xref: /oneTBB/include/oneapi/tbb/enumerable_thread_specific.h (revision 54a22a62)

/*
    Copyright (c) 2005-2024 Intel Corporation

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.
*/

#ifndef __TBB_enumerable_thread_specific_H
#define __TBB_enumerable_thread_specific_H

#include "detail/_config.h"
#include "detail/_namespace_injection.h"
#include "detail/_assert.h"
#include "detail/_template_helpers.h"
#include "detail/_aligned_space.h"

#include "concurrent_vector.h"
#include "tbb_allocator.h"
#include "cache_aligned_allocator.h"
#include "profiling.h"

#include <atomic>
#include <thread>
#include <cstring> // memcpy
#include <cstddef> // std::ptrdiff_t

#include "task.h" // for task::suspend_point

#if _WIN32 || _WIN64
#ifndef NOMINMAX
#define NOMINMAX
#define __TBB_DEFINED_NOMINMAX 1
#endif
#include <windows.h>
#if __TBB_DEFINED_NOMINMAX
#undef NOMINMAX
#undef __TBB_DEFINED_NOMINMAX
#endif
#else
#include <pthread.h>
#endif

namespace tbb {
namespace detail {
namespace d1 {

//! enum for selecting between single key and key-per-instance versions
enum ets_key_usage_type {
    ets_key_per_instance
    , ets_no_key
#if __TBB_RESUMABLE_TASKS
    , ets_suspend_aware
#endif
};

// Forward declaration to use in internal classes
template <typename T, typename Allocator, ets_key_usage_type ETS_key_type>
class enumerable_thread_specific;

template <std::size_t ThreadIDSize>
struct internal_ets_key_selector {
    using key_type = std::thread::id;
    static key_type current_key() {
        return std::this_thread::get_id();
    }
};

// Intel Compiler on OSX cannot create atomics objects that instantiated from non-fundamental types
#if __INTEL_COMPILER && __APPLE__
template<>
struct internal_ets_key_selector<sizeof(std::size_t)> {
    using key_type = std::size_t;
    static key_type current_key() {
        auto id = std::this_thread::get_id();
        return reinterpret_cast<key_type&>(id);
    }
};
#endif

template <ets_key_usage_type ETS_key_type>
struct ets_key_selector : internal_ets_key_selector<sizeof(std::thread::id)> {};

#if __TBB_RESUMABLE_TASKS
template <>
struct ets_key_selector<ets_suspend_aware> {
    using key_type = suspend_point;
    static key_type current_key() {
        return r1::current_suspend_point();
    }
};
#endif

template<ets_key_usage_type ETS_key_type>
class ets_base : detail::no_copy {
protected:
    using key_type = typename ets_key_selector<ETS_key_type>::key_type;

public:
    struct slot;
    struct array {
        array* next;
        std::size_t lg_size;
        slot& at( std::size_t k ) {
            return (reinterpret_cast<slot*>(reinterpret_cast<void*>(this+1)))[k];
        }
        std::size_t size() const { return std::size_t(1) << lg_size; }
        std::size_t mask() const { return size() - 1; }
        std::size_t start( std::size_t h ) const {
            return h >> (8 * sizeof(std::size_t) - lg_size);
        }
    };
    struct slot {
        std::atomic<key_type> key;
        void* ptr;
        bool empty() const { return key.load(std::memory_order_relaxed) == key_type(); }
        bool match( key_type k ) const { return key.load(std::memory_order_relaxed) == k; }
        bool claim( key_type k ) {
            // TODO: maybe claim ptr, because key_type is not guaranteed to fit into word size
            key_type expected = key_type();
            return key.compare_exchange_strong(expected, k);
        }
    };

protected:
    //! Root of linked list of arrays of decreasing size.
    /** nullptr if and only if my_count==0.
        Each array in the list is half the size of its predecessor. */
    std::atomic<array*> my_root;
    std::atomic<std::size_t> my_count;

    virtual void* create_local() = 0;
    virtual void* create_array(std::size_t _size) = 0;  // _size in bytes
    virtual void free_array(void* ptr, std::size_t _size) = 0; // _size in bytes

    array* allocate( std::size_t lg_size ) {
        std::size_t n = std::size_t(1) << lg_size;
        array* a = static_cast<array*>(create_array(sizeof(array) + n * sizeof(slot)));
        a->lg_size = lg_size;
        std::memset( a + 1, 0, n * sizeof(slot) );
        return a;
    }
    void deallocate(array* a) {
        std::size_t n = std::size_t(1) << (a->lg_size);
        free_array( static_cast<void*>(a), std::size_t(sizeof(array) + n * sizeof(slot)) );
    }

    ets_base() : my_root{nullptr}, my_count{0} {}
    virtual ~ets_base();  // g++ complains if this is not virtual

