-1
\$\begingroup\$

I wrote a Singleton template, with examples, google tests and README https://github.com/erez-strauss/init_singleton/blob/master/singleton.h

The usage can be as simple as:

#include <singleton.h>
int main()
{
    es::init::singleton<int>::instance() =
        4;  // a singleton of a type int, initialized before main() starts, destroyed after main() exits.
    return 0;
}

That is a singleton of a type int, it is being created before the main() starts, and destroyed after main() exists.

A bit more useful example would be:

#include <app_singletons.h>
class DataA { public: DataA(){std::cout << "DataA()\n";
   es::init::args.for_each(
        [](int index, auto& arg) { std::cout << "DataA: arg[" << index << "]: '" << arg << "'" << std::endl; });
   es::init::env.for_each(
        [](int index, auto& earg) { std::cout << "DataA: env[" << index << "]: '" << earg << "'" << std::endl; });
} ~DataA() {std::cout << "~DataA()\n";}};
int main()
{
    std::cout << "main() start\n";
    es::init::singleton<DataA>::instance(); // a singleton of a type DataA, initialized before main() starts, destroyed after main() exits.
    std::cout << "main() end\n";
    return 0;
}

$ ./singleton8 a b c
DataA()
DataA: arg[0]: './singleton8'
DataA: arg[1]: 'a'
DataA: arg[2]: 'b'
DataA: arg[3]: 'c'
DataA: env[0]: 'SHELL=/bin/bash'
.....
DataA: env[63]: '_=./singleton8'
main() start
main() end
~DataA()

The singleton template is using an atomic<> pointer to a function, which returns the address of the singleton object. This atomic pointer first points to an initialization function, which creates the singleton and changes the pointer to point to another function which just return the reference to the singleton with no condition.

The above example shows two features of the singleton early_initializer

  1. access to std::cout/std::cerr is available from the early initialized constructors - for this the init function need to create instance of the std::ios_base::Init object, which initializes them, on first such object creation.

  2. the access to command line arguments and environment from early initialized constructors, by two singletons that are defined in app_singleton.h, the args singleton init function is a constructor function that expects the argc, argv parameters.

Singleton objects can be of any type, and if you choose a class that depends on other singletons, that is fully supported, so they will be crated as needed, there are examples in github that show it.

If there is a circular dependency of the singletons, the singleton template will detect it at initialization time and will throw an exception.

This singleton implementation also solves the multiple compile units / object files initialization order, as the required order is specified, by accessing the other singletons es::init::singleton<>::instance(). The full header file also takes care of the destruction of the singleton objects.

//
// The efficient Singleton with proper initialization and destruction order
//
// 1. Efficient, with no condition on every access
// 2. Multi dependency on other Singleton(s)
// 3. Multi thread safe
// 4. Proper initialization order with multiple compile units
// 5. Early initialization, lazy initialization
// 6. Detects circular dependency - generates exception
// 7. Proper destruction order
// 8. None intrusive, requires only default constructor, supports native types.
//
// See README.md for details.
// Discussion Proper Initialization / Destruction order:
// Simple Singleton can initialize properly when accessing one by one the singletons that require / depend on others.
// As each one of them is being initialized on the first instance() call.
// In order to solve their destructors order, their respective destructors need to be pushed into an atomic stack, and
// counter of active singletons should be increased. The SpecialDeleter of each one of the uniq_ptr<> of them,
// just reduces the counter of live singletons, and calls the deleter in the stack when the counter gets to Zero.
//
// reference to a singleton will be valid in the block scope.
// *** handle multiple calls to firstTimeGetInstance():
//    -- one after the other() , as the unique pointer objects are "reinitialized" by the compiler.
//         --> should return pointer to the same object.
//    -- one call during the firstTimeActive() - if same-thread - circular dependency
//
// MIT License
//
// Copyright (c) 2019,2020 Erez Strauss, erez@erezstrauss.com
//  http://github.com/erez-strauss/init_singleton/
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
// SOFTWARE.
//

