Memory

Memory contains three general areas. First, function and operator calls via new and delete operator or member function calls. Second, allocation via allocator. And finally, smart pointer and intelligent pointer abstractions.

Allocators

Memory management for Standard Library entities is encapsulated in a class template called allocator. The allocator abstraction is used throughout the library in string, container classes, algorithms, and parts of iostreams. This class, and base classes of it, are the superset of available free store (heap) management classes.

Requirements

The C++ standard only gives a few directives in this area:

  • When you add elements to a container, and the container must allocate more memory to hold them, the container makes the request via its Allocator template parameter, which is usually aliased to allocator_type. This includes adding chars to the string class, which acts as a regular STL container in this respect.

  • The default Allocator argument of every container-of-T is allocator<T>.

  • The interface of the allocator<T> class is extremely simple. It has about 20 public declarations (nested typedefs, member functions, etc), but the two which concern us most are:

    	 T*    allocate   (size_type n, const void* hint = 0);
    	 void  deallocate (T* p, size_type n);
           

    The n arguments in both those functions is a count of the number of T's to allocate space for, not their total size. (This is a simplification; the real signatures use nested typedefs.)

  • The storage is obtained by calling ::operator new, but it is unspecified when or how often this function is called. The use of the hint is unspecified, but intended as an aid to locality if an implementation so desires. [20.4.1.1]/6

Complete details can be found in the C++ standard, look in [20.4 Memory].

Design Issues

The easiest way of fulfilling the requirements is to call operator new each time a container needs memory, and to call operator delete each time the container releases memory. This method may be slower than caching the allocations and re-using previously-allocated memory, but has the advantage of working correctly across a wide variety of hardware and operating systems, including large clusters. The __gnu_cxx::new_allocator implements the simple operator new and operator delete semantics, while __gnu_cxx::malloc_allocator implements much the same thing, only with the C language functions std::malloc and std::free.

Another approach is to use intelligence within the allocator class to cache allocations. This extra machinery can take a variety of forms: a bitmap index, an index into an exponentially increasing power-of-two-sized buckets, or simpler fixed-size pooling cache. The cache is shared among all the containers in the program: when your program's std::vector<int> gets cut in half and frees a bunch of its storage, that memory can be reused by the private std::list<WonkyWidget> brought in from a KDE library that you linked against. And operators new and delete are not always called to pass the memory on, either, which is a speed bonus. Examples of allocators that use these techniques are __gnu_cxx::bitmap_allocator, __gnu_cxx::pool_allocator, and __gnu_cxx::__mt_alloc.

Depending on the implementation techniques used, the underlying operating system, and compilation environment, scaling caching allocators can be tricky. In particular, order-of-destruction and order-of-creation for memory pools may be difficult to pin down with certainty, which may create problems when used with plugins or loading and unloading shared objects in memory. As such, using caching allocators on systems that do not support abi::__cxa_atexit is not recommended.

Implementation

Interface Design

The only allocator interface that is supported is the standard C++ interface. As such, all STL containers have been adjusted, and all external allocators have been modified to support this change.

The class allocator just has typedef, constructor, and rebind members. It inherits from one of the high-speed extension allocators, covered below. Thus, all allocation and deallocation depends on the base class.

The base class that allocator is derived from may not be user-configurable.

Selecting Default Allocation Policy

It's difficult to pick an allocation strategy that will provide maximum utility, without excessively penalizing some behavior. In fact, it's difficult just deciding which typical actions to measure for speed.

Three synthetic benchmarks have been created that provide data that is used to compare different C++ allocators. These tests are:

  1. Insertion.

    Over multiple iterations, various STL container objects have elements inserted to some maximum amount. A variety of allocators are tested. Test source for sequence and associative containers.

  2. Insertion and erasure in a multi-threaded environment.

    This test shows the ability of the allocator to reclaim memory on a per-thread basis, as well as measuring thread contention for memory resources. Test source here.

  3. A threaded producer/consumer model.

    Test source for sequence and associative containers.

The current default choice for allocator is __gnu_cxx::new_allocator.

Disabling Memory Caching

In use, allocator may allocate and deallocate using implementation-specific strategies and heuristics. Because of this, a given call to an allocator object's allocate member function may not actually call the global operator new and a given call to to the deallocate member function may not call operator delete.

