Smart Pointer Programming Techniques

Using incomplete classes for implementation hiding
The "Pimpl" idiom
Using abstract classes for implementation hiding
Preventing delete px.get()
Using a shared_ptr to hold a pointer to an array
Encapsulating allocation details, wrapping factory functions
Using a shared_ptr to hold a pointer to a statically allocated object
Using a shared_ptr to hold a pointer to a COM object
Using a shared_ptr to hold a pointer to an object with an embedded reference count
Using a shared_ptr to hold another shared ownership smart pointer
Obtaining a shared_ptr from a raw pointer
Obtaining a shared_ptr (weak_ptr) to this in a constructor
Obtaining a shared_ptr to this
Using shared_ptr as a smart counted handle
Using shared_ptr to execute code on block exit
Using shared_ptr<void> to hold an arbitrary object
Associating arbitrary data with heterogeneous shared_ptr instances
Using shared_ptr as a CopyConstructible mutex lock
Using shared_ptr to wrap member function calls
Delayed deallocation
Weak pointers to objects not managed by a shared_ptr

Using incomplete classes for implementation hiding

A proven technique (that works in C, too) for separating interface from implementation is to use a pointer to an incomplete class as an opaque handle:

class FILE;

FILE * fopen(char const * name, char const * mode);
void fread(FILE * f, void * data, size_t size);
void fclose(FILE * f);

It is possible to express the above interface using shared_ptr, eliminating the need to manually call fclose:

class FILE;

shared_ptr<FILE> fopen(char const * name, char const * mode);
void fread(shared_ptr<FILE> f, void * data, size_t size);

This technique relies on shared_ptr's ability to execute a custom deleter, eliminating the explicit call to fclose, and on the fact that shared_ptr<X> can be copied and destroyed when X is incomplete.

The "Pimpl" idiom

A C++ specific variation of the incomplete class pattern is the "Pimpl" idiom. The incomplete class is not exposed to the user; it is hidden behind a forwarding facade. shared_ptr can be used to implement a "Pimpl":

// file.hpp:

class file
{
private:

    class impl;
    shared_ptr<impl> pimpl_;

public:

    file(char const * name, char const * mode);

    // compiler generated members are fine and useful

    void read(void * data, size_t size);
};

// file.cpp:

#include "file.hpp"

class file::impl
{
private:

    impl(impl const &);
    impl & operator=(impl const &);

    // private data

public:

    impl(char const * name, char const * mode) { ... }
    ~impl() { ... }
    void read(void * data, size_t size) { ... }
};

file::file(char const * name, char const * mode): pimpl_(new impl(name, mode))
{
}

void file::read(void * data, size_t size)
{
    pimpl_->read(data, size);
}

The key thing to note here is that the compiler-generated copy constructor, assignment operator, and destructor all have a sensible meaning. As a result, file is CopyConstructible and Assignable, allowing its use in standard containers.

Using abstract classes for implementation hiding

Another widely used C++ idiom for separating inteface and implementation is to use abstract base classes and factory functions. The abstract classes are sometimes called "interfaces" and the pattern is known as "interface-based programming". Again, shared_ptr can be used as the return type of the factory functions:

// X.hpp:

class X
{
public:

    virtual void f() = 0;
    virtual void g() = 0;

protected:

    ~X() {}
};

shared_ptr<X> createX();

-- X.cpp:

class X_impl: public X
{
private:

    X_impl(X_impl const &);
    X_impl & operator=(X_impl const &);

public:

    virtual void f()
    {
      // ...
    }

    virtual void g()
    {
      // ...
    }
};

shared_ptr<X> createX()
{
    shared_ptr<X> px(new X_impl);
    return px;
}

A key property of shared_ptr is that the allocation, construction, deallocation, and destruction details are captured at the point of construction, inside the factory function. Note the protected and nonvirtual destructor in the example above. The client code cannot, and does not need to, delete a pointer to X; the shared_ptr<X> instance returned from createX will correctly call ~X_impl.

Preventing `delete px.get()`

It is often desirable to prevent client code from deleting a pointer that is being managed by shared_ptr. The previous technique showed one possible approach, using a protected destructor. Another alternative is to use a private deleter:

class X
{
private:

    ~X();

    class deleter;
    friend class deleter;

    class deleter
    {
    public:

        void operator()(X * p) { delete p; }
    };

public:

    static shared_ptr<X> create()
    {
        shared_ptr<X> px(new X, X::deleter());
        return px;
    }
};

Using a `shared_ptr` to hold a pointer to an array

A shared_ptr can be used to hold a pointer to an array allocated with new[]:

shared_ptr<X> px(new X[1], checked_array_deleter<X>());

Note, however, that shared_array is often preferable, if this is an option. It has an array-specific interface, without operator* and operator->, and does not allow pointer conversions.