    void* table_lookup( bool& exists );
    void  table_clear();
    // The following functions are not used in concurrent context,
    // so we don't need synchronization and ITT annotations there.
    template <ets_key_usage_type E2>
    void table_elementwise_copy( const ets_base& other,
                                 void*(*add_element)(ets_base<E2>&, void*) ) {
        __TBB_ASSERT(!my_root.load(std::memory_order_relaxed), nullptr);
        __TBB_ASSERT(!my_count.load(std::memory_order_relaxed), nullptr);
        if( !other.my_root.load(std::memory_order_relaxed) ) return;
        array* root = allocate(other.my_root.load(std::memory_order_relaxed)->lg_size);
        my_root.store(root, std::memory_order_relaxed);
        root->next = nullptr;
        my_count.store(other.my_count.load(std::memory_order_relaxed), std::memory_order_relaxed);
        std::size_t mask = root->mask();
        for( array* r = other.my_root.load(std::memory_order_relaxed); r; r = r->next ) {
            for( std::size_t i = 0; i < r->size(); ++i ) {
                slot& s1 = r->at(i);
                if( !s1.empty() ) {
                    for( std::size_t j = root->start(std::hash<key_type>{}(s1.key.load(std::memory_order_relaxed))); ; j = (j+1)&mask ) {
                        slot& s2 = root->at(j);
                        if( s2.empty() ) {
                            s2.ptr = add_element(static_cast<ets_base<E2>&>(*this), s1.ptr);
                            s2.key.store(s1.key.load(std::memory_order_relaxed), std::memory_order_relaxed);
                            break;
                        }
                        else if( s2.match(s1.key.load(std::memory_order_relaxed)) )
                            break;
                    }
                }
            }
        }
    }
    void table_swap( ets_base& other ) {
       __TBB_ASSERT(this!=&other, "Don't swap an instance with itself");
       swap_atomics_relaxed(my_root, other.my_root);
       swap_atomics_relaxed(my_count, other.my_count);
    }
};

template<ets_key_usage_type ETS_key_type>
ets_base<ETS_key_type>::~ets_base() {
    __TBB_ASSERT(!my_root.load(std::memory_order_relaxed), nullptr);
}

template<ets_key_usage_type ETS_key_type>
void ets_base<ETS_key_type>::table_clear() {
    while ( array* r = my_root.load(std::memory_order_relaxed) ) {
        my_root.store(r->next, std::memory_order_relaxed);
        deallocate(r);
    }
    my_count.store(0, std::memory_order_relaxed);
}

template<ets_key_usage_type ETS_key_type>
void* ets_base<ETS_key_type>::table_lookup( bool& exists ) {
    const key_type k = ets_key_selector<ETS_key_type>::current_key();

    __TBB_ASSERT(k != key_type(), nullptr);
    void* found;
    std::size_t h = std::hash<key_type>{}(k);
    for( array* r = my_root.load(std::memory_order_acquire); r; r = r->next ) {
        call_itt_notify(acquired,r);
        std::size_t mask=r->mask();
        for(std::size_t i = r->start(h); ;i=(i+1)&mask) {
            slot& s = r->at(i);
            if( s.empty() ) break;
            if( s.match(k) ) {
                if( r == my_root.load(std::memory_order_acquire) ) {
                    // Success at top level
                    exists = true;
                    return s.ptr;
                } else {
                    // Success at some other level.  Need to insert at top level.
                    exists = true;
                    found = s.ptr;
                    goto insert;
                }
            }
        }
    }
    // Key does not yet exist.  The density of slots in the table does not exceed 0.5,
    // for if this will occur a new table is allocated with double the current table
    // size, which is swapped in as the new root table.  So an empty slot is guaranteed.
    exists = false;
    found = create_local();
    {
        std::size_t c = ++my_count;
        array* r = my_root.load(std::memory_order_acquire);
        call_itt_notify(acquired,r);
        if( !r || c > r->size()/2 ) {
            std::size_t s = r ? r->lg_size : 2;
            while( c > std::size_t(1)<<(s-1) ) ++s;
            array* a = allocate(s);
            for(;;) {
                a->next = r;
                call_itt_notify(releasing,a);
                array* new_r = r;
                if( my_root.compare_exchange_strong(new_r, a) ) break;
                call_itt_notify(acquired, new_r);
                __TBB_ASSERT(new_r != nullptr, nullptr);
                if( new_r->lg_size >= s ) {
                    // Another thread inserted an equal or  bigger array, so our array is superfluous.
                    deallocate(a);
                    break;
                }
                r = new_r;
            }
        }
    }
    insert:
    // Whether a slot has been found in an older table, or if it has been inserted at this level,
    // it has already been accounted for in the total.  Guaranteed to be room for it, and it is
    // not present, so search for empty slot and use it.
    array* ir = my_root.load(std::memory_order_acquire);
    call_itt_notify(acquired, ir);
    std::size_t mask = ir->mask();
    for(std::size_t i = ir->start(h);; i = (i+1)&mask) {
        slot& s = ir->at(i);
        if( s.empty() ) {
            if( s.claim(k) ) {
                s.ptr = found;
                return found;
            }
        }
    }
}

