Improving shared_ptr for C++0x, Revision 2

• Overview
• Allocator Support
• Aliasing Support
• Object Creation
• Move Support
• Atomic Access
• Cycle Collection

Changes in Revision 1

• Expanded the Allocator Support section with more rationale.
• Clarified that the proposed changes are source- and binary-compatible.
• Added rationale for the make_shared name.
• Changed the make_shared proposed wording to use variadic templates and changed the Throws clause to reflect the semantics. Remarks added.

I. Overview

While shared_ptr has already proven its utility time and again, we have observed frequent requests for enhancements in several legitimate and key areas:

• Ability to control its internal allocations;
• Aliasing support;
• A factory function that eliminates the need for an explicit new expression;
• Reduced reference counting overhead when using shared_ptr in "mostly unique" ownership situations such as containers;
• A variant of shared_ptr that is atomic, that is, safe to be manipulated from multiple threads without synchronization;
• A way to reclaim cyclic shared_ptr structures;
• A variant of shared_ptr that is confined to a single thread and uses unsynchronized reference count updates.

This document proposes additions to the C++0x standard to address the first six. Some of the proposed additions are essentially a subset of those presented in N1851 by Kliatchko and Rocha. We refer the interested reader to that paper for a more extensive rationale.

This proposal makes use of variadic templates and rvalue references. These C++0x features markedly improve the usability and effectiveness of several of the proposed additions.

II. Allocator Support

The default behavior of shared_ptr is to allocate its control block using new. This precludes its use in contexts where uncontrolled dynamic allocations are not allowed. The proposed addition allows the user to supply an allocator that shared_ptr will use.

Boost users have repeatedly been asking for the ability to control the internal control block allocation; one particularly amusing occurence was the desire to use a shared_ptr in the implementation of ::operator new. Typical environments where shared_ptr cannot be used as-is include embedded systems and computer and video games. Paul Pedriana (one of the primary authors of the Electronic Arts STL implementation) writes in N2271 that

EASTL's shared_ptr/weak_ptr allow the user to specify an allocator instead of implicitly using global new

and explains that

As described in the game software issues section, global new usage is often verboten in game software development, at least for console platforms. Thus any library facility which uses global operator new or any memory allocation that cannot be controlled by the user is unacceptable.

Asked to comment on the proposed addition, Paul kindly offered the following quote:

Certainly we (I can speak for all of Electronic Arts here) are in support of your proposal and you can document it as such. You could probably get other game developers and embedded developers to agree as well.

Vladimir Kliatchko and Ilougino Rocha offer additional arguments in favor of this functionality in N1851, complete with proposed wording. The suggested changes in this paper are a subset of those.

The key part of this proposal is that the allocator is not made part of shared_ptr's type. It is instead supplied as a parameter to the shared_ptr constructor in the same manner as the deleter is. This is made possible by the already existing deleter support infrastructure. Coupled with the fact that shared_ptr only allocates memory once in its constructor, and only deallocates it when the last instance in an ownership group is destroyed, this allows us to avoid all common sources of allocator problems. The allocator used for the creation of a particular shared_ptr instance remains purely an implementation detail for the user of that shared_ptr.

Why use the std::allocator interface if it's widely perceived as "broken"? The simple answer is that it works for our (and our users') purposes and is already in the Standard.

[Acknowledgments: this allocator interface has been independently suggested by Doug Gregor, Joe Gottman, Greg Colvin and others on the Boost list, as well as in N1851.]

Impact:

This feature extends the interface of shared_ptr in a backward-compatible way, allowing its broader use, and is therefore strongly recommended to be added to the C++0x standard. It does not impact existing uses, nor does it introduce binary compatibility issues.

Proposed text:

Add to shared_ptr [util.smartptr.shared] the following constructor:

template<class Y, class D, class A> shared_ptr( Y * p, D d, A a );

and the following member function:

template<class Y, class D, class A> void reset( Y * p, D d, A a );

Change the section:

template<class Y, class D> shared_ptr( Y * p, D d );

Requires: p shall be convertible to T*. D shall be CopyConstructible. The copy constructor and destructor of D shall not throw exceptions. The expression d(p) shall be well-formed, shall have well defined behavior, and shall not throw exceptions.

