4
votes

In this thread, the following is said about singleton instances:

The static variable can be static to the GetInstance() function, or it can be static in the Singleton class. There's interesting tradeoffs there.

What are these trade-offs? I am aware that, if declared as a static function variable, the singleton won't be constructed until the function is first called. I've also read something about thread-safety, but am unaware of what exactly that entails, or how the two approaches differ in that regard.

Are there any other major differences between the two? Which approach is better?

In my concrete example, I have a factory class set up as a singleton, and I'm storing the instance as a static const field in the class. I don't have a getInstance() method, but rather expect the user to access the instance directly, like so: ItemFactory::factory. The default constructor is private, and the instance is allocated statically.

Addendum: how good of an idea is it to overload operator() to call the createItem() method for the singleton, such that Items can be created like so: ItemFactory::factory("id")?

3
See this for thread-safety for static function variables. I would say static function variables are the way to go, because with that way you are assured of the correct initialization order.Neijwiert

3 Answers

2
votes

What are these trade-offs?

This is the most important consideration:

The static data member is initialized during the static initialization at the start of the program. If any static object depends on the singleton, then there will be a static initialization order fiasco.

The function local static object is initialized when the function is first called. Since whoever depends on the singleton will call the function, the singleton will be appropriately initialized and is not susceptible to the fiasco. There is still a - very subtle - problem with the destruction. If a destructor of a static object depends on the singleton, but the constructor of that object does not, then you'll end up with undefined behaviour.

Also, being initialized on the first time the function is called, means that the function may be called after the static initialization is done and main has been called. And therefore, the program may have spawned multiple threads. There could be a race condition on the initialization of the static local, resulting in multiple instances being constructed. Luckily, since C++11, the standard guarantees that the initialization is thread safe and this tradeoff no longer exists in conforming compilers.

Thread safety is not an issue with the static data member.

Which approach is better?

That depends on what your requirements are and what version of the standard you support.

2
votes

I vote for static function variable. The newer C++ standard require automatic thread safety for initialization of such variables. It's implemented in GNU C++ for about ten years already. Visual Studio 2015 also supports this. If you make a static pointer variable holding reference to your singleton object, you'll have to deal with thread issues manually.

In the other hand, if you make a static member pointer field like shown in in the snippet below, you will be able to change it from other static methods, maybe re-init this field with other instance upon handling request to change program configuration. However, the snippet below contains a bug just to remind you how difficult multithreading is.

class ItemFactory {
static std::atomic_flag initialized = ATOMIC_FLAG_INIT;
static std::unique_ptr<ItemFactory> theFactoryInstance;
public:
static ItemFactory& getInstance() {
  if (!initialized.test_and_set(std::memory_order_acquire)) {
    theFactoryInstance = std::make_unique<ItemFactory>();
  }
  return *theFactoryInstance;
}
};

I wouldn't advise you to implement your singleton as a global non-pointer variable initialized before entry to the main() function. Thread safety issues will go away along with implicit cache coherency overhead, but you're not able to control the initialization order of your global variables in any precise or portable way.

Anyway, this choice doesn't force any permanent design implications. Since this instance will reside in the private section of your class you may always change it.

I don't think overloading of operator() for a factory is a good idea. operator() have "execute" semantics while in factory it's gonna stand for "create".

1
votes

What is the best approach to a singleton in c++?

Hide the fact that it's a singleton and give it value semantics.

How?

All singleton-ness ought to be an implementation detail. In this way, consumers of your class need not refactor their programs if you need to change the way you implement your singleton (or indeed if you decide that it should not really be a singleton after all).

Why ?

Because now your program never has to worry itself with references, pointers, lifetimes and whatnot. It just uses an instance of the object as if it were a value. Safe in the knowledge that the singleton will take care of whatever lifetime/resource requirements it has.

What about a singleton that releases resources when not in use?

no problem.