//! Specialization that exploits native TLS
template <>
class ets_base<ets_key_per_instance>: public ets_base<ets_no_key> {
    using super = ets_base<ets_no_key>;
#if _WIN32||_WIN64
#if __TBB_WIN8UI_SUPPORT
    using tls_key_t = DWORD;
    void create_key() { my_key = FlsAlloc(nullptr); }
    void destroy_key() { FlsFree(my_key); }
    void set_tls(void * value) { FlsSetValue(my_key, (LPVOID)value); }
    void* get_tls() { return (void *)FlsGetValue(my_key); }
#else
    using tls_key_t = DWORD;
    void create_key() { my_key = TlsAlloc(); }
    void destroy_key() { TlsFree(my_key); }
    void set_tls(void * value) { TlsSetValue(my_key, (LPVOID)value); }
    void* get_tls() { return (void *)TlsGetValue(my_key); }
#endif
#else
    using tls_key_t = pthread_key_t;
    void create_key() { pthread_key_create(&my_key, nullptr); }
    void destroy_key() { pthread_key_delete(my_key); }
    void set_tls( void * value ) const { pthread_setspecific(my_key, value); }
    void* get_tls() const { return pthread_getspecific(my_key); }
#endif
    tls_key_t my_key;
    virtual void* create_local() override = 0;
    virtual void* create_array(std::size_t _size) override = 0;  // _size in bytes
    virtual void free_array(void* ptr, std::size_t _size) override = 0; // size in bytes
protected:
    ets_base() {create_key();}
    ~ets_base() {destroy_key();}
    void* table_lookup( bool& exists ) {
        void* found = get_tls();
        if( found ) {
            exists=true;
        } else {
            found = super::table_lookup(exists);
            set_tls(found);
        }
        return found;
    }
    void table_clear() {
        destroy_key();
        create_key();
        super::table_clear();
    }
    void table_swap( ets_base& other ) {
       using std::swap;
       __TBB_ASSERT(this!=&other, "Don't swap an instance with itself");
       swap(my_key, other.my_key);
       super::table_swap(other);
    }
};

//! Random access iterator for traversing the thread local copies.
template< typename Container, typename Value >
class enumerable_thread_specific_iterator
{
    //! current position in the concurrent_vector

    Container *my_container;
    typename Container::size_type my_index;
    mutable Value *my_value;

    template<typename C, typename T, typename U>
    friend bool operator==( const enumerable_thread_specific_iterator<C, T>& i,
                     const enumerable_thread_specific_iterator<C, U>& j );

    template<typename C, typename T, typename U>
    friend bool operator<( const enumerable_thread_specific_iterator<C,T>& i,
                           const enumerable_thread_specific_iterator<C,U>& j );

    template<typename C, typename T, typename U>
    friend std::ptrdiff_t operator-( const enumerable_thread_specific_iterator<C,T>& i,
                                const enumerable_thread_specific_iterator<C,U>& j );

    template<typename C, typename U>
    friend class enumerable_thread_specific_iterator;

public:
    //! STL support
    using difference_type = std::ptrdiff_t;
    using value_type = Value;
    using pointer = Value*;
    using reference = Value&;
    using iterator_category = std::random_access_iterator_tag;

    enumerable_thread_specific_iterator( const Container &container, typename Container::size_type index ) :
        my_container(&const_cast<Container &>(container)), my_index(index), my_value(nullptr) {}

    //! Default constructor
    enumerable_thread_specific_iterator() : my_container(nullptr), my_index(0), my_value(nullptr) {}

    template<typename U>
    enumerable_thread_specific_iterator( const enumerable_thread_specific_iterator<Container, U>& other ) :
            my_container( other.my_container ), my_index( other.my_index), my_value( const_cast<Value *>(other.my_value) ) {}

    enumerable_thread_specific_iterator operator+( std::ptrdiff_t offset ) const {
        return enumerable_thread_specific_iterator(*my_container, my_index + offset);
    }

    friend enumerable_thread_specific_iterator operator+( std::ptrdiff_t offset, enumerable_thread_specific_iterator v ) {
        return enumerable_thread_specific_iterator(*v.my_container, v.my_index + offset);
    }

    enumerable_thread_specific_iterator &operator+=( std::ptrdiff_t offset ) {
        my_index += offset;
        my_value = nullptr;
        return *this;
    }

    enumerable_thread_specific_iterator operator-( std::ptrdiff_t offset ) const {
        return enumerable_thread_specific_iterator( *my_container, my_index-offset );
    }

    enumerable_thread_specific_iterator &operator-=( std::ptrdiff_t offset ) {
        my_index -= offset;
        my_value = nullptr;
        return *this;
    }

    Value& operator*() const {
        Value* value = my_value;
        if( !value ) {
            value = my_value = (*my_container)[my_index].value();
        }
        __TBB_ASSERT( value==(*my_container)[my_index].value(), "corrupt cache" );
        return *value;
    }

    Value& operator[]( std::ptrdiff_t k ) const {
       return *(*my_container)[my_index + k].value();
    }

    Value* operator->() const {return &operator*();}

    enumerable_thread_specific_iterator& operator++() {
        ++my_index;
        my_value = nullptr;
        return *this;
    }

    enumerable_thread_specific_iterator& operator--() {
        --my_index;
        my_value = nullptr;
        return *this;
    }