Effects: Constructs a shared_ptr object that owns the pointer p and the deleter d.

in [util.smartptr.shared.const] to:

template<class Y, class D> shared_ptr( Y * p, D d );
template<class Y, class D, class A> shared_ptr( Y * p, D d, A a );

Requires: p shall be convertible to T*. D shall be CopyConstructible. The copy constructor and destructor of D shall not throw. The expression d(p) shall be well-formed, shall have well defined behavior, and shall not throw. A shall be an allocator [allocator.requirements]. The copy constructor and destructor of A shall not throw.

Effects: Constructs a shared_ptr object that owns the pointer p and the deleter d. The second constructor shall use a copy of a to allocate memory for internal use.

Add the following to [util.smartptr.shared.mod]:

template<class Y, class D, class A> void reset( Y * p, D d, A a );

Effects: Equivalent to shared_ptr( p, d, a ).swap( *this ).

Implementability:

This feature has been added to boost::shared_ptr and will be part of Boost 1.35. See:

http://boost.cvs.sourceforge.net/*checkout*/boost/boost/boost/shared_ptr.hpp
http://boost.cvs.sourceforge.net/*checkout*/boost/boost/libs/smart_ptr/test/shared_ptr_alloc2_test.cpp

for reference.

III. Aliasing Support

Advanced users often require the ability to create a shared_ptr instance p that shares ownership with another (master) shared_ptr q but points to an object that is not a base of *q. *p may be a member or an element of *q, for example. This section proposes an additional constructor that can be used for this purpose.

An interesting side effect of this increase of expressive power is that now the *_pointer_cast functions can be implemented in user code. The make_shared factory function presented later in this document can also be implemented using only the public interface of shared_ptr via the aliasing constructor.

Impact:

This feature extends the interface of shared_ptr in a backward-compatible way that increases its expressive power and is therefore strongly recommended to be added to the C++0x standard. It introduces no source- and binary compatibility issues.

Proposed text:

Add to shared_ptr [util.smartptr.shared] the following constructor:

template<class Y> shared_ptr( shared_ptr<Y> const & r, T * p );

Add the following to [util.smartptr.shared.const]:

template<class Y> shared_ptr( shared_ptr<Y> const & r, T * p );

Effects: Constructs a shared_ptr instance that stores p and shares ownership with r.

Postconditions: get() == p && use_count() == r.use_count().

Throws: nothing.

[Note: To avoid the possibility of a dangling pointer, the user of this constructor must ensure that p remains valid at least until the ownership group of r is destroyed. --end note.]

[Note: This constructor allows creation of an empty shared_ptr instance with a non-NULL stored pointer. --end note.]

Implementability:

This feature has been added to boost::shared_ptr and will be part of Boost 1.35. See:

http://boost.cvs.sourceforge.net/*checkout*/boost/boost/boost/shared_ptr.hpp
http://boost.cvs.sourceforge.net/*checkout*/boost/boost/libs/smart_ptr/test/shared_ptr_alias_test.cpp

for reference.

IV. Object Creation

Consistent use of shared_ptr can eliminate the need to use an explicit delete, but it currently provides no support in avoiding explicit new. There have been repeated requests from users for a factory function that creates an object of a given type and returns a shared_ptr to it. Besides convenience and style, such a function is also exception safe and considerably faster because it can use a single allocation for the object and its corresponding control block, eliminating a significant portion of shared_ptr's construction overhead. This function eliminates one of the major efficiency complaints about shared_ptr.

This section proposes a family of overloaded function templates, make_shared<T> and allocate_shared<T>, to address this need. make_shared uses the global operator new to allocate memory, whereas allocate_shared uses an user-supplied allocator, allowing finer control consistent with section II of this document.

The rationale for choosing the name make_shared is that the expression make_shared<Widget>() can be read aloud and conveys the intended meaning. A free function also enables a non-intrusive implementation that can be delivered without modifications to an existing shared_ptr implementation.

