)runtime;
#endif
    }

    /// Checks if the task was scheduled and throws transwarp::control_error if it's not
    void ensure_task_was_scheduled() const {
        if (!this->future_.valid()) {
            throw transwarp::control_error{"task was not scheduled: " + transwarp::to_string(*this, " ")};
        }
    }

    /// Schedules this task for execution using the provided executor.
    /// The task-specific executor gets precedence if it exists.
    /// Runs the task on the same thread as the caller if neither the global
    /// nor the task-specific executor is found.
    void schedule_impl(bool reset, transwarp::executor* executor=nullptr) override {
        if (schedule_mode_ && (reset || !this->future_.valid())) {
            if (reset) {
                cancel(false);
            }
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
            refcount_ = childcount_;
#endif
            std::weak_ptr<task_impl_base> self = std::static_pointer_cast<task_impl_base>(this->shared_from_this());
            using runner_t = transwarp::detail::runner<result_type, task_type, task_impl_base, decltype(parents_)>;
            std::shared_ptr<runner_t> runner = std::shared_ptr<runner_t>(new runner_t(this->id(), self, parents_));
            this->raise_event(transwarp::event_type::before_scheduled);
            this->future_ = runner->future();
            this->raise_event(transwarp::event_type::after_future_changed);
            if (this->executor_) {
                this->executor_->execute([runner]{ (*runner)(); }, *this);
            } else if (executor) {
                executor->execute([runner]{ (*runner)(); }, *this);
            } else {
                (*runner)();
            }
        }
    }

    /// Schedules all tasks in the graph for execution using the provided executor.
    /// The task-specific executors get precedence if they exist.
    /// Runs tasks on the same thread as the caller if neither the global
    /// nor a task-specific executor is found.
    void schedule_all_impl(bool reset_all, transwarp::executor* executor=nullptr) {
        transwarp::detail::schedule_visitor visitor{reset_all, executor};
        visit_all(visitor);
    }

    /// Visits each task in a depth-first traversal
    void visit(const std::function<void(transwarp::itask&)>& visitor) override {
        if (!this->visited_) {
            transwarp::detail::call_with_each(transwarp::detail::visit_visitor{visitor}, parents_);
            visitor(*this);
            this->visited_ = true;
        }
    }

    /// Traverses through each task and marks them as not visited.
    void unvisit() noexcept override {
        if (this->visited_) {
            this->visited_ = false;
            transwarp::detail::call_with_each(transwarp::detail::unvisit_visitor{}, parents_);
        }
    }

    /// Visits all tasks
    template<typename Visitor>
    void visit_all(Visitor& visitor) {
        finalize();
        for (transwarp::itask* t : *this->tasks_) {
            visitor(*t);
        }
    }

    void increment_childcount() noexcept override {
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
        ++childcount_;
#endif
    }

    void decrement_refcount() override {
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
        if (--refcount_ == 0) {
            this->raise_event(transwarp::event_type::after_satisfied);
        }
#endif
    }

    void reset_future() override {
        this->future_ = std::shared_future<result_type>{};
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    std::size_t level_ = 0;
    transwarp::task_type type_ = transwarp::task_type::root;
    std::shared_ptr<transwarp::executor> executor_;
    std::atomic<bool> canceled_{false};
    bool schedule_mode_ = true;
#ifndef TRANSWARP_DISABLE_TASK_TIME
    std::atomic<std::int64_t> avg_idletime_us_{-1};
    std::atomic<std::int64_t> avg_waittime_us_{-1};
    std::atomic<std::int64_t> avg_runtime_us_{-1};
#endif
    std::unique_ptr<Functor> functor_;
    typename transwarp::detail::parents<ParentResults...>::type parents_;
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
    std::size_t childcount_ = 0;
    std::atomic<std::size_t> refcount_{0};
#endif
};


/// A task proxy
template<typename ResultType, typename TaskType, typename Functor, typename... ParentResults>
class task_impl_proxy : public transwarp::detail::task_impl_base<ResultType, TaskType, Functor, ParentResults...> {
public:
    /// The task type
    using task_type = TaskType;

    /// The result type of this task
    using result_type = ResultType;