    //! Post increment
    enumerable_thread_specific_iterator operator++(int) {
        enumerable_thread_specific_iterator result = *this;
        ++my_index;
        my_value = nullptr;
        return result;
    }

    //! Post decrement
    enumerable_thread_specific_iterator operator--(int) {
        enumerable_thread_specific_iterator result = *this;
        --my_index;
        my_value = nullptr;
        return result;
    }
};

template<typename Container, typename T, typename U>
bool operator==( const enumerable_thread_specific_iterator<Container, T>& i,
                 const enumerable_thread_specific_iterator<Container, U>& j ) {
    return i.my_index == j.my_index && i.my_container == j.my_container;
}

template<typename Container, typename T, typename U>
bool operator!=( const enumerable_thread_specific_iterator<Container,T>& i,
                 const enumerable_thread_specific_iterator<Container,U>& j ) {
    return !(i==j);
}

template<typename Container, typename T, typename U>
bool operator<( const enumerable_thread_specific_iterator<Container,T>& i,
                const enumerable_thread_specific_iterator<Container,U>& j ) {
    return i.my_index<j.my_index;
}

template<typename Container, typename T, typename U>
bool operator>( const enumerable_thread_specific_iterator<Container,T>& i,
                const enumerable_thread_specific_iterator<Container,U>& j ) {
    return j<i;
}

template<typename Container, typename T, typename U>
bool operator>=( const enumerable_thread_specific_iterator<Container,T>& i,
                 const enumerable_thread_specific_iterator<Container,U>& j ) {
    return !(i<j);
}

template<typename Container, typename T, typename U>
bool operator<=( const enumerable_thread_specific_iterator<Container,T>& i,
                 const enumerable_thread_specific_iterator<Container,U>& j ) {
    return !(j<i);
}

template<typename Container, typename T, typename U>
std::ptrdiff_t operator-( const enumerable_thread_specific_iterator<Container,T>& i,
                     const enumerable_thread_specific_iterator<Container,U>& j ) {
    return i.my_index-j.my_index;
}

template<typename SegmentedContainer, typename Value >
class segmented_iterator
{
    template<typename C, typename T, typename U>
    friend bool operator==(const segmented_iterator<C,T>& i, const segmented_iterator<C,U>& j);

    template<typename C, typename T, typename U>
    friend bool operator!=(const segmented_iterator<C,T>& i, const segmented_iterator<C,U>& j);

    template<typename C, typename U>
    friend class segmented_iterator;

public:
    segmented_iterator() {my_segcont = nullptr;}

    segmented_iterator( const SegmentedContainer& _segmented_container ) :
        my_segcont(const_cast<SegmentedContainer*>(&_segmented_container)),
        outer_iter(my_segcont->end()) { }

    ~segmented_iterator() {}

    using InnerContainer = typename SegmentedContainer::value_type;
    using inner_iterator = typename InnerContainer::iterator;
    using outer_iterator = typename SegmentedContainer::iterator;

    // STL support
    // TODO: inherit all types from segmented container?
    using difference_type = std::ptrdiff_t;
    using value_type = Value;
    using size_type = typename SegmentedContainer::size_type;
    using pointer = Value*;
    using reference = Value&;
    using iterator_category = std::input_iterator_tag;

    // Copy Constructor
    template<typename U>
    segmented_iterator(const segmented_iterator<SegmentedContainer, U>& other) :
        my_segcont(other.my_segcont),
        outer_iter(other.outer_iter),
        // can we assign a default-constructed iterator to inner if we're at the end?
        inner_iter(other.inner_iter)
    {}

    // assignment
    template<typename U>
    segmented_iterator& operator=( const segmented_iterator<SegmentedContainer, U>& other) {
        my_segcont = other.my_segcont;
        outer_iter = other.outer_iter;
        if(outer_iter != my_segcont->end()) inner_iter = other.inner_iter;
        return *this;
    }

    // allow assignment of outer iterator to segmented iterator.  Once it is
    // assigned, move forward until a non-empty inner container is found or
    // the end of the outer container is reached.
    segmented_iterator& operator=(const outer_iterator& new_outer_iter) {
        __TBB_ASSERT(my_segcont != nullptr, nullptr);
        // check that this iterator points to something inside the segmented container
        for(outer_iter = new_outer_iter ;outer_iter!=my_segcont->end(); ++outer_iter) {
            if( !outer_iter->empty() ) {
                inner_iter = outer_iter->begin();
                break;
            }
        }
        return *this;
    }

    // pre-increment
    segmented_iterator& operator++() {
        advance_me();
        return *this;
    }

    // post-increment
    segmented_iterator operator++(int) {
        segmented_iterator tmp = *this;
        operator++();
        return tmp;
    }

    bool operator==(const outer_iterator& other_outer) const {
        __TBB_ASSERT(my_segcont != nullptr, nullptr);
        return (outer_iter == other_outer &&
                (outer_iter == my_segcont->end() || inner_iter == outer_iter->begin()));
    }

    bool operator!=(const outer_iterator& other_outer) const {
        return !operator==(other_outer);