Encapsulating allocation details, wrapping factory functions

shared_ptr can be used in creating C++ wrappers over existing C style library interfaces that return raw pointers from their factory functions to encapsulate allocation details. As an example, consider this interface, where CreateX might allocate X from its own private heap, ~X may be inaccessible, or X may be incomplete:

X * CreateX();
void DestroyX(X *);

The only way to reliably destroy a pointer returned by CreateX is to call DestroyX.

Here is how a shared_ptr-based wrapper may look like:

shared_ptr<X> createX()
{
    shared_ptr<X> px(CreateX(), DestroyX);
    return px;
}

Client code that calls createX still does not need to know how the object has been allocated, but now the destruction is automatic.

Using a `shared_ptr` to hold a pointer to a statically allocated object

Sometimes it is desirable to create a shared_ptr to an already existing object, so that the shared_ptr does not attempt to destroy the object when there are no more references left. As an example, the factory function:

shared_ptr<X> createX();

in certain situations may need to return a pointer to a statically allocated X instance.

The solution is to use a custom deleter that does nothing:

struct null_deleter
{
    void operator()(void const *) const
    {
    }
};

static X x;

shared_ptr<X> createX()
{
    shared_ptr<X> px(&x, null_deleter());
    return px;
}

The same technique works for any object known to outlive the pointer.

Using a `shared_ptr` to hold a pointer to a COM Object

Background: COM objects have an embedded reference count and two member functions that manipulate it. AddRef() increments the count. Release() decrements the count and destroys itself when the count drops to zero.

It is possible to hold a pointer to a COM object in a shared_ptr:

shared_ptr<IWhatever> make_shared_from_COM(IWhatever * p)
{
    p->AddRef();
    shared_ptr<IWhatever> pw(p, mem_fn(&IWhatever::Release));
    return pw;
}

Note, however, that shared_ptr copies created from pw will not "register" in the embedded count of the COM object; they will share the single reference created in make_shared_from_COM. Weak pointers created from pw will be invalidated when the last shared_ptr is destroyed, regardless of whether the COM object itself is still alive.

As explained in the mem_fn documentation, you need to #define BOOST_MEM_FN_ENABLE_STDCALL first.

Using a `shared_ptr` to hold a pointer to an object with an embedded reference count

This is a generalization of the above technique. The example assumes that the object implements the two functions required by intrusive_ptr, intrusive_ptr_add_ref and intrusive_ptr_release:

template<class T> struct intrusive_deleter
{
    void operator()(T * p)
    {
        if(p) intrusive_ptr_release(p);
    }
};

shared_ptr<X> make_shared_from_intrusive(X * p)
{
    if(p) intrusive_ptr_add_ref(p);
    shared_ptr<X> px(p, intrusive_deleter<X>());
    return px;
}

Using a `shared_ptr` to hold another shared ownership smart pointer

One of the design goals of shared_ptr is to be used in library interfaces. It is possible to encounter a situation where a library takes a shared_ptr argument, but the object at hand is being managed by a different reference counted or linked smart pointer.

It is possible to exploit shared_ptr's custom deleter feature to wrap this existing smart pointer behind a shared_ptr facade:

template<class P> struct smart_pointer_deleter
{
private:

    P p_;

public:

    smart_pointer_deleter(P const & p): p_(p)
    {
    }

    void operator()(void const *)
    {
        p_.reset();
    }
    
    P const & get() const
    {
        return p_;
    }
};

shared_ptr<X> make_shared_from_another(another_ptr<X> qx)
{
    shared_ptr<X> px(qx.get(), smart_pointer_deleter< another_ptr<X> >(qx));
    return px;
}

One subtle point is that deleters are not allowed to throw exceptions, and the above example as written assumes that p_.reset() doesn't throw. If this is not the case, p_.reset() should be wrapped in a try {} catch(...) {} block that ignores exceptions. In the (usually unlikely) event when an exception is thrown and ignored, p_ will be released when the lifetime of the deleter ends. This happens when all references, including weak pointers, are destroyed or reset.