Here's an example of the two approaches hidden behind the facade of an object with value semantics.

imagine this use case:

auto j1 = jobbie();
auto j2 = jobbie();
auto j3 = jobbie();

j1.log("doh");
j2.log("ray");
j3.log("me");

{
    shared_file f;
    f.log("hello");
}

{
    shared_file().log("goodbye");
}

shared_file().log("here's another");

shared_file f2;
{
    shared_file().log("no need to reopen");
    shared_file().log("or here");
    shared_file().log("or even here");
}
f2.log("all done");

where a jobbie object is just a facade for a singleton, but the shared_file object wants to flush/close itself when not in use.

so the output should look like this:

doh
ray
me
opening file
logging to file: hello
closing file
opening file
logging to file: goodbye
closing file
opening file
logging to file: here's another
closing file
opening file
logging to file: no need to reopen
logging to file: or here
logging to file: or even here
logging to file: all done
closing file

We can achieve this using the idiom, which I'll call 'value-semantics-is-a-facade-for-singleton':

#include <iostream>
#include <vector>

// interface
struct jobbie
{
    void log(const std::string& s);

private:
    // if we decide to make jobbie less singleton-like in future
    // then as far as the interface is concerned the only change is here
    // and since these items are private, it won't matter to consumers of the class
    struct impl;
    static impl& get();
};

// implementation

struct jobbie::impl
{
    void log(const std::string& s) {
        std::cout << s << std::endl;
    }
};

auto jobbie::get() -> impl& {
    //
    // NOTE
    // now you can change the singleton storage strategy simply by changing this code
    // alternative 1:
    static impl _;
    return _;

    // for example, we could use a weak_ptr which we lock and store the shared_ptr in the outer
    // jobbie class. This would give us a shared singleton which releases resources when not in use

}

// implement non-singleton interface

void jobbie::log(const std::string& s)
{
    get().log(s);
}

struct shared_file
{
    shared_file();

    void log(const std::string& s);

private:
    struct impl;
    static std::shared_ptr<impl> get();
    std::shared_ptr<impl> _impl;
};

// private implementation

struct shared_file::impl {

    // in a multithreaded program
    // we require a condition variable to ensure that the shared resource is closed
    // when we try to re-open it (race condition)
    struct statics {
        std::mutex m;
        std::condition_variable cv;
        bool still_open = false;
        std::weak_ptr<impl> cache;
    };

    static statics& get_statics() {
        static statics _;
        return _;
    }

    impl() {
        std::cout << "opening file\n";
    }
    ~impl() {
        std::cout << "closing file\n";
        // close file here
        // and now that it's closed, we can signal the singleton state that it can be
        // reopened

        auto& stats = get_statics();

        // we *must* use a lock otherwise the compiler may re-order memory access
        // across the memory fence
        auto lock = std::unique_lock<std::mutex>(stats.m);
        stats.still_open = false;
        lock.unlock();
        stats.cv.notify_one();
    }
    void log(const std::string& s) {
        std::cout << "logging to file: " << s << std::endl;
    }
};

auto shared_file::get() -> std::shared_ptr<impl>
{
    auto& statics = impl::get_statics();

    auto lock = std::unique_lock<std::mutex>(statics.m);
    std::shared_ptr<impl> candidate;
    statics.cv.wait(lock, [&statics, &candidate] {
        return bool(candidate = statics.cache.lock())
        or not statics.still_open;
    });
    if (candidate)
        return candidate;

    statics.cache = candidate = std::make_shared<impl>();
    statics.still_open = true;
    return candidate;
}


// interface implementation

shared_file::shared_file() : _impl(get()) {}
void shared_file::log(const std::string& s) { _impl->log(s); }

// test our class
auto main() -> int
{
    using namespace std;

    auto j1 = jobbie();
    auto j2 = jobbie();
    auto j3 = jobbie();

    j1.log("doh");
    j2.log("ray");
    j3.log("me");

    {
        shared_file f;
        f.log("hello");
    }

    {
        shared_file().log("goodbye");
    }

    shared_file().log("here's another");

    shared_file f2;
    {
        shared_file().log("no need to reopen");
        shared_file().log("or here");
        shared_file().log("or even here");
    }
    f2.log("all done");


    return 0;
}