    }

    // (i)* RHS
    reference operator*() const {
        __TBB_ASSERT(my_segcont != nullptr, nullptr);
        __TBB_ASSERT(outer_iter != my_segcont->end(), "Dereferencing a pointer at end of container");
        __TBB_ASSERT(inner_iter != outer_iter->end(), nullptr); // should never happen
        return *inner_iter;
    }

    // i->
    pointer operator->() const { return &operator*();}

private:
    SegmentedContainer* my_segcont;
    outer_iterator outer_iter;
    inner_iterator inner_iter;

    void advance_me() {
        __TBB_ASSERT(my_segcont != nullptr, nullptr);
        __TBB_ASSERT(outer_iter != my_segcont->end(), nullptr); // not true if there are no inner containers
        __TBB_ASSERT(inner_iter != outer_iter->end(), nullptr); // not true if the inner containers are all empty.
        ++inner_iter;
        while(inner_iter == outer_iter->end() && ++outer_iter != my_segcont->end()) {
            inner_iter = outer_iter->begin();
        }
    }
};    // segmented_iterator

template<typename SegmentedContainer, typename T, typename U>
bool operator==( const segmented_iterator<SegmentedContainer,T>& i,
                 const segmented_iterator<SegmentedContainer,U>& j ) {
    if(i.my_segcont != j.my_segcont) return false;
    if(i.my_segcont == nullptr) return true;
    if(i.outer_iter != j.outer_iter) return false;
    if(i.outer_iter == i.my_segcont->end()) return true;
    return i.inner_iter == j.inner_iter;
}

// !=
template<typename SegmentedContainer, typename T, typename U>
bool operator!=( const segmented_iterator<SegmentedContainer,T>& i,
                 const segmented_iterator<SegmentedContainer,U>& j ) {
    return !(i==j);
}

template<typename T>
struct construct_by_default: no_assign {
    void construct(void*where) {new(where) T();} // C++ note: the () in T() ensure zero initialization.
    construct_by_default( int ) {}
};

template<typename T>
struct construct_by_exemplar: no_assign {
    const T exemplar;
    void construct(void*where) {new(where) T(exemplar);}
    construct_by_exemplar( const T& t ) : exemplar(t) {}
    construct_by_exemplar( T&& t ) : exemplar(std::move(t)) {}
};

template<typename T, typename Finit>
struct construct_by_finit: no_assign {
    Finit f;
    void construct(void* where) {new(where) T(f());}
    construct_by_finit( Finit&& f_ ) : f(std::move(f_)) {}
};

template<typename T, typename... P>
struct construct_by_args: no_assign {
    stored_pack<P...> pack;
    void construct(void* where) {
        call( [where](const typename std::decay<P>::type&... args ){
           new(where) T(args...);
        }, pack );
    }
    construct_by_args( P&& ... args ) : pack(std::forward<P>(args)...) {}
};

// storage for initialization function pointer
// TODO: consider removing the template parameter T here and in callback_leaf
class callback_base {
public:
    // Clone *this
    virtual callback_base* clone() const = 0;
    // Destruct and free *this
    virtual void destroy() = 0;
    // Need virtual destructor to satisfy GCC compiler warning
    virtual ~callback_base() { }
    // Construct T at where
    virtual void construct(void* where) = 0;
};

template <typename Constructor>
class callback_leaf: public callback_base, Constructor {
    template<typename... P> callback_leaf( P&& ... params ) : Constructor(std::forward<P>(params)...) {}
    // TODO: make the construction/destruction consistent (use allocator.construct/destroy)
    using my_allocator_type = typename tbb::tbb_allocator<callback_leaf>;

    callback_base* clone() const override {
        return make(*this);
    }

    void destroy() override {
        my_allocator_type alloc;
        tbb::detail::allocator_traits<my_allocator_type>::destroy(alloc, this);
        tbb::detail::allocator_traits<my_allocator_type>::deallocate(alloc, this, 1);
    }

    void construct(void* where) override {
        Constructor::construct(where);
    }

public:
    template<typename... P>
    static callback_base* make( P&& ... params ) {
        void* where = my_allocator_type().allocate(1);
        return new(where) callback_leaf( std::forward<P>(params)... );
    }
};

//! Template for recording construction of objects in table
/** All maintenance of the space will be done explicitly on push_back,
    and all thread local copies must be destroyed before the concurrent
    vector is deleted.

    The flag is_built is initialized to false.  When the local is
    successfully-constructed, set the flag to true or call value_committed().
    If the constructor throws, the flag will be false.
*/
template<typename U>
struct ets_element {
    detail::aligned_space<U> my_space;
    bool is_built;
    ets_element() { is_built = false; }  // not currently-built
    U* value() { return my_space.begin(); }
    U* value_committed() { is_built = true; return my_space.begin(); }
    ~ets_element() {
        if(is_built) {
            my_space.begin()->~U();
            is_built = false;
        }
    }
};

// A predicate that can be used for a compile-time compatibility check of ETS instances
// Ideally, it should have been declared inside the ETS class, but unfortunately
// in that case VS2013 does not enable the variadic constructor.
template<typename T, typename ETS> struct is_compatible_ets : std::false_type {};
template<typename T, typename U, typename A, ets_key_usage_type C>
struct is_compatible_ets< T, enumerable_thread_specific<U,A,C> > : std::is_same<T, U> {};