    /// Assigns a value to this task. Scheduling will have no effect after a value
    /// has been set. Calling reset() will remove the value and re-enable scheduling.
    void set_value(const typename transwarp::decay<result_type>::type& value) override {
        this->ensure_task_not_running();
        this->future_ = transwarp::detail::make_future_with_value<result_type>(value);
        this->schedule_mode_ = false;
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Assigns a value to this task. Scheduling will have no effect after a value
    /// has been set. Calling reset() will remove the value and re-enable scheduling.
    void set_value(typename transwarp::decay<result_type>::type&& value) override {
        this->ensure_task_not_running();
        this->future_ = transwarp::detail::make_future_with_value<result_type>(std::move(value));
        this->schedule_mode_ = false;
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Returns the result of this task. Throws any exceptions that the underlying
    /// functor throws. Should only be called if was_scheduled() is true,
    /// throws transwarp::control_error otherwise
    typename transwarp::result<result_type>::type get() const override {
        this->ensure_task_was_scheduled();
        return this->future_.get();
    }

protected:

    task_impl_proxy() = default;

    template<typename F>
    task_impl_proxy(F&& functor, std::shared_ptr<transwarp::task<ParentResults>>... parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents)...)
    {}

    template<typename F, typename P>
    task_impl_proxy(F&& functor, std::vector<std::shared_ptr<transwarp::task<P>>> parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents))
    {}

};

/// A task proxy for reference result type.
template<typename ResultType, typename TaskType, typename Functor, typename... ParentResults>
class task_impl_proxy<ResultType&, TaskType, Functor, ParentResults...> : public transwarp::detail::task_impl_base<ResultType&, TaskType, Functor, ParentResults...> {
public:
    /// The task type
    using task_type = TaskType;

    /// The result type of this task
    using result_type = ResultType&;

    /// Assigns a value to this task. Scheduling will have no effect after a value
    /// has been set. Calling reset() will remove the value and re-enable scheduling.
    void set_value(typename transwarp::decay<result_type>::type& value) override {
        this->ensure_task_not_running();
        this->future_ = transwarp::detail::make_future_with_value<result_type>(value);
        this->schedule_mode_ = false;
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Returns the result of this task. Throws any exceptions that the underlying
    /// functor throws. Should only be called if was_scheduled() is true,
    /// throws transwarp::control_error otherwise
    typename transwarp::result<result_type>::type get() const override {
        this->ensure_task_was_scheduled();
        return this->future_.get();
    }

protected:

    task_impl_proxy() = default;

    template<typename F>
    task_impl_proxy(F&& functor, std::shared_ptr<transwarp::task<ParentResults>>... parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents)...)
    {}

    template<typename F, typename P>
    task_impl_proxy(F&& functor, std::vector<std::shared_ptr<transwarp::task<P>>> parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents))
    {}

};

/// A task proxy for void result type.
template<typename TaskType, typename Functor, typename... ParentResults>
class task_impl_proxy<void, TaskType, Functor, ParentResults...> : public transwarp::detail::task_impl_base<void, TaskType, Functor, ParentResults...> {
public:
    /// The task type
    using task_type = TaskType;

    /// The result type of this task
    using result_type = void;

    /// Assigns a value to this task. Scheduling will have no effect after a call
    /// to this. Calling reset() will reset this and re-enable scheduling.
    void set_value() override {
        this->ensure_task_not_running();
        this->future_ = transwarp::detail::make_ready_future();
        this->schedule_mode_ = false;
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Blocks until the task finishes. Throws any exceptions that the underlying
    /// functor throws. Should only be called if was_scheduled() is true,
    /// throws transwarp::control_error otherwise
    void get() const override {
        this->ensure_task_was_scheduled();
        this->future_.get();
    }

protected:

    task_impl_proxy() = default;

    template<typename F>
    task_impl_proxy(F&& functor, std::shared_ptr<transwarp::task<ParentResults>>... parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents)...)
    {}

    template<typename F, typename P>
    task_impl_proxy(F&& functor, std::vector<std::shared_ptr<transwarp::task<P>>> parents)
    : transwarp::detail::task_impl_base<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents))
    {}

};

} // detail


/// A task representing a piece of work given by functor and parent tasks.
/// By connecting tasks a directed acyclic graph is built.
/// Tasks should be created using the make_task factory functions.
template<typename TaskType, typename Functor, typename... ParentResults>
class task_impl : public transwarp::detail::task_impl_proxy<typename transwarp::detail::functor_result<TaskType, Functor, ParentResults...>::type, TaskType, Functor, ParentResults...> {
public:
    /// The task type
    using task_type = TaskType;

    /// The result type of this task
    using result_type = typename transwarp::detail::functor_result<TaskType, Functor, ParentResults...>::type;

    /// A task is defined by functor and parent tasks.
    /// Note: Don't use this constructor directly, use transwarp::make_task
    template<typename F>
    task_impl(F&& functor, std::shared_ptr<transwarp::task<ParentResults>>... parents)
    : transwarp::detail::task_impl_proxy<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents)...)
    {}