Another twist is that it is possible, given the above shared_ptr instance, to recover the original smart pointer, using get_deleter:

void extract_another_from_shared(shared_ptr<X> px)
{
    typedef smart_pointer_deleter< another_ptr<X> > deleter;

    if(deleter const * pd = get_deleter<deleter>(px))
    {
        another_ptr<X> qx = pd->get();
    }
    else
    {
        // not one of ours
    }
}

Obtaining a `shared_ptr` from a raw pointer

Sometimes it is necessary to obtain a shared_ptr given a raw pointer to an object that is already managed by another shared_ptr instance. Example:

void f(X * p)
{
    shared_ptr<X> px(???);
}

Inside f, we'd like to create a shared_ptr to *p.

In the general case, this problem has no solution. One approach is to modify f to take a shared_ptr, if possible:

void f(shared_ptr<X> px);

The same transformation can be used for nonvirtual member functions, to convert the implicit this:

void X::f(int m);

would become a free function with a shared_ptr first argument:

void f(shared_ptr<X> this_, int m);

If f cannot be changed, but X uses intrusive counting, use make_shared_from_intrusive described above. Or, if it's known that the shared_ptr created in f will never outlive the object, use a null deleter.

Obtaining a `shared_ptr` (`weak_ptr`) to `this` in a constructor

Some designs require objects to register themselves on construction with a central authority. When the registration routines take a shared_ptr, this leads to the question how could a constructor obtain a shared_ptr to this:

class X
{
public:

    X()
    {
        shared_ptr<X> this_(???);
    }
};

In the general case, the problem cannot be solved. The X instance being constructed can be an automatic variable or a static variable; it can be created on the heap:

shared_ptr<X> px(new X);

but at construction time, px does not exist yet, and it is impossible to create another shared_ptr instance that shares ownership with it.

Depending on context, if the inner shared_ptr this_ doesn't need to keep the object alive, use a null_deleter as explained here and here. If X is supposed to always live on the heap, and be managed by a shared_ptr, use a static factory function:

class X
{
private:

    X() { ... }

public:

    static shared_ptr<X> create()
    {
        shared_ptr<X> px(new X);
        // use px as 'this_'
        return px;
    }
};

Obtaining a `shared_ptr` to `this`

Sometimes it is needed to obtain a shared_ptr from this in a virtual member function under the assumption that this is already managed by a shared_ptr. The transformations described in the previous technique cannot be applied.

A typical example:

class X
{
public:

    virtual void f() = 0;

protected:

    ~X() {}
};

class Y
{
public:

    virtual shared_ptr<X> getX() = 0;

protected:

    ~Y() {}
};

// --

class impl: public X, public Y
{
public:

    impl() { ... }

    virtual void f() { ... }

    virtual shared_ptr<X> getX()
    {
        shared_ptr<X> px(???);
        return px;
    }
};

The solution is to keep a weak pointer to this as a member in impl:

class impl: public X, public Y
{
private:

    weak_ptr<impl> weak_this;

    impl(impl const &);
    impl & operator=(impl const &);

    impl() { ... }

public:

    static shared_ptr<impl> create()
    {
        shared_ptr<impl> pi(new impl);
        pi->weak_this = pi;
        return pi;
    }

    virtual void f() { ... }

    virtual shared_ptr<X> getX()
    {
        shared_ptr<X> px(weak_this);
        return px;
    }
};

The library now includes a helper class template enable_shared_from_this that can be used to encapsulate the solution:

class impl: public X, public Y, public enable_shared_from_this<impl>
{
public:

    impl(impl const &);
    impl & operator=(impl const &);

public:

    virtual void f() { ... }

    virtual shared_ptr<X> getX()
    {
        return shared_from_this();
    }
}

Note that you no longer need to manually initialize the weak_ptr member in enable_shared_from_this. Constructing a shared_ptr to impl takes care of that.

Using `shared_ptr` as a smart counted handle

Some library interfaces use opaque handles, a variation of the incomplete class technique described above. An example:

typedef void * HANDLE;
HANDLE CreateProcess();
void CloseHandle(HANDLE);

Instead of a raw pointer, it is possible to use shared_ptr as the handle and get reference counting and automatic resource management for free:

typedef shared_ptr<void> handle;

handle createProcess()
{
    shared_ptr<void> pv(CreateProcess(), CloseHandle);
    return pv;
}

Using `shared_ptr` to execute code on block exit

shared_ptr<void> can automatically execute cleanup code when control leaves a scope.

Executing f(p), where p is a pointer:

    shared_ptr<void> guard(p, f);

Executing arbitrary code: f(x, y):

    shared_ptr<void> guard(static_cast<void*>(0), bind(f, x, y));

For a more thorough treatment, see the article "Simplify Your Exception-Safe Code" by Andrei Alexandrescu and Petru Marginean, available online at http://www.cuj.com/experts/1812/alexandr.htm?topic=experts.