// A predicate that checks whether, for a variable 'foo' of type T, foo() is a valid expression
template <typename T> using has_empty_braces_operator = decltype(std::declval<T>()());
template <typename T> using is_callable_no_args = supports<T, has_empty_braces_operator>;

//! The enumerable_thread_specific container
/** enumerable_thread_specific has the following properties:
    - thread-local copies are lazily created, with default, exemplar or function initialization.
    - thread-local copies do not move (during lifetime, and excepting clear()) so the address of a copy is invariant.
    - the contained objects need not have operator=() defined if combine is not used.
    - enumerable_thread_specific containers may be copy-constructed or assigned.
    - thread-local copies can be managed by hash-table, or can be accessed via TLS storage for speed.
    - outside of parallel contexts, the contents of all thread-local copies are accessible by iterator or using combine or combine_each methods

@par Segmented iterator
    When the thread-local objects are containers with input_iterators defined, a segmented iterator may
    be used to iterate over all the elements of all thread-local copies.

@par combine and combine_each
    - Both methods are defined for enumerable_thread_specific.
    - combine() requires the type T have operator=() defined.
    - neither method modifies the contents of the object (though there is no guarantee that the applied methods do not modify the object.)
    - Both are evaluated in serial context (the methods are assumed to be non-benign.)

@ingroup containers */
template <typename T, typename Allocator=cache_aligned_allocator<T>,
          ets_key_usage_type ETS_key_type=ets_no_key >
class enumerable_thread_specific: ets_base<ETS_key_type> {

    template<typename U, typename A, ets_key_usage_type C> friend class enumerable_thread_specific;

    using padded_element = padded<ets_element<T>>;

    //! A generic range, used to create range objects from the iterators
    template<typename I>
    class generic_range_type: public blocked_range<I> {
    public:
        using value_type = T;
        using reference = T&;
        using const_reference = const T&;
        using iterator = I;
        using difference_type = std::ptrdiff_t;

        generic_range_type( I begin_, I end_, std::size_t grainsize_ = 1) : blocked_range<I>(begin_,end_,grainsize_) {}
        template<typename U>
        generic_range_type( const generic_range_type<U>& r) : blocked_range<I>(r.begin(),r.end(),r.grainsize()) {}
        generic_range_type( generic_range_type& r, split ) : blocked_range<I>(r,split()) {}
    };

    using allocator_traits_type = tbb::detail::allocator_traits<Allocator>;

    using padded_allocator_type = typename allocator_traits_type::template rebind_alloc<padded_element>;
    using internal_collection_type = tbb::concurrent_vector< padded_element, padded_allocator_type >;

    callback_base *my_construct_callback;

    internal_collection_type my_locals;

    // TODO: consider unifying the callback mechanism for all create_local* methods below
    //   (likely non-compatible and requires interface version increase)
    void* create_local() override {
        padded_element& lref = *my_locals.grow_by(1);
        my_construct_callback->construct(lref.value());
        return lref.value_committed();
    }

    static void* create_local_by_copy( ets_base<ETS_key_type>& base, void* p ) {
        enumerable_thread_specific& ets = static_cast<enumerable_thread_specific&>(base);
        padded_element& lref = *ets.my_locals.grow_by(1);
        new(lref.value()) T(*static_cast<T*>(p));
        return lref.value_committed();
    }

    static void* create_local_by_move( ets_base<ETS_key_type>& base, void* p ) {
        enumerable_thread_specific& ets = static_cast<enumerable_thread_specific&>(base);
        padded_element& lref = *ets.my_locals.grow_by(1);
        new(lref.value()) T(std::move(*static_cast<T*>(p)));
        return lref.value_committed();
    }

    using array_allocator_type = typename allocator_traits_type::template rebind_alloc<uintptr_t>;

    // _size is in bytes
    void* create_array(std::size_t _size) override {
        std::size_t nelements = (_size + sizeof(uintptr_t) -1) / sizeof(uintptr_t);
        return array_allocator_type().allocate(nelements);
    }

    void free_array( void* _ptr, std::size_t _size) override {
        std::size_t nelements = (_size + sizeof(uintptr_t) -1) / sizeof(uintptr_t);
        array_allocator_type().deallocate( reinterpret_cast<uintptr_t *>(_ptr),nelements);
    }

public:

    //! Basic types
    using value_type = T;
    using allocator_type = Allocator;
    using size_type = typename internal_collection_type::size_type;
    using difference_type = typename internal_collection_type::difference_type;
    using reference = value_type&;
    using const_reference = const value_type&;

    using pointer = typename allocator_traits_type::pointer;
    using const_pointer = typename allocator_traits_type::const_pointer;

    // Iterator types
    using iterator = enumerable_thread_specific_iterator<internal_collection_type, value_type>;
    using const_iterator = enumerable_thread_specific_iterator<internal_collection_type, const value_type>;