    /// A task is defined by functor and parent tasks.
    /// Note: Don't use this constructor directly, use transwarp::make_task
    template<typename F, typename P>
    task_impl(F&& functor, std::vector<std::shared_ptr<transwarp::task<P>>> parents)
    : transwarp::detail::task_impl_proxy<result_type, task_type, Functor, ParentResults...>(std::forward<F>(functor), std::move(parents))
    {}

    // delete copy/move semantics
    task_impl(const task_impl&) = delete;
    task_impl& operator=(const task_impl&) = delete;
    task_impl(task_impl&&) = delete;
    task_impl& operator=(task_impl&&) = delete;

    /// Gives this task a name and returns a ptr to itself
    std::shared_ptr<task_impl> named(std::string name) {
#ifndef TRANSWARP_DISABLE_TASK_NAME
#ifdef TRANSWARP_CPP11
        this->set_name(transwarp::option_str{std::move(name)});
#else
        this->set_name(std::make_optional(std::move(name)));
#endif
#else
        (void)name;
#endif
        return std::static_pointer_cast<task_impl>(this->shared_from_this());
    }

    /// Creates a continuation to this task
    template<typename TaskType_, typename Functor_>
    auto then(TaskType_, Functor_&& functor) -> std::shared_ptr<transwarp::task_impl<TaskType_, typename std::decay<Functor_>::type, result_type>> {
        using task_t = transwarp::task_impl<TaskType_, typename std::decay<Functor_>::type, result_type>;
        return std::shared_ptr<task_t>(new task_t(std::forward<Functor_>(functor), std::static_pointer_cast<task_impl>(this->shared_from_this())));
    }

    /// Clones this task and casts the result to a ptr to task_impl
    std::shared_ptr<task_impl> clone_cast() const {
        return std::static_pointer_cast<task_impl>(this->clone());
    }

private:

    task_impl() = default;

    std::shared_ptr<transwarp::task<result_type>> clone_impl(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>& task_cache) const override {
        auto t = std::shared_ptr<task_impl>{new task_impl};
        t->copy_from(*this);
        t->level_ = this->level_;
        t->type_ = this->type_;
        t->executor_ = this->executor_;
        t->canceled_ = this->canceled_.load();
        t->schedule_mode_ = this->schedule_mode_;
#ifndef TRANSWARP_DISABLE_TASK_TIME
        t->avg_idletime_us_ = this->avg_idletime_us_.load();
        t->avg_waittime_us_ = this->avg_waittime_us_.load();
        t->avg_runtime_us_ = this->avg_runtime_us_.load();
#endif
        t->functor_.reset(new Functor(*this->functor_));
        t->parents_ = transwarp::detail::parents<ParentResults...>::clone(task_cache, this->parents_);
        t->executor_ = this->executor_;
#ifndef TRANSWARP_DISABLE_TASK_REFCOUNT
        t->childcount_ = this->childcount_;
#endif
        return t;
    }

};


/// A value task that stores a single value and doesn't require scheduling.
/// Value tasks should be created using the make_value_task factory functions.
template<typename ResultType>
class value_task : public transwarp::detail::task_common<ResultType> {
public:
    /// The task type
    using task_type = transwarp::root_type;

    /// The result type of this task
    using result_type = ResultType;

    /// A value task is defined by a given value.
    /// Note: Don't use this constructor directly, use transwarp::make_value_task
    template<typename T>
    value_task(T&& value)
    {
        this->future_ = transwarp::detail::make_future_with_value<result_type>(std::forward<T>(value));
        this->tasks_.reset(new typename transwarp::detail::task_common<result_type>::tasks_t{this});
    }

    // delete copy/move semantics
    value_task(const value_task&) = delete;
    value_task& operator=(const value_task&) = delete;
    value_task(value_task&&) = delete;
    value_task& operator=(value_task&&) = delete;

    /// Gives this task a name and returns a ptr to itself
    std::shared_ptr<value_task> named(std::string name) {
#ifndef TRANSWARP_DISABLE_TASK_NAME
#ifdef TRANSWARP_CPP11
        this->set_name(transwarp::option_str{std::move(name)});
#else
        this->set_name(std::make_optional(std::move(name)));
#endif
#else
        (void)name;
#endif
        return std::static_pointer_cast<value_task>(this->shared_from_this());
    }

    /// Creates a continuation to this task
    template<typename TaskType_, typename Functor_>
    auto then(TaskType_, Functor_&& functor) -> std::shared_ptr<transwarp::task_impl<TaskType_, typename std::decay<Functor_>::type, result_type>> {
        using task_t = transwarp::task_impl<TaskType_, typename std::decay<Functor_>::type, result_type>;
        return std::shared_ptr<task_t>(new task_t(std::forward<Functor_>(functor), std::static_pointer_cast<value_task>(this->shared_from_this())));
    }