This can be confusing.

In particular, this can make debugging memory errors more difficult, especially when using third-party tools like valgrind or debug versions of new.

There are various ways to solve this problem. One would be to use a custom allocator that just called operators new and delete directly, for every allocation. (See the default allocator, include/ext/new_allocator.h, for instance.) However, that option may involve changing source code to use a non-default allocator. Another option is to force the default allocator to remove caching and pools, and to directly allocate with every call of allocate and directly deallocate with every call of deallocate, regardless of efficiency. As it turns out, this last option is also available.

To globally disable memory caching within the library for some of the optional non-default allocators, merely set GLIBCXX_FORCE_NEW (with any value) in the system's environment before running the program. If your program crashes with GLIBCXX_FORCE_NEW in the environment, it likely means that you linked against objects built against the older library (objects which might still using the cached allocations...).

Using a Specific Allocator

You can specify different memory management schemes on a per-container basis, by overriding the default Allocator template parameter. For example, an easy (but non-portable) method of specifying that only malloc or free should be used instead of the default node allocator is:

    std::list <int, __gnu_cxx::malloc_allocator<int> >  malloc_list;

Likewise, a debugging form of whichever allocator is currently in use:

    std::deque <int, __gnu_cxx::debug_allocator<std::allocator<int> > >  debug_deque;
      

Custom Allocators

Writing a portable C++ allocator would dictate that the interface would look much like the one specified for allocator. Additional member functions, but not subtractions, would be permissible.

Probably the best place to start would be to copy one of the extension allocators: say a simple one like new_allocator.

Extension Allocators

Several other allocators are provided as part of this implementation. The location of the extension allocators and their names have changed, but in all cases, functionality is equivalent. Starting with gcc-3.4, all extension allocators are standard style. Before this point, SGI style was the norm. Because of this, the number of template arguments also changed. Here's a simple chart to track the changes.

More details on each of these extension allocators follows.

  1. new_allocator

    Simply wraps ::operator new and ::operator delete.

  2. malloc_allocator

    Simply wraps malloc and free. There is also a hook for an out-of-memory handler (for new/delete this is taken care of elsewhere).

  3. array_allocator

    Allows allocations of known and fixed sizes using existing global or external storage allocated via construction of std::tr1::array objects. By using this allocator, fixed size containers (including std::string) can be used without instances calling ::operator new and ::operator delete. This capability allows the use of STL abstractions without runtime complications or overhead, even in situations such as program startup. For usage examples, please consult the testsuite.

  4. debug_allocator

    A wrapper around an arbitrary allocator A. It passes on slightly increased size requests to A, and uses the extra memory to store size information. When a pointer is passed to deallocate(), the stored size is checked, and assert() is used to guarantee they match.

  5. throw_allocator

    Includes memory tracking and marking abilities as well as hooks for throwing exceptions at configurable intervals (including random, all, none).

  6. __pool_alloc

    A high-performance, single pool allocator. The reusable memory is shared among identical instantiations of this type. It calls through ::operator new to obtain new memory when its lists run out. If a client container requests a block larger than a certain threshold size, then the pool is bypassed, and the allocate/deallocate request is passed to ::operator new directly.

    Older versions of this class take a boolean template parameter, called thr, and an integer template parameter, called inst.

    The inst number is used to track additional memory pools. The point of the number is to allow multiple instantiations of the classes without changing the semantics at all. All three of

        typedef  __pool_alloc<true,0>    normal;
        typedef  __pool_alloc<true,1>    private;
        typedef  __pool_alloc<true,42>   also_private;
       

    behave exactly the same way. However, the memory pool for each type (and remember that different instantiations result in different types) remains separate.

    The library uses 0 in all its instantiations. If you wish to keep separate free lists for a particular purpose, use a different number.

    The thr boolean determines whether the pool should be manipulated atomically or not. When thr = true, the allocator is thread-safe, while thr = false, is slightly faster but unsafe for multiple threads.

    For thread-enabled configurations, the pool is locked with a single big lock. In some situations, this implementation detail may result in severe performance degradation.

    (Note that the GCC thread abstraction layer allows us to provide safe zero-overhead stubs for the threading routines, if threads were disabled at configuration time.)