#pragma once

#include <atomic>
#include <exception>
#include <iostream>
#include <mutex>
#include <string_view>
#include <thread>
#include <type_traits>

namespace es::init {

__extension__ using uint128_t = unsigned __int128;

constexpr const bool verbose_singletons{false};

static constexpr bool USE_BUILTIN_16B
{
#if defined(__GNUC__) && defined(__clang__)
    false  // clang++ uses builtin for std::atomic<__int128>
#elif defined(__GNUC__)
    true  // enable the __sync... builtin functions instead of std::atomic<> that call atomic library.
#else
    false  // other compilers
#endif
};

template<typename T>
struct sequenced_ptr
{
    union
    {
        uint128_t _unit;
        struct
        {
            T*       _p;
            uint64_t _seq;
        } _s;
    } _u;
    sequenced_ptr load() noexcept
    {
        if constexpr (USE_BUILTIN_16B)
            return sequenced_ptr{__sync_val_compare_and_swap(&this->_u._unit, 0, 0)};
        else
            return sequenced_ptr{reinterpret_cast<std::atomic<uint128_t>*>(this)->load()};
    }
    bool cas(sequenced_ptr expected_value, const sequenced_ptr new_value)
    {
        if constexpr (USE_BUILTIN_16B)
            return __sync_bool_compare_and_swap(&this->_u._unit, expected_value._u._unit, new_value._u._unit);
        else
            return reinterpret_cast<std::atomic<uint128_t>*>(this)->compare_exchange_strong(expected_value._u._unit,
                                                                                            new_value._u._unit);
    }
};
static_assert(sizeof(sequenced_ptr<int>) == 16, "Wrong size for sequenced_ptr");

struct singleton_base
{
};

struct singletons_meta_data
{
    singletons_meta_data* _next{nullptr};
    void (*_func)(singletons_meta_data*){nullptr};
    void*       _p{nullptr};
    const char* _func_name{nullptr};
    uint32_t    _useCount{0};
    uint32_t    _flags{0};
};

class singletons_counter
{
    inline static std::atomic<int64_t> global_counter{0};

public:
    static auto get() { return global_counter.load(); };

    template<typename T, template<typename TT> class EI, typename M, typename InitT>
    friend class singleton;
};

template<typename T>
struct static_obj_stack
{
    inline static sequenced_ptr<T> top{0UL};
    inline static std::mutex       stack_mutex{};

    static void push(T* p)
    {
        sequenced_ptr<T> stack_top;
        sequenced_ptr<T> new_top;
        do
        {
            stack_top = top.load();
            new_top   = stack_top;

            p->_next         = stack_top._u._s._p;
            new_top._u._s._p = p;
            ++new_top._u._s._seq;
        } while (!top.cas(stack_top, new_top));
    }
    static T* pop()
    {
        sequenced_ptr<T> stack_top;
        sequenced_ptr<T> new_top;
        do
        {
            stack_top = top.load();
            if (!stack_top._u._s._p) return nullptr;
            new_top          = stack_top;
            new_top._u._s._p = stack_top._u._s._p->_next;
            ++new_top._u._s._seq;
        } while (!top.cas(stack_top, new_top));
        return stack_top._u._s._p;
    }

    static uint64_t size()
    {
        std::lock_guard<std::mutex> guard(stack_mutex);
        uint64_t                    n{0};
        for (T* p = top._u._s._p; p != nullptr; p = p->_next) ++n;
        return n;
    }
};

using stack = static_obj_stack<singletons_meta_data>;
static inline std::atomic<bool> clean_up_phase{false};
static inline int               app_argc{0};
static inline char**            app_argv{nullptr};

inline void emptyStack()
{
    std::lock_guard<std::mutex> guard(stack::stack_mutex);

    clean_up_phase = true;

    uint64_t n{0};
    while (auto p = stack::pop())
    {
        if constexpr (es::init::verbose_singletons)
        {
            std::cout << "emptyStack[" << n++ << "]: node: " << (void*)p << " func_name: " << p->_func_name
                      << std::endl;
        }
        if (auto f = p->_func)
        {
            p->_func = nullptr;
            f(p);
        }
    }
}