    /// Clones this task and casts the result to a ptr to value_task
    std::shared_ptr<value_task> clone_cast() const {
        return std::static_pointer_cast<value_task>(this->clone());
    }

    /// Nothing to be done to finalize a value task
    void finalize() override {}

    /// The task's level
    std::size_t level() const noexcept override {
        return 0;
    }

    /// The task's type
    transwarp::task_type type() const noexcept override {
        return transwarp::task_type::root;
    }

    /// Value tasks don't have executors as they don't run
    std::shared_ptr<transwarp::executor> executor() const noexcept override {
        return nullptr;
    }

    /// Value tasks cannot be canceled
    bool canceled() const noexcept override {
        return false;
    }

    /// Returns -1 as value tasks don't run
    std::int64_t avg_idletime_us() const noexcept override {
        return -1;
    }

    /// Returns -1 as value tasks don't run
    std::int64_t avg_waittime_us() const noexcept override {
        return -1;
    }

    /// Returns -1 as value tasks don't run
    std::int64_t avg_runtime_us() const noexcept override {
        return -1;
    }

    /// No-op because a value task never runs
    void set_executor(std::shared_ptr<transwarp::executor>) override {}

    /// No-op because a value task never runs and doesn't have parents
    void set_executor_all(std::shared_ptr<transwarp::executor>) override {}

    /// No-op because a value task never runs
    void remove_executor() override {}

    /// No-op because a value task never runs and doesn't have parents
    void remove_executor_all() override {}

    /// Sets a priority to all tasks (defaults to 0). transwarp will not directly use this.
    /// This is only useful if something else is using the priority
    void set_priority_all(std::int64_t priority) override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        this->set_priority(priority);
#else
        (void)priority;
#endif
    }

    /// Resets the priority of all tasks to 0
    void reset_priority_all() override {
#ifndef TRANSWARP_DISABLE_TASK_PRIORITY
        this->reset_priority();
#endif
    }

    /// Assigns custom data to all tasks. transwarp will not directly use this.
    /// This is only useful if something else is using this custom data
    void set_custom_data_all(transwarp::any_data custom_data) override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        this->set_custom_data(std::move(custom_data));
#else
        (void)custom_data;
#endif
    }

    /// Removes custom data from all tasks
    void remove_custom_data_all() override {
#ifndef TRANSWARP_DISABLE_TASK_CUSTOM_DATA
        this->remove_custom_data();
#endif
    }

    /// No-op because a value task never runs
    void schedule() override {}

    /// No-op because a value task never runs
    void schedule(transwarp::executor&) override {}

    /// No-op because a value task never runs
    void schedule(bool) override {}

    /// No-op because a value task never runs
    void schedule(transwarp::executor&, bool) override {}

    /// No-op because a value task never runs and doesn't have parents
    void schedule_all() override {}

    /// No-op because a value task never runs and doesn't have parents
    void schedule_all(transwarp::executor&) override {}

    /// No-op because a value task never runs and doesn't have parents
    void schedule_all(bool) override {}

    /// No-op because a value task never runs and doesn't have parents
    void schedule_all(transwarp::executor&, bool) override {}

    /// Assigns a value to this task
    void set_value(const typename transwarp::decay<result_type>::type& value) override {
        this->future_ = transwarp::detail::make_future_with_value<result_type>(value);
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Assigns a value to this task
    void set_value(typename transwarp::decay<result_type>::type&& value) override {
        this->future_ = transwarp::detail::make_future_with_value<result_type>(std::move(value));
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Assigns an exception to this task
    void set_exception(std::exception_ptr exception) override {
        this->future_ = transwarp::detail::make_future_with_exception<result_type>(exception);
        this->raise_event(transwarp::event_type::after_future_changed);
    }

    /// Returns the value of this task. Throws an exception if this task has an exception assigned to it
    typename transwarp::result<result_type>::type get() const override {
        return this->future_.get();
    }

    /// Returns true because a value task is scheduled once on construction
    bool was_scheduled() const noexcept override {
        return true;
    }

    /// No-op because a value task never runs
    void wait() const override {}

    /// Returns true because a value task is always ready
    bool is_ready() const override {
        return true;
    }

    /// Returns true because a value task always contains a result
    bool has_result() const noexcept override {
        return true;
    }

    /// No-op because a value task never runs
    void reset() override {}

    /// No-op because a value task never runs and doesn't have parents
    void reset_all() override {}

    /// No-op because a value task never runs
    void cancel(bool) noexcept override {}