  7. __mt_alloc

    A high-performance fixed-size allocator with exponentially-increasing allocations. It has its own chapter in the documentation.

  8. bitmap_allocator

    A high-performance allocator that uses a bit-map to keep track of the used and unused memory locations. It has its own chapter in the documentation.

Bibliography

ISO/IEC 14882:1998 Programming languages - C++ . isoc++_1998 20.4 Memory.

The Standard Librarian: What Are Allocators Good For? . Matt Austern. C/C++ Users Journal .

Reconsidering Custom Memory Allocation . Emery Berger. Ben Zorn. Kathryn McKinley. Copyright © 2002 OOPSLA.

Allocator Types . Klaus Kreft. Angelika Langer. C/C++ Users Journal .

The C++ Programming Language. Bjarne Stroustrup. Copyright © 2000 . 19.4 Allocators. Addison Wesley .

Yalloc: A Recycling C++ Allocator. Felix Yen.

auto_ptr

Limitations

Explaining all of the fun and delicious things that can happen with misuse of the auto_ptr class template (called AP here) would take some time. Suffice it to say that the use of AP safely in the presence of copying has some subtleties.

The AP class is a really nifty idea for a smart pointer, but it is one of the dumbest of all the smart pointers -- and that's fine.

AP is not meant to be a supersmart solution to all resource leaks everywhere. Neither is it meant to be an effective form of garbage collection (although it can help, a little bit). And it can notbe used for arrays!

AP is meant to prevent nasty leaks in the presence of exceptions. That's all. This code is AP-friendly:

    // Not a recommend naming scheme, but good for web-based FAQs.
    typedef std::auto_ptr<MyClass>  APMC;

    extern function_taking_MyClass_pointer (MyClass*);
    extern some_throwable_function ();

    void func (int data)
    {
	APMC  ap (new MyClass(data));

	some_throwable_function();   // this will throw an exception

	function_taking_MyClass_pointer (ap.get());
    }
   

When an exception gets thrown, the instance of MyClass that's been created on the heap will be delete'd as the stack is unwound past func().

Changing that code as follows is not AP-friendly:

	APMC  ap (new MyClass[22]);
   

You will get the same problems as you would without the use of AP:

	char*  array = new char[10];       // array new...
	...
	delete array;                      // ...but single-object delete
   

AP cannot tell whether the pointer you've passed at creation points to one or many things. If it points to many things, you are about to die. AP is trivial to write, however, so you could write your own auto_array_ptr for that situation (in fact, this has been done many times; check the mailing lists, Usenet, Boost, etc).

Use in Containers

All of the containers described in the standard library require their contained types to have, among other things, a copy constructor like this:

    struct My_Type
    {
	My_Type (My_Type const&);
    };
   

Note the const keyword; the object being copied shouldn't change. The template class auto_ptr (called AP here) does not meet this requirement. Creating a new AP by copying an existing one transfers ownership of the pointed-to object, which means that the AP being copied must change, which in turn means that the copy ctors of AP do not take const objects.

The resulting rule is simple: Never ever use a container of auto_ptr objects. The standard says that undefined behavior is the result, but it is guaranteed to be messy.

To prevent you from doing this to yourself, the concept checks built in to this implementation will issue an error if you try to compile code like this:

    #include <vector>
    #include <memory>

    void f()
    {
	std::vector< std::auto_ptr<int> >   vec_ap_int;
    }
   

Should you try this with the checks enabled, you will see an error.

shared_ptr

The shared_ptr class template stores a pointer, usually obtained via new, and implements shared ownership semantics.

Requirements

The standard deliberately doesn't require a reference-counted implementation, allowing other techniques such as a circular-linked-list.

Design Issues

The shared_ptr code is kindly donated to GCC by the Boost project and the original authors of the code. The basic design and algorithms are from Boost, the notes below describe details specific to the GCC implementation. Names have been uglified in this implementation, but the design should be recognisable to anyone familiar with the Boost 1.32 shared_ptr.

The basic design is an abstract base class, _Sp_counted_base that does the reference-counting and calls virtual functions when the count drops to zero. Derived classes override those functions to destroy resources in a context where the correct dynamic type is known. This is an application of the technique known as type erasure.