    // Parallel range types
    using range_type = generic_range_type<iterator>;
    using const_range_type = generic_range_type<const_iterator>;

    //! Default constructor.  Each local instance of T is default constructed.
    enumerable_thread_specific() : my_construct_callback(
        callback_leaf<construct_by_default<T> >::make(/*dummy argument*/0)
    ){}

    //! Constructor with initializer functor. Each local instance of T is constructed by T(finit()).
    template <typename Finit , typename = typename std::enable_if<is_callable_no_args<typename std::decay<Finit>::type>::value>::type>
    explicit enumerable_thread_specific( Finit finit ) : my_construct_callback(
        callback_leaf<construct_by_finit<T,Finit> >::make( std::move(finit) )
    ){}

    //! Constructor with exemplar. Each local instance of T is copy-constructed from the exemplar.
    explicit enumerable_thread_specific( const T& exemplar ) : my_construct_callback(
        callback_leaf<construct_by_exemplar<T> >::make( exemplar )
    ){}

    explicit enumerable_thread_specific( T&& exemplar ) : my_construct_callback(
        callback_leaf<construct_by_exemplar<T> >::make( std::move(exemplar) )
    ){}

    //! Variadic constructor with initializer arguments.  Each local instance of T is constructed by T(args...)
    template <typename P1, typename... P,
              typename = typename std::enable_if<!is_callable_no_args<typename std::decay<P1>::type>::value
                                                      && !is_compatible_ets<T, typename std::decay<P1>::type>::value
                                                      && !std::is_same<T, typename std::decay<P1>::type>::value
                                                     >::type>
    enumerable_thread_specific( P1&& arg1, P&& ... args ) : my_construct_callback(
        callback_leaf<construct_by_args<T,P1,P...> >::make( std::forward<P1>(arg1), std::forward<P>(args)... )
    ){}

    //! Destructor
    ~enumerable_thread_specific() {
        if(my_construct_callback) my_construct_callback->destroy();
        // Deallocate the hash table before overridden free_array() becomes inaccessible
        this->ets_base<ETS_key_type>::table_clear();
    }

    //! returns reference to local, discarding exists
    reference local() {
        bool exists;
        return local(exists);
    }

    //! Returns reference to calling thread's local copy, creating one if necessary
    reference local(bool& exists)  {
        void* ptr = this->table_lookup(exists);
        return *(T*)ptr;
    }

    //! Get the number of local copies
    size_type size() const { return my_locals.size(); }

    //! true if there have been no local copies created
    bool empty() const { return my_locals.empty(); }

    //! begin iterator
    iterator begin() { return iterator( my_locals, 0 ); }
    //! end iterator
    iterator end() { return iterator(my_locals, my_locals.size() ); }

    //! begin const iterator
    const_iterator begin() const { return const_iterator(my_locals, 0); }

    //! end const iterator
    const_iterator end() const { return const_iterator(my_locals, my_locals.size()); }

    //! Get range for parallel algorithms
    range_type range( std::size_t grainsize=1 ) { return range_type( begin(), end(), grainsize ); }

    //! Get const range for parallel algorithms
    const_range_type range( std::size_t grainsize=1 ) const { return const_range_type( begin(), end(), grainsize ); }

    //! Destroys local copies
    void clear() {
        my_locals.clear();
        this->table_clear();
        // callback is not destroyed
    }

private:
    template<typename A2, ets_key_usage_type C2>
    void internal_copy(const enumerable_thread_specific<T, A2, C2>& other) {
        // this tests is_compatible_ets
        static_assert( (is_compatible_ets<T, typename std::decay<decltype(other)>::type>::value), "is_compatible_ets fails" );
        // Initialize my_construct_callback first, so that it is valid even if rest of this routine throws an exception.
        my_construct_callback = other.my_construct_callback->clone();
        __TBB_ASSERT(my_locals.size()==0, nullptr);
        my_locals.reserve(other.size());
        this->table_elementwise_copy( other, create_local_by_copy );
    }

    void internal_swap(enumerable_thread_specific& other) {
        using std::swap;
        __TBB_ASSERT( this!=&other, nullptr);
        swap(my_construct_callback, other.my_construct_callback);
        // concurrent_vector::swap() preserves storage space,
        // so addresses to the vector kept in ETS hash table remain valid.
        swap(my_locals, other.my_locals);
        this->ets_base<ETS_key_type>::table_swap(other);
    }

    template<typename A2, ets_key_usage_type C2>
    void internal_move(enumerable_thread_specific<T, A2, C2>&& other) {
        static_assert( (is_compatible_ets<T, typename std::decay<decltype(other)>::type>::value), "is_compatible_ets fails" );
        my_construct_callback = other.my_construct_callback;
        other.my_construct_callback = nullptr;
        __TBB_ASSERT(my_locals.size()==0, nullptr);
        my_locals.reserve(other.size());
        this->table_elementwise_copy( other, create_local_by_move );
    }

public:
    enumerable_thread_specific( const enumerable_thread_specific& other )
    : ets_base<ETS_key_type>() /* prevents GCC warnings with -Wextra */
    {
        internal_copy(other);
    }