    /// No-op because a value task never runs and doesn't have parents
    void cancel_all(bool) noexcept override {}

    /// Adds a new listener for all event types and for all parents
    void add_listener_all(std::shared_ptr<transwarp::listener> listener) override {
        this->add_listener(listener);
    }

    /// Adds a new listener for the given event type only and for all parents
    void add_listener_all(transwarp::event_type event, std::shared_ptr<transwarp::listener> listener) override {
        this->add_listener(event, listener);
    }

    /// Removes the listener for all event types and for all parents
    void remove_listener_all(const std::shared_ptr<transwarp::listener>& listener) override {
        this->remove_listener(listener);
    }

    /// Removes the listener for the given event type only and for all parents
    void remove_listener_all(transwarp::event_type event, const std::shared_ptr<transwarp::listener>& listener) override {
        this->remove_listener(event, listener);
    }

    /// Removes all listeners and for all parents
    void remove_listeners_all() override {
        this->remove_listeners();
    }

    /// Removes all listeners for the given event type and for all parents
    void remove_listeners_all(transwarp::event_type event) override {
        this->remove_listeners(event);
    }

    /// Empty because a value task doesn't have parents
    std::vector<transwarp::itask*> parents() const override {
        return {};
    }

    /// Returns all tasks in the graph in breadth order
    const std::vector<transwarp::itask*>& tasks() override {
        return *this->tasks_;
    }

    /// Returns empty edges because a value task doesn't have parents
    std::vector<transwarp::edge> edges() override {
        return {};
    }

private:

    value_task()
    {
        this->tasks_.reset(new typename transwarp::detail::task_common<result_type>::tasks_t{this});
    }

    std::shared_ptr<transwarp::task<result_type>> clone_impl(std::unordered_map<std::shared_ptr<transwarp::itask>, std::shared_ptr<transwarp::itask>>&) const override {
        auto t = std::shared_ptr<value_task>(new value_task);
        t->copy_from(*this);
        return t;
    }

    /// No-op as value tasks are always at level 0
    void set_level(std::size_t) noexcept override {}

    /// No-op as value tasks are always root tasks
    void set_type(transwarp::task_type) noexcept override {}

    /// No-op as value tasks don't run
    void set_avg_idletime_us(std::int64_t) noexcept override {}

    /// No-op as value tasks don't run
    void set_avg_waittime_us(std::int64_t) noexcept override {}

    /// No-op as value tasks don't run
    void set_avg_runtime_us(std::int64_t) noexcept override {}

    /// No-op because a value task never runs
    void schedule_impl(bool, transwarp::executor*) override {}

    /// Visits this task
    void visit(const std::function<void(transwarp::itask&)>& visitor) override {
        if (!this->visited_) {
            visitor(*this);
            this->visited_ = true;
        }
    }

    /// Marks this task as not visited
    void unvisit() noexcept override {
        this->visited_ = false;
    }

    void increment_childcount() noexcept override {}

    void decrement_refcount() override {}

    void reset_future() override {}

};


/// A factory function to create a new task
template<typename TaskType, typename Functor, typename... Parents>
auto make_task(TaskType, Functor&& functor, std::shared_ptr<Parents>... parents) -> std::shared_ptr<transwarp::task_impl<TaskType, typename std::decay<Functor>::type, typename Parents::result_type...>> {
    using task_t = transwarp::task_impl<TaskType, typename std::decay<Functor>::type, typename Parents::result_type...>;
    return std::shared_ptr<task_t>(new task_t(std::forward<Functor>(functor), std::move(parents)...));
}


/// A factory function to create a new task with vector parents
template<typename TaskType, typename Functor, typename ParentType>
auto make_task(TaskType, Functor&& functor, std::vector<ParentType> parents) -> std::shared_ptr<transwarp::task_impl<TaskType, typename std::decay<Functor>::type, std::vector<ParentType>>> {
    using task_t = transwarp::task_impl<TaskType, typename std::decay<Functor>::type, std::vector<ParentType>>;
    return std::shared_ptr<task_t>(new task_t(std::forward<Functor>(functor), std::move(parents)));
}


/// A factory function to create a new value task
template<typename Value>
auto make_value_task(Value&& value) -> std::shared_ptr<transwarp::value_task<typename transwarp::decay<Value>::type>> {
    using task_t = transwarp::value_task<typename transwarp::decay<Value>::type>;
    return std::shared_ptr<task_t>(new task_t(std::forward<Value>(value)));
}