Implementation

Class Hierarchy

A shared_ptr<T> contains a pointer of type T* and an object of type __shared_count. The shared_count contains a pointer of type _Sp_counted_base* which points to the object that maintains the reference-counts and destroys the managed resource.

_Sp_counted_base<Lp>

The base of the hierarchy is parameterized on the lock policy (see below.) _Sp_counted_base doesn't depend on the type of pointer being managed, it only maintains the reference counts and calls virtual functions when the counts drop to zero. The managed object is destroyed when the last strong reference is dropped, but the _Sp_counted_base itself must exist until the last weak reference is dropped.

_Sp_counted_base_impl<Ptr, Deleter, Lp>

Inherits from _Sp_counted_base and stores a pointer of type Ptr and a deleter of type Deleter. _Sp_deleter is used when the user doesn't supply a custom deleter. Unlike Boost's, this default deleter is not "checked" because GCC already issues a warning if delete is used with an incomplete type. This is the only derived type used by tr1::shared_ptr<Ptr> and it is never used by std::shared_ptr, which uses one of the following types, depending on how the shared_ptr is constructed.

_Sp_counted_ptr<Ptr, Lp>

Inherits from _Sp_counted_base and stores a pointer of type Ptr, which is passed to delete when the last reference is dropped. This is the simplest form and is used when there is no custom deleter or allocator.

_Sp_counted_deleter<Ptr, Deleter, Alloc>

Inherits from _Sp_counted_ptr and adds support for custom deleter and allocator. Empty Base Optimization is used for the allocator. This class is used even when the user only provides a custom deleter, in which case allocator is used as the allocator.

_Sp_counted_ptr_inplace<Tp, Alloc, Lp>

Used by allocate_shared and make_shared. Contains aligned storage to hold an object of type Tp, which is constructed in-place with placement new. Has a variadic template constructor allowing any number of arguments to be forwarded to Tp's constructor. Unlike the other _Sp_counted_* classes, this one is parameterized on the type of object, not the type of pointer; this is purely a convenience that simplifies the implementation slightly.

C++11-only features are: rvalue-ref/move support, allocator support, aliasing constructor, make_shared & allocate_shared. Additionally, the constructors taking auto_ptr parameters are deprecated in C++11 mode.

Thread Safety

The Thread Safety section of the Boost shared_ptr documentation says "shared_ptr objects offer the same level of thread safety as built-in types." The implementation must ensure that concurrent updates to separate shared_ptr instances are correct even when those instances share a reference count e.g.

shared_ptr<A> a(new A);
shared_ptr<A> b(a);

// Thread 1     // Thread 2
   a.reset();      b.reset();

The dynamically-allocated object must be destroyed by exactly one of the threads. Weak references make things even more interesting. The shared state used to implement shared_ptr must be transparent to the user and invariants must be preserved at all times. The key pieces of shared state are the strong and weak reference counts. Updates to these need to be atomic and visible to all threads to ensure correct cleanup of the managed resource (which is, after all, shared_ptr's job!) On multi-processor systems memory synchronisation may be needed so that reference-count updates and the destruction of the managed resource are race-free.

The function _Sp_counted_base::_M_add_ref_lock(), called when obtaining a shared_ptr from a weak_ptr, has to test if the managed resource still exists and either increment the reference count or throw bad_weak_ptr. In a multi-threaded program there is a potential race condition if the last reference is dropped (and the managed resource destroyed) between testing the reference count and incrementing it, which could result in a shared_ptr pointing to invalid memory.

The Boost shared_ptr (as used in GCC) features a clever lock-free algorithm to avoid the race condition, but this relies on the processor supporting an atomic Compare-And-Swap instruction. For other platforms there are fall-backs using mutex locks. Boost (as of version 1.35) includes several different implementations and the preprocessor selects one based on the compiler, standard library, platform etc. For the version of shared_ptr in libstdc++ the compiler and library are fixed, which makes things much simpler: we have an atomic CAS or we don't, see Lock Policy below for details.