inline void report_singletons_stack()
{
    std::lock_guard<std::mutex> guard(stack::stack_mutex);
    uint64_t                    n{0};
    for (singletons_meta_data* p = stack::top._u._s._p; p != nullptr; p = p->_next)
    {
        std::cout << "singletons_stack_meta_data_node[" << ++n << "]: " << (void*)p << " func_name: " << p->_func_name
                  << std::endl;
    }
}

template<typename T>
struct early_initializer_no_args
{
    [[using gnu: used, constructor]] static void early_init()
    {
        std::ios_base::Init z;
        T::instance();
    }
};

template<typename T>
struct early_initializer
{
    [[using gnu: used, constructor]] static void early_init(int argc, char** argv)
    {
        std::ios_base::Init z;
        if (es::init::app_argc != argc) es::init::app_argc = argc;
        if (es::init::app_argv != argv) es::init::app_argv = argv;
        T::instance();
    }
};

template<typename T>
struct lazy_initializer
{  // empty - do nothing.
};

template<typename T, template<typename TT> class EI = early_initializer, typename M = void,
         typename InitT = ::std::ios_base::Init>
class singleton : public singleton_base, EI<singleton<T, EI, M, InitT>>
{
    struct SpecialDeleter
    {
        void operator()(T* p) const
        {
            static bool activated{false};
            if constexpr (verbose_singletons)
            {
                std::cout << "SpecialDeleter start " << singletons_counter::global_counter.load() << " "
                          << __PRETTY_FUNCTION__ << " " << (void*)p << " " << (activated ? 'T' : 'F') << std::endl;
            }
            activated = true;
            if (p)
            {
                std::lock_guard<std::mutex> guard(_mutex);
                if (_getInstance == optGetInstance) _getInstance = firstTimeGetInstance;
                singleton_meta_data_node._p = (void*)p;
                --singletons_counter::global_counter;
                if constexpr (es::init::verbose_singletons)
                {
                    std::cout << "SpecialDeleter - done " << singletons_counter::global_counter.load() << " "
                              << __PRETTY_FUNCTION__ << std::endl;
                }
                if (!singletons_counter::global_counter) emptyStack();
            }
        }
    };

    static void activeDelete(singletons_meta_data* np)
    {
        static uint64_t active_delete_count{0};
        ++active_delete_count;
        if constexpr (es::init::verbose_singletons)
        {
            std::cerr << "activeDelete - " << __PRETTY_FUNCTION__ << " " << (void*)np->_p << " " << active_delete_count
                      << std::endl;
        }
        delete static_cast<T*>(np->_p);
        np->_p = nullptr;
    }

    static T& firstTimeGetInstance()
    {
        volatile InitT init_object{};  // make sure, one can use the std::cout std::cerr streams from Singletons code

        if (clean_up_phase)
            std::cerr << "Warning: initializing at clean up phase - " << __PRETTY_FUNCTION__ << std::endl;

        if (!_instance)
        {
            if (singleton_meta_data_node._flags & 0x1)
            {
                throw std::logic_error(std::string{"Error: circular dependency "} + __PRETTY_FUNCTION__);
            }
            std::lock_guard<std::mutex> guard(_mutex);
            if (!_instance)
            {
                singleton_meta_data_node._flags = 0x1;
                _instance                       = std::unique_ptr<T, SpecialDeleter>{new T{}, SpecialDeleter{}};
                ++singletons_counter::global_counter;

                if (!singleton_meta_data_node._p)
                {
                    singleton_meta_data_node._func      = activeDelete;
                    singleton_meta_data_node._p         = (void*)&*_instance;
                    singleton_meta_data_node._func_name = __PRETTY_FUNCTION__;
                    stack::push(&singleton_meta_data_node);
                }
            }
        }
        _getInstance = optGetInstance;
        singleton_meta_data_node._flags &= ~0x1U;

        return *_instance;
    }

    static T& optGetInstance() { return *_instance; }

    inline static std::atomic<T& (*)()>              _getInstance{firstTimeGetInstance};
    inline static std::unique_ptr<T, SpecialDeleter> _instance{nullptr};
    inline static std::mutex                         _mutex{};
    inline static singletons_meta_data               singleton_meta_data_node{nullptr, nullptr, nullptr, nullptr, 0, 0};

    static_assert(sizeof(_instance) == 8, "instace size should be 8 bytes");

public:
    [[using gnu: hot]] static T& instance()
    {
        auto f = _getInstance.load();
        return f();
    }
};