/// A function similar to std::for_each but returning a transwarp task for
/// deferred, possibly asynchronous execution. This function creates a graph
/// with std::distance(first, last) root tasks
template<typename InputIt, typename UnaryOperation>
auto for_each(InputIt first, InputIt last, UnaryOperation unary_op) -> std::shared_ptr<transwarp::task_impl<transwarp::wait_type, transwarp::no_op_functor, std::vector<std::shared_ptr<transwarp::task<void>>>>> {
    const auto distance = std::distance(first, last);
    if (distance <= 0) {
        throw transwarp::invalid_parameter{"first or last"};
    }
    std::vector<std::shared_ptr<transwarp::task<void>>> tasks;
    tasks.reserve(static_cast<std::size_t>(distance));
    for (; first != last; ++first) {
        tasks.push_back(transwarp::make_task(transwarp::root, [unary_op,first]{ unary_op(*first); }));
    }
    return transwarp::make_task(transwarp::wait, transwarp::no_op, tasks);
}

/// A function similar to std::for_each but returning a transwarp task for
/// deferred, possibly asynchronous execution. This function creates a graph
/// with std::distance(first, last) root tasks.
/// Overload for automatic scheduling by passing an executor.
template<typename InputIt, typename UnaryOperation>
auto for_each(transwarp::executor& executor, InputIt first, InputIt last, UnaryOperation unary_op) -> std::shared_ptr<transwarp::task_impl<transwarp::wait_type, transwarp::no_op_functor, std::vector<std::shared_ptr<transwarp::task<void>>>>> {
    auto task = transwarp::for_each(first, last, unary_op);
    task->schedule_all(executor);
    return task;
}


/// A function similar to std::transform but returning a transwarp task for
/// deferred, possibly asynchronous execution. This function creates a graph
/// with std::distance(first1, last1) root tasks
template<typename InputIt, typename OutputIt, typename UnaryOperation>
auto transform(InputIt first1, InputIt last1, OutputIt d_first, UnaryOperation unary_op) -> std::shared_ptr<transwarp::task_impl<transwarp::wait_type, transwarp::no_op_functor, std::vector<std::shared_ptr<transwarp::task<void>>>>> {
    const auto distance = std::distance(first1, last1);
    if (distance <= 0) {
        throw transwarp::invalid_parameter{"first1 or last1"};
    }
    std::vector<std::shared_ptr<transwarp::task<void>>> tasks;
    tasks.reserve(static_cast<std::size_t>(distance));
    for (; first1 != last1; ++first1, ++d_first) {
        tasks.push_back(transwarp::make_task(transwarp::root, [unary_op,first1,d_first]{ *d_first = unary_op(*first1); }));
    }
    return transwarp::make_task(transwarp::wait, transwarp::no_op, tasks);
}

/// A function similar to std::transform but returning a transwarp task for
/// deferred, possibly asynchronous execution. This function creates a graph
/// with std::distance(first1, last1) root tasks.
/// Overload for automatic scheduling by passing an executor.
template<typename InputIt, typename OutputIt, typename UnaryOperation>
auto transform(transwarp::executor& executor, InputIt first1, InputIt last1, OutputIt d_first, UnaryOperation unary_op) -> std::shared_ptr<transwarp::task_impl<transwarp::wait_type, transwarp::no_op_functor, std::vector<std::shared_ptr<transwarp::task<void>>>>> {
    auto task = transwarp::transform(first1, last1, d_first, unary_op);
    task->schedule_all(executor);
    return task;
}


/// A task pool that allows running multiple instances of the same task in parallel.
template<typename ResultType>
class task_pool {
public:

    /// Constructs a task pool
    task_pool(std::shared_ptr<transwarp::task<ResultType>> task,
              std::size_t minimum_size,
              std::size_t maximum_size)
    : task_(std::move(task)),
      minimum_(minimum_size),
      maximum_(maximum_size),
      finished_(maximum_size)
    {
        if (minimum_ < 1) {
            throw transwarp::invalid_parameter{"minimum size"};
        }
        if (minimum_ > maximum_) {
            throw transwarp::invalid_parameter{"minimum or maximum size"};
        }
        task_->add_listener(transwarp::event_type::after_finished, listener_);
        for (std::size_t i=0; i<minimum_; ++i) {
            idle_.push(task_->clone());
        }
    }

    /// Constructs a task pool with reasonable defaults for minimum and maximum
    explicit
    task_pool(std::shared_ptr<transwarp::task<ResultType>> task)
    : task_pool(std::move(task), 32, 65536)
    {}

    // delete copy/move semantics
    task_pool(const task_pool&) = delete;
    task_pool& operator=(const task_pool&) = delete;
    task_pool(task_pool&&) = delete;
    task_pool& operator=(task_pool&&) = delete;