Impact:

This feature does not affect the interface of shared_ptr. It is possible to implement in a non-intrusive way using only the public interface, as long as aliasing support is present. Access to implementation details can eliminate between 5 and 8 bytes of storage overhead on a typical 32 bit platform.

The addition is a strong candidate for the C++0x standard, but can be relegated to a technical report.

Proposed text:

Synopsis:

namespace std {
  template<class T, class... Args> shared_ptr<T> make_shared( Args && ... args );
  template<class T, class... Args> shared_ptr<T> allocate_shared( A const & a, Args && ... args );
}

Description:

template<class T, class... Args> shared_ptr<T> make_shared( Args && ... args );
template<class T, class... Args> shared_ptr<T> allocate_shared( A const & a, Args && ... args );

Requires: The expression new( pv ) T( std::forward<Args>(args)... ), where pv is a void* pointing to storage suitable to hold an object of type T, shall be well-formed. A shall be an allocator [allocator.requirements]. The copy constructor and destructor of A shall not throw.

Effects: Allocates memory suitable for an object of type T and constructs an object in it via the placement new expression new( pv ) T() or new( pv ) T( std::forward<Args>(args)... ). allocate_shared uses a copy of a to allocate memory. If an exception is thrown, has no effect.

Returns: A shared_ptr instance that stores and owns the address of the newly constructed object of type T.

Postconditions: get() != 0 && use_count() == 1.

Throws: bad_alloc, or an exception thrown from A::allocate or the constructor of T.

Remarks: Implementations are encouraged, but not required, to perform no more than one memory allocation. [Note: This provides efficiency equivalent to an intrusive smart pointer. --end note]

[Note: These functions will typically allocate more memory than sizeof(T) to allow for internal bookkeeping structures such as the reference counts. --end note]

Implementability:

A proof of concept non-intrusive implementation is available at:

http://www.pdimov.com/cpp/make_shared.cpp

This implementation uses variadic templates and rvalue references. When these features are not available, it falls back on a family of overloaded function templates taking arguments by const reference.

V. Move Support

Users often express concerns over the cost of copying a shared_ptr in situations where the source of the copy is no longer needed. To address this use case, N1851 proposes a separate smart pointer, managed_ptr, that is convertible from and to shared_ptr and enforces unique ownership.

The current proposal does not take this approach. Instead, we propose that move constructors and move assignment operators be added to shared_ptr. This allows a shared_ptr to be as efficient as an auto_ptr or the proposed unique_ptr when the source of the copy or assignment is a temporary or no longer needed. Move-aware standard containers will automatically take advantage of this optimization. As an example of the consequences, reallocating a vector< shared_ptr<T> > will no longer entail any reference count updates.

Impact:

This feature affects the interface of shared_ptr in a way that reduces its copy overhead and is in line with the rvalue recommendations for the standard library presented in N1859-N1862. We believe that it is a strong candidate for addition to the C++0x standard.

Proposed text:

Add to shared_ptr [util.smartptr.shared] the following:

shared_ptr( shared_ptr && r );
template<class Y> shared_ptr( shared_ptr<Y> && r );

shared_ptr& operator=( shared_ptr && r );
template<class Y> shared_ptr& operator=( shared_ptr<Y> && r );

Add the following to [util.smartptr.shared.const]:

shared_ptr( shared_ptr && r );
template<class Y> shared_ptr( shared_ptr<Y> && r );

Requires: For the second constructor Y* shall be convertible to T*.

Effects: Move-constructs a shared_ptr instance from r.

Postconditions: *this contains the old value of r. r is empty.

Throws: nothing.

Add the following to [util.smartptr.shared.assign]:

shared_ptr& operator=( shared_ptr && r );
template<class Y> shared_ptr& operator=( shared_ptr<Y> && r );

Effects: Equivalent to shared_ptr( move( r ) ).swap( *this ).

Returns: *this.

Implementability:

This feature has been added to boost::shared_ptr and will be part of Boost 1.35.

Thanks to Joe Gottman for his comments on the move support.

Thanks to Jens Maurer for his comments on the make_shared wording.

--end