    template<typename Alloc, ets_key_usage_type Cachetype>
    enumerable_thread_specific( const enumerable_thread_specific<T, Alloc, Cachetype>& other )
    {
        internal_copy(other);
    }

    enumerable_thread_specific( enumerable_thread_specific&& other ) : my_construct_callback()
    {
        // TODO: use internal_move correctly here
        internal_swap(other);
    }

    template<typename Alloc, ets_key_usage_type Cachetype>
    enumerable_thread_specific( enumerable_thread_specific<T, Alloc, Cachetype>&& other ) : my_construct_callback()
    {
        internal_move(std::move(other));
    }

    enumerable_thread_specific& operator=( const enumerable_thread_specific& other )
    {
        if( this != &other ) {
            this->clear();
            my_construct_callback->destroy();
            internal_copy( other );
        }
        return *this;
    }

    template<typename Alloc, ets_key_usage_type Cachetype>
    enumerable_thread_specific& operator=( const enumerable_thread_specific<T, Alloc, Cachetype>& other )
    {
        __TBB_ASSERT( static_cast<void*>(this)!=static_cast<const void*>(&other), nullptr); // Objects of different types
        this->clear();
        my_construct_callback->destroy();
        internal_copy(other);
        return *this;
    }

    enumerable_thread_specific& operator=( enumerable_thread_specific&& other )
    {
        if( this != &other ) {
            // TODO: use internal_move correctly here
            internal_swap(other);
        }
        return *this;
    }

    template<typename Alloc, ets_key_usage_type Cachetype>
    enumerable_thread_specific& operator=( enumerable_thread_specific<T, Alloc, Cachetype>&& other )
    {
        __TBB_ASSERT( static_cast<void*>(this)!=static_cast<const void*>(&other), nullptr); // Objects of different types
        this->clear();
        my_construct_callback->destroy();
        internal_move(std::move(other));
        return *this;
    }

    // CombineFunc has signature T(T,T) or T(const T&, const T&)
    template <typename CombineFunc>
    T combine(CombineFunc f_combine) {
        if(begin() == end()) {
            ets_element<T> location;
            my_construct_callback->construct(location.value());
            return *location.value_committed();
        }
        const_iterator ci = begin();
        T my_result = *ci;
        while(++ci != end())
            my_result = f_combine( my_result, *ci );
        return my_result;
    }

    // combine_func_t takes T by value or by [const] reference, and returns nothing
    template <typename CombineFunc>
    void combine_each(CombineFunc f_combine) {
        for(iterator ci = begin(); ci != end(); ++ci) {
            f_combine( *ci );
        }
    }

}; // enumerable_thread_specific

template< typename Container >
class flattened2d {
    // This intermediate typedef is to address issues with VC7.1 compilers
    using conval_type = typename Container::value_type;

public:
    //! Basic types
    using size_type = typename conval_type::size_type;
    using difference_type = typename conval_type::difference_type;
    using allocator_type = typename conval_type::allocator_type;
    using value_type = typename conval_type::value_type;
    using reference = typename conval_type::reference;
    using const_reference = typename conval_type::const_reference;
    using pointer = typename conval_type::pointer;
    using const_pointer = typename conval_type::const_pointer;

    using iterator = segmented_iterator<Container, value_type>;
    using const_iterator = segmented_iterator<Container, const value_type>;

    flattened2d( const Container &c, typename Container::const_iterator b, typename Container::const_iterator e ) :
        my_container(const_cast<Container*>(&c)), my_begin(b), my_end(e) { }

    explicit flattened2d( const Container &c ) :
        my_container(const_cast<Container*>(&c)), my_begin(c.begin()), my_end(c.end()) { }

    iterator begin() { return iterator(*my_container) = my_begin; }
    iterator end() { return iterator(*my_container) = my_end; }
    const_iterator begin() const { return const_iterator(*my_container) = my_begin; }
    const_iterator end() const { return const_iterator(*my_container) = my_end; }

    size_type size() const {
        size_type tot_size = 0;
        for(typename Container::const_iterator i = my_begin; i != my_end; ++i) {
            tot_size += i->size();
        }
        return tot_size;
    }

private:
    Container *my_container;
    typename Container::const_iterator my_begin;
    typename Container::const_iterator my_end;
};

template <typename Container>
flattened2d<Container> flatten2d(const Container &c, const typename Container::const_iterator b, const typename Container::const_iterator e) {
    return flattened2d<Container>(c, b, e);
}

template <typename Container>
flattened2d<Container> flatten2d(const Container &c) {
    return flattened2d<Container>(c);
}

} // namespace d1
} // namespace detail

inline namespace v1 {
using detail::d1::enumerable_thread_specific;
using detail::d1::flattened2d;
using detail::d1::flatten2d;
// ets enum keys
using detail::d1::ets_key_usage_type;
using detail::d1::ets_key_per_instance;
using detail::d1::ets_no_key;
#if __TBB_RESUMABLE_TASKS
using detail::d1::ets_suspend_aware;
#endif
} // inline namespace v1

} // namespace tbb

#endif // __TBB_enumerable_thread_specific_H