    /// Returns the next idle task.
    /// If there are no idle tasks then it will attempt to double the
    /// pool size. If that fails then it will return a nullptr. On successful
    /// retrieval of an idle task the function will mark that task as busy.
    std::shared_ptr<transwarp::task<ResultType>> next_task(bool maybe_resize=true) {
        const transwarp::itask* finished_task{};
        {
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            if (!finished_.empty()) {
                finished_task = finished_.front(); finished_.pop();
            }
        }

        std::shared_ptr<transwarp::task<ResultType>> task;
        if (finished_task) {
            task = busy_.find(finished_task)->second;
        } else {
            if (maybe_resize && idle_.empty()) {
                resize(size() * 2); // double pool size
            }
            if (idle_.empty()) {
                return nullptr;
            }
            task = idle_.front(); idle_.pop();
            busy_.emplace(task.get(), task);
        }

        auto future = task->future();
        if (future.valid()) {
            future.wait(); // will return immediately
        }
        return task;
    }

    /// Just like next_task() but waits for a task to become available.
    /// The returned task will always be a valid pointer
    std::shared_ptr<transwarp::task<ResultType>> wait_for_next_task(bool maybe_resize=true) {
        for (;;) {
            std::shared_ptr<transwarp::task<ResultType>> g = next_task(maybe_resize);
            if (g) {
                return g;
            }
        }
    }

    /// Returns the current total size of the pool (sum of idle and busy tasks)
    std::size_t size() const {
        return idle_.size() + busy_.size();
    }

    /// Returns the minimum size of the pool
    std::size_t minimum_size() const {
        return minimum_;
    }

    /// Returns the maximum size of the pool
    std::size_t maximum_size() const {
        return maximum_;
    }

    /// Returns the number of idle tasks in the pool
    std::size_t idle_count() const {
        std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
        return idle_.size() + finished_.size();
    }

    /// Returns the number of busy tasks in the pool
    std::size_t busy_count() const {
        std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
        return busy_.size() - finished_.size();
    }

    /// Resizes the task pool to the given new size if possible
    void resize(std::size_t new_size) {
        reclaim();
        if (new_size > size()) { // grow
            const std::size_t count = new_size - size();
            for (std::size_t i=0; i<count; ++i) {
                if (size() == maximum_) {
                    break;
                }
                idle_.push(task_->clone());
            }
        } else if (new_size < size()) { // shrink
            const std::size_t count = size() - new_size;
            for (std::size_t i=0; i<count; ++i) {
                if (idle_.empty() || size() == minimum_) {
                    break;
                }
                idle_.pop();
            }
        }
    }

    /// Reclaims finished tasks by marking them as idle again
    void reclaim() {
        decltype(finished_) finished{finished_.capacity()};
        {
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            finished_.swap(finished);
        }
        while (!finished.empty()) {
            const transwarp::itask* task = finished.front(); finished.pop();
            const auto it = busy_.find(task);
            idle_.push(it->second);
            busy_.erase(it);
        }
    }

private:

    class finished_listener : public transwarp::listener {
    public:

        explicit
        finished_listener(task_pool<ResultType>& pool)
        : pool_(pool)
        {}

        // Called on a potentially high-priority thread
        void handle_event(transwarp::event_type, transwarp::itask& task) override {
            std::lock_guard<transwarp::detail::spinlock> lock{pool_.spinlock_};
            pool_.finished_.push(static_cast<const transwarp::itask*>(&task));
        }

    private:
        task_pool<ResultType>& pool_;
    };

    std::shared_ptr<transwarp::task<ResultType>> task_;
    std::size_t minimum_;
    std::size_t maximum_;
    mutable transwarp::detail::spinlock spinlock_; // protecting finished_
    transwarp::detail::circular_buffer<const transwarp::itask*> finished_;
    std::queue<std::shared_ptr<transwarp::task<ResultType>>> idle_;
    std::unordered_map<const transwarp::itask*, std::shared_ptr<transwarp::task<ResultType>>> busy_;
    std::shared_ptr<transwarp::listener> listener_{new finished_listener(*this)};
};


/// A timer that tracks the average idle, wait, and run time of each task it listens to.
/// - idle = time between scheduling and starting the task (executor dependent)
/// - wait = time between starting and invoking the task's functor, i.e. wait for parent tasks to finish
/// - run = time between invoking and finishing the task's computations
class timer : public transwarp::listener {
public:
    timer() = default;

    // delete copy/move semantics
    timer(const timer&) = delete;
    timer& operator=(const timer&) = delete;
    timer(timer&&) = delete;
    timer& operator=(timer&&) = delete;