Selecting Lock Policy

There is a single _Sp_counted_base class, which is a template parameterized on the enum __gnu_cxx::_Lock_policy. The entire family of classes is parameterized on the lock policy, right up to __shared_ptr, __weak_ptr and __enable_shared_from_this. The actual std::shared_ptr class inherits from __shared_ptr with the lock policy parameter selected automatically based on the thread model and platform that libstdc++ is configured for, so that the best available template specialization will be used. This design is necessary because it would not be conforming for shared_ptr to have an extra template parameter, even if it had a default value. The available policies are:

  1. _S_Atomic

    Selected when GCC supports a builtin atomic compare-and-swap operation on the target processor (see Atomic Builtins.) The reference counts are maintained using a lock-free algorithm and GCC's atomic builtins, which provide the required memory synchronisation.

  2. _S_Mutex

    The _Sp_counted_base specialization for this policy contains a mutex, which is locked in add_ref_lock(). This policy is used when GCC's atomic builtins aren't available so explicit memory barriers are needed in places.

  3. _S_Single

    This policy uses a non-reentrant add_ref_lock() with no locking. It is used when libstdc++ is built without --enable-threads.

For all three policies, reference count increments and decrements are done via the functions in ext/atomicity.h, which detect if the program is multi-threaded. If only one thread of execution exists in the program then less expensive non-atomic operations are used.

Related functions and classes
dynamic_pointer_cast, static_pointer_cast, const_pointer_cast

As noted in N2351, these functions can be implemented non-intrusively using the alias constructor. However the aliasing constructor is only available in C++11 mode, so in TR1 mode these casts rely on three non-standard constructors in shared_ptr and __shared_ptr. In C++11 mode these constructors and the related tag types are not needed.

enable_shared_from_this

The clever overload to detect a base class of type enable_shared_from_this comes straight from Boost. There is an extra overload for __enable_shared_from_this to work smoothly with __shared_ptr<Tp, Lp> using any lock policy.

make_shared, allocate_shared

make_shared simply forwards to allocate_shared with std::allocator as the allocator. Although these functions can be implemented non-intrusively using the alias constructor, if they have access to the implementation then it is possible to save storage and reduce the number of heap allocations. The newly constructed object and the _Sp_counted_* can be allocated in a single block and the standard says implementations are "encouraged, but not required," to do so. This implementation provides additional non-standard constructors (selected with the type _Sp_make_shared_tag) which create an object of type _Sp_counted_ptr_inplace to hold the new object. The returned shared_ptr<A> needs to know the address of the new A object embedded in the _Sp_counted_ptr_inplace, but it has no way to access it. This implementation uses a "covert channel" to return the address of the embedded object when get_deleter<_Sp_make_shared_tag>() is called. Users should not try to use this. As well as the extra constructors, this implementation also needs some members of _Sp_counted_deleter to be protected where they could otherwise be private.

Use

Examples

Examples of use can be found in the testsuite, under testsuite/tr1/2_general_utilities/shared_ptr, testsuite/20_util/shared_ptr and testsuite/20_util/weak_ptr.

Unresolved Issues

The shared_ptr atomic access clause in the C++11 standard is not implemented in GCC.

The _S_single policy uses atomics when used in MT code, because it uses the same dispatcher functions that check __gthread_active_p(). This could be addressed by providing template specialisations for some members of _Sp_counted_base<_S_single>.

Unlike Boost, this implementation does not use separate classes for the pointer+deleter and pointer+deleter+allocator cases in C++11 mode, combining both into _Sp_counted_deleter and using allocator when the user doesn't specify an allocator. If it was found to be beneficial an additional class could easily be added. With the current implementation, the _Sp_counted_deleter and __shared_count constructors taking a custom deleter but no allocator are technically redundant and could be removed, changing callers to always specify an allocator. If a separate pointer+deleter class was added the __shared_count constructor would be needed, so it has been kept for now.

The hack used to get the address of the managed object from _Sp_counted_ptr_inplace::_M_get_deleter() is accessible to users. This could be prevented if get_deleter<_Sp_make_shared_tag>() always returned NULL, since the hack only needs to work at a lower level, not in the public API. This wouldn't be difficult, but hasn't been done since there is no danger of accidental misuse: users already know they are relying on unsupported features if they refer to implementation details such as _Sp_make_shared_tag.

tr1::_Sp_deleter could be a private member of tr1::__shared_count but it would alter the ABI.

Acknowledgments

The original authors of the Boost shared_ptr, which is really nice code to work with, Peter Dimov in particular for his help and invaluable advice on thread safety. Phillip Jordan and Paolo Carlini for the lock policy implementation.