}  // namespace es::init

Here are two use cases, more examples are available on the github repo:

  • Initialization order of application components holding references or using other components at init time, using singlton<> of different types. The following example (examples/singleton10.cpp) shows few classes (I use struct to simpify the example text)

    • ComponentE - has no dependencies - used as singleton inside ComponentC constructor
    • ComponentD - has no dependencies - used as singleton as initialization value to ComponenD reference in componentC
    • ComponentC - depends on D and E
    • ComponentB - depend on C, as it access the C singleton
    • ComponentA - depends on B, and indirectly on all the others
    • Engine - depends on ComponentA

    All the above component can be in a single cpp file or each in its own cpp files, and their initialization order is well defined, and does not depent on linkage order.

#include <app_singletons.h>
#include <iostream>

template<typename T>
struct CDReporter
{
CDReporter() { std::cout << "inside: " << __PRETTY_FUNCTION__ << " this: " << (void*)this << std::endl; }
~CDReporter() { std::cout << "inside: " << __PRETTY_FUNCTION__ << " this: " << (void*)this << std::endl; }
};

struct ComponentE : public CDReporter<ComponentE> { ComponentE() { } };
struct ComponentD : public CDReporter<ComponentD> { ComponentD() { } };
struct ComponentC : public CDReporter<ComponentC> { ComponentC() : _d(es::init::singleton<ComponentD>::instance()) { es::init::singleton<ComponentE>::instance(); } ComponentD& _d; };
struct ComponentB : public CDReporter<ComponentB> { ComponentB() { es::init::singleton<ComponentC>::instance(); } };
struct ComponentA : public CDReporter<ComponentA> { ComponentA() { es::init::singleton<ComponentB>::instance(); } };
class Engine { public: Engine() : _a(es::init::singleton<ComponentA>::instance()) {} ComponentA& _a; };

int main() { Engine e{}; return 0; }
  • Deep nested function that internally needs access to a singleton information for decision

    This example9.cpp shows access to command-line arguments from the SetupInfo object. It eliminate the need to pass references to singleton<> objects from containing Objects down to their internal objects (ProcessorA - internal to B, while B internal to C) As this implementation is faster than others it reduces the overhead of accessing singletons objects.

#include <app_singletons.h>
#include <iostream>

class SetupInfo {
    bool _verbose{false};
public:
    SetupInfo() {
        es::init::args.for_each([&](auto, auto opt) {
            if (opt && opt[0] == '-' && opt[1] == 'v' && !opt[2]) _verbose = true; });
    }
    bool verbose() { return _verbose; }
};

class ProcessorA {
public:
    int action() {
        if (es::init::singleton<SetupInfo>::instance().verbose()) {
            // real action.
            std::cout << "verbose\n";
            return 1;
        }
        return 0;
    }
};

class ProcessorB {
public:
    ProcessorB(ProcessorA& pa) : _processorA(pa) {}
    int         action() { return _processorA.action(); }
    ProcessorA& _processorA;
};

class ProcessorC {
public:
    ProcessorC(ProcessorB& pb) : _processorB(pb) {}
    int         action() { return _processorB.action(); }
    ProcessorB& _processorB;
};

int main()
{
    ProcessorA a;
    ProcessorB b{a};
    ProcessorC c{b};

    c.action();
}

My questions:

  1. Please review the above code and suggest improvements and changes.
  2. Does the above singleton pattern implementation follows the current C++17/C++2a best practice.
  3. How to improve it, and make it more useful?
  4. How to make it portable, Windows environment and intel compiler?