    /// Performs the actual timing and populates the task's timing members
    void handle_event(const transwarp::event_type event, transwarp::itask& task) override {
        switch (event) {
        case transwarp::event_type::before_scheduled: {
            const std::chrono::time_point<std::chrono::steady_clock> now = std::chrono::steady_clock::now();
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            track.startidle = now;
        }
        break;
        case transwarp::event_type::before_started: {
            const std::chrono::time_point<std::chrono::steady_clock> now = std::chrono::steady_clock::now();
            track_idletime(task, now);
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            track.startwait = now;
        }
        break;
        case transwarp::event_type::after_canceled: {
            const std::chrono::time_point<std::chrono::steady_clock> now = std::chrono::steady_clock::now();
            track_waittime(task, now);
        }
        break;
        case transwarp::event_type::before_invoked: {
            const std::chrono::time_point<std::chrono::steady_clock> now = std::chrono::steady_clock::now();
            track_waittime(task, now);
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            track.running = true;
            track.startrun = now;
        }
        break;
        case transwarp::event_type::after_finished: {
            const std::chrono::time_point<std::chrono::steady_clock> now = std::chrono::steady_clock::now();
            track_runtime(task, now);
        }
        break;
        default: break;
        }
    }

    /// Resets all timing information
    void reset() {
        std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
        tracks_.clear();
    }

private:

    void track_idletime(transwarp::itask& task, const std::chrono::time_point<std::chrono::steady_clock>& now) {
        std::int64_t avg_idletime_us;
        {
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            track.idletime += std::chrono::duration_cast<std::chrono::microseconds>(now - track.startidle).count();
            ++track.idlecount;
            avg_idletime_us = static_cast<std::int64_t>(track.idletime / track.idlecount);
        }
        task.set_avg_idletime_us(avg_idletime_us);
    }

    void track_waittime(transwarp::itask& task, const std::chrono::time_point<std::chrono::steady_clock>& now) {
        std::int64_t avg_waittime_us;
        {
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            track.waittime += std::chrono::duration_cast<std::chrono::microseconds>(now - track.startwait).count();
            ++track.waitcount;
            avg_waittime_us = static_cast<std::int64_t>(track.waittime / track.waitcount);
        }
        task.set_avg_waittime_us(avg_waittime_us);
    }

    void track_runtime(transwarp::itask& task, const std::chrono::time_point<std::chrono::steady_clock>& now) {
        std::int64_t avg_runtime_us;
        {
            std::lock_guard<transwarp::detail::spinlock> lock{spinlock_};
            auto& track = tracks_[&task];
            if (!track.running) {
                return;
            }
            track.running = false;
            track.runtime += std::chrono::duration_cast<std::chrono::microseconds>(now - track.startrun).count();
            ++track.runcount;
            avg_runtime_us = static_cast<std::int64_t>(track.runtime / track.runcount);
        }
        task.set_avg_runtime_us(avg_runtime_us);
    }

    struct track {
        bool running = false;
        std::chrono::time_point<std::chrono::steady_clock> startidle;
        std::chrono::time_point<std::chrono::steady_clock> startwait;
        std::chrono::time_point<std::chrono::steady_clock> startrun;
        std::chrono::microseconds::rep idletime = 0;
        std::chrono::microseconds::rep idlecount = 0;
        std::chrono::microseconds::rep waittime = 0;
        std::chrono::microseconds::rep waitcount = 0;
        std::chrono::microseconds::rep runtime = 0;
        std::chrono::microseconds::rep runcount = 0;
    };

    transwarp::detail::spinlock spinlock_; // protecting tracks_
    std::unordered_map<const transwarp::itask*, track> tracks_;
};


/// The releaser will release a task's future when the task's `after_satisfied`
/// event was received which happens when all children received the task's result.
/// The releaser should be used in cases where the task's result is only needed
/// for consumption by its children and can then be discarded.
class releaser : public transwarp::listener {
public:
    releaser() = default;

    /// The executor gives control over where a task's future is released
    explicit releaser(std::shared_ptr<transwarp::executor> executor)
    : executor_(std::move(executor))
    {}

    // delete copy/move semantics
    releaser(const releaser&) = delete;
    releaser& operator=(const releaser&) = delete;
    releaser(releaser&&) = delete;
    releaser& operator=(releaser&&) = delete;

    void handle_event(const transwarp::event_type event, transwarp::itask& task) override {
        if (event == transwarp::event_type::after_satisfied) {
            if (executor_) {
                executor_->execute([&task]{ task.reset_future(); }, task);
            } else {
                task.reset_future();
            }
        }
    }

private:
    std::shared_ptr<transwarp::executor> executor_;
};


} // transwarp