Using `shared_ptr<void>` to hold an arbitrary object

shared_ptr<void> can act as a generic object pointer similar to void*. When a shared_ptr<void> instance constructed as:

    shared_ptr<void> pv(new X);

is destroyed, it will correctly dispose of the X object by executing ~X.

This propery can be used in much the same manner as a raw void* is used to temporarily strip type information from an object pointer. A shared_ptr<void> can later be cast back to the correct type by using static_pointer_cast.

Associating arbitrary data with heterogeneous `shared_ptr` instances

shared_ptr and weak_ptr support operator< comparisons required by standard associative containers such as std::map. This can be used to non-intrusively associate arbitrary data with objects managed by shared_ptr:

typedef int Data;

std::map< shared_ptr<void>, Data > userData;
// or std::map< weak_ptr<void>, Data > userData; to not affect the lifetime

shared_ptr<X> px(new X);
shared_ptr<int> pi(new int(3));

userData[px] = 42;
userData[pi] = 91;

Using `shared_ptr` as a CopyConstructible mutex lock

Sometimes it's necessary to return a mutex lock from a function, and a noncopyable lock cannot be returned by value. It is possible to use shared_ptr as a mutex lock:

class mutex
{
public:

    void lock();
    void unlock();
};

shared_ptr<mutex> lock(mutex & m)
{
    m.lock();
    return shared_ptr<mutex>(&m, mem_fn(&mutex::unlock));
}

Better yet, the shared_ptr instance acting as a lock can be encapsulated in a dedicated shared_lock class:

class shared_lock
{
private:

    shared_ptr<void> pv;

public:

    template<class Mutex> explicit shared_lock(Mutex & m): pv((m.lock(), &m), mem_fn(&Mutex::unlock)) {}
};

shared_lock can now be used as:

    shared_lock lock(m);

Note that shared_lock is not templated on the mutex type, thanks to shared_ptr<void>'s ability to hide type information.

Using `shared_ptr` to wrap member function calls

shared_ptr implements the ownership semantics required from the Wrap/CallProxy scheme described in Bjarne Stroustrup's article "Wrapping C++ Member Function Calls" (available online at http://www.research.att.com/~bs/wrapper.pdf). An implementation is given below:

template<class T> class pointer
{
private:

    T * p_;

public:

    explicit pointer(T * p): p_(p)
    {
    }

    shared_ptr<T> operator->() const
    {
        p_->prefix();
        return shared_ptr<T>(p_, mem_fn(&T::suffix));
    }
};

class X
{
private:

    void prefix();
    void suffix();
    friend class pointer<X>;
    
public:

    void f();
    void g();
};

int main()
{
    X x;

    pointer<X> px(&x);

    px->f();
    px->g();
}

Delayed deallocation

In some situations, a single px.reset() can trigger an expensive deallocation in a performance-critical region:

class X; // ~X is expensive

class Y
{
    shared_ptr<X> px;

public:

    void f()
    {
        px.reset();
    }
};

The solution is to postpone the potential deallocation by moving px to a dedicated free list that can be periodically emptied when performance and response times are not an issue:

vector< shared_ptr<void> > free_list;

class Y
{
    shared_ptr<X> px;

public:

    void f()
    {
        free_list.push_back(px);
        px.reset();
    }
};

// periodically invoke free_list.clear() when convenient

Another variation is to move the free list logic to the construction point by using a delayed deleter:

struct delayed_deleter
{
    template<class T> void operator()(T * p)
    {
        try
        {
            shared_ptr<void> pv(p);
            free_list.push_back(pv);
        }
        catch(...)
        {
        }
    }
};

Weak pointers to objects not managed by a `shared_ptr`

Make the object hold a shared_ptr to itself, using a null_deleter:

class X
{
private:

    shared_ptr<X> this_;
    int i_;

public:

    explicit X(int i): this_(this, null_deleter()), i_(i)
    {
    }

    // repeat in all constructors (including the copy constructor!)

    X(X const & rhs): this_(this, null_deleter()), i_(rhs.i_)
    {
    }

    // do not forget to not assign this_ in the copy assignment

    X & operator=(X const & rhs)
    {
	    i_ = rhs.i_;
    }

    weak_ptr<X> get_weak_ptr() const { return this_; }
};

When the object's lifetime ends, X::this_ will be destroyed, and all weak pointers will automatically expire.

$Date: 2006/11/09 20:22:24 $