Thanks, Erez

\$\endgroup\$
  • 2
    \$\begingroup\$ Can you also specify some concrete use cases, where using that code should be used in production please? \$\endgroup\$ – πάντα ῥεῖ Jan 4 at 17:22
  • \$\begingroup\$ I added the use cases to the question text above, thank you. \$\endgroup\$ – user5673656 Jan 5 at 21:24
  • \$\begingroup\$ The atomic pointer doesn't do much. Simply follow Scott Meyer's singleton, approach to achieve thread safety of creation of the instance. \$\endgroup\$ – πάντα ῥεῖ Jan 5 at 21:31
  • \$\begingroup\$ πάντα ῥεῖ - using atomic<> is not intended to provide locking, the creation of the specific object is done under proper mutex lock, only once per Singleton at init time. The atomic<> usage prevents the compiler from over optimization, and in case of lazy_initializer, one thread will to the firstTimeGetInstance() and the other threads will do the optimized version. \$\endgroup\$ – user5673656 Jan 6 at 2:17
  • \$\begingroup\$ "That is a singleton of a type int, it is being created before the main() starts, and destroyed after main() exists." Before main starts? How are the other examples processing command line arguments if they don't get them from main? early_init never seems to be called either? What compiler / platform does this compile on? \$\endgroup\$ – user673679 Jan 6 at 17:50
1
\$\begingroup\$

style nit:

inline static singletons_meta_data singleton_meta_data_node{nullptr, nullptr, nullptr, nullptr, 0, 0};

Specifying all the member values like this decreases readability. Setting everything to nullptrs doesn't tell the reader anything useful. Either explicitly set the members like ._next = nullptr, or better, just let it default construct.

Also, does inline do anything here?


Isn't it a problem when static_obj_stack is destroyed? The mutex destructor would be called and nothing else would be able to be use it anymore. _instances would call emptyStack which uses static_obj_stack. AFAIK, static destruction order is not guaranteed.


    [[using gnu: hot]] static T& instance()
    {
        auto f = _getInstance.load();
        return f();
    }

Can you explain how this works? If two threads enter here, they could both load firstTimeGetInstance.

Then inside that method it looks like it will erroneously throw a circular dependency error, because one thread might set the flag to 1 and the other might test it:

        if (!_instance)
        {
            if (singleton_meta_data_node._flags & 0x1)
            {
                throw std::logic_error(std::string{"Error: circular dependency "} + __PRETTY_FUNCTION__);
            }
            std::lock_guard<std::mutex> guard(_mutex);
            if (!_instance)
            {
                singleton_meta_data_node._flags = 0x1;
                _instance                       = std::unique_ptr<T, SpecialDeleter>{new T{}, SpecialDeleter{}};
                ++singletons_counter::global_counter;

                if (!singleton_meta_data_node._p)
                {
                    singleton_meta_data_node._func      = activeDelete;
                    singleton_meta_data_node._p         = (void*)&*_instance;
                    singleton_meta_data_node._func_name = __PRETTY_FUNCTION__;
                    stack::push(&singleton_meta_data_node);
                }
            }
        }
| improve this answer | |
\$\endgroup\$
  • \$\begingroup\$ the singleton_meta_data_node should be plain old data, with no function initializer, but simple data. the 'inline' enable multiple header file to user these class static variable with initialization value. regarding the _flags=01 - thanks, fixed it in github repo, unsetting the flag bit, inside the guard scope. \$\endgroup\$ – user5673656 Jan 7 at 3:20
  • \$\begingroup\$ singletons_meta_data_node still has a generated constructor that you can use, and you're even explicitly setting all the members to 0s. Both of these are unnecessary: en.cppreference.com/w/cpp/language/initialization For all other non-local static and thread-local variables, Zero initialization takes place \$\endgroup\$ – butt Jan 7 at 22:17
  • \$\begingroup\$ I changed the mutex to a trivially constructed spin lock, so there is no generated function that is being called. Regarding the initialization to zero, thank you, I'm in this case prefer to state the obvious, and be explicit, to avoid later on someone "fixing" it with constructor values which will mess things. regarding two threads getting into the firstTimeGetInstance() function, only one of them will get the lock (on the spin lock, see github code) and will create the object, the other will spin, waiting for the object to be created, then return the reference to it. \$\endgroup\$ – user5673656 Jan 14 at 3:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.