C++11 Smart Pointers

Introduction

Ooops. Yet another article on smart pointers of C++11. Nowadays I hear a lot of people talking about the new C++ standard which is nothing but C++0x/C++11. I went through some of the language features of C++11 and it's really an amazing work. I'll focus only on the smart pointers section of C++11.

Background

What are the issues with normal/raw/naked pointers?

Let's go one by one.

People refrain to use pointers as they give a lot of issues if not handled properly. That's why the newbie programmers dislike pointers. Many issues are involved with pointers like ensuring the lifetime of objects referred to by pointers, dangling references, and memory leaks.

Dangling reference is caused if a memory block is pointed by more than one pointer variable and if one of the pointers is released without letting know the other pointer. As all of you know, memory leaks occur when a block of memory is fetched from the heap and is not released back.

People say, I write clean and error proof code, why should I use smart pointers? And a programmer asked me, "Hey, here is my code. I fetched the memory from the heap, manipulated it, and after that I released it properly. What is the need of a smart pointer? "

Hide   Copy Code

void Foo( )
{
    int* iPtr = new int[5];
 
    //manipulate the memory block
    .
    .
    .
 
    delete[ ] iPtr;    
}

The above code works fine and memory is released properly under ideal circumstances. But think of the practical environment of code execution. The instructions between memory allocation and releasing can do nasty things like accessing an invalid memory location, dividing by zero, or say another programmer pitching into your program to fix a bug and adding a premature return statement based on some condition.

In all the above cases, you will never reach the point where the memory is released. This is because the first two cases throw an exception whereas the third one is a premature return. So the memory gets leaked while the program is running.

The one stop solution for all of the above issues is Smart Pointers [if they are really smart enough].

What is a smart pointer?

Smart pointer is a RAII modeled class to manage dynamically allocated memory. It provides all the interfaces provided by normal pointers with a few exceptions. During construction, it owns the memory and releases the same when it goes out of scope. In this way, the programmer is free about managing dynamically allocated memory.

C++98 has introduced the first of its kind called auto_ptr.

auto_ptr

Let's see the use of auto_ptr and how smart it is to resolve the above issues.

Hide   Copy Code

class Test
{
public:
 Test(int a = 0 ) : m_a(a)
 {
 }
 ~Test( )
 {
  cout<<"Calling destructor"<<endl;
 }
public:
 int m_a;
};
 
void main( )
{
 std::auto_ptr<Test> p( new Test(5) ); 
 cout<<p->m_a<<endl;
}

The above code is smart to release the memory associated with it. What we did is, we fetched a memory block to hold an object of type Test and associated it with auto_ptr p. So when p goes out of scope, the associated memory block is also released.

Hide   Shrink    Copy Code

//***************************************************************
class Test
{
public:
 Test(int a = 0 ) : m_a(a)
 {
 }
 ~Test( )
 {
  cout<<"Calling destructor"<<endl;
 }
public:
 int m_a;
};
//***************************************************************
void Fun( )
{
 int a = 0, b= 5, c;
 if( a ==0 )
 {
  throw "Invalid divisor";
 }
 c = b/a;
 return;
}
//***************************************************************
void main( )
{
 try
 {
  std::auto_ptr<Test> p( new Test(5) ); 
  Fun( );
  cout<<p->m_a<<endl;
 }
 catch(...)
 {
  cout<<"Something has gone wrong"<<endl;
 }
}

In the above case, an exception is thrown but still the pointer is released properly. This is because of stack unwinding which happens when an exception is thrown. As all local objects belonging to the try block are destroyed, p goes out of scope and it releases the associated memory.

Issue 1: So far auto_ptr is smart. But it has more fundamental flaws over its smartness. auto_ptr transfers the ownership when it is assigned to another auto_ptr. This is really an issue while passing the auto_ptr between the functions. Say, I have an auto_ptr in Foo( ) and this pointer is passed another function say Fun( ) fromFoo. Now once Fun( ) completes its execution, the ownership is not returned back to Foo.

Hide   Copy Code

//***************************************************************
class Test
{
public:
 Test(int a = 0 ) : m_a(a)
 {
 }
 ~Test( )
 {
  cout<<"Calling destructor"<<endl;
 }
public:
 int m_a;
};
 
 
//***************************************************************
void Fun(auto_ptr<Test> p1 )
{
 cout<<p1->m_a<<endl;
}
//***************************************************************
void main( )
{
 std::auto_ptr<Test> p( new Test(5) ); 
 Fun(p);
 cout<<p->m_a<<endl;
} 

The above code causes a program crash because of the weird behavior of auto_ptr. What happens is that, powns a memory block and when Fun is called, p transfers the ownership of its associated memory block to theauto_ptr p1 which is the copy of p. Now p1 owns the memory block which was previously owned by p. So far it is fine. Now fun has completed its execution, and p1 goes out of scope and the memory blocked is released. How about p? p does not own anything, that is why it causes a crash when the next line is executed which accesses p thinking that it owns some resource.

Issue 2: Yet another flaw. auto_ptr cannot be used with an array of objects. I mean it cannot be used with the operator new[].

Hide   Copy Code

//***************************************************************
void main( )
{
 std::auto_ptr<Test> p(new Test[5]);
}

The above code gives a runtime error. This is because when auto_ptr goes out of scope, delete is called on the associated memory block. This is fine if auto_ptr owns only a single object. But in the above code, we have created an array of objects on the heap which should be destroyed using delete[ ] and not delete.

Issue 3: auto_ptr cannot be used with standard containers like vector, list, map, etc.

As auto_ptr is more error prone and it will be deprecated, C++ 11 has come with a new set of smart pointers, each has its own purpose.

• shared_ptr
• unique_ptr
• weak_ptr

shared_ptr

OK, get ready to enjoy the real smartness. The first of its kind is shared_ptr which has the notion called shared ownership. The goal of shared_ptr is very simple: Multiple shared pointers can refer to a single object and when the last shared pointer goes out of scope, memory is released automatically.

Creation:

Hide   Copy Code

void main( )
{
 shared_ptr<int> sptr1( new int );
}

Make use of the make_shared macro which expedites the creation process. As shared_ptr allocates memory internally, to hold the reference count, make_shared( ) is implemented in a way to do this job effectively.

Hide   Copy Code

void main( )
{
 shared_ptr<int> sptr1 = make_shared<int>(100);
}

The above code creates a shared_ptr which points to a memory block to hold an integer with value 100 and reference count 1. If another shared pointer is created out of sptr1, the reference count goes up to 2. This count is known as strong reference. Apart from this, the shared pointer has another reference count known as weak reference, which will be explained while visiting weak pointers.

You can find out the number of shared_ptrs referring to the resource by just getting the reference count by calling use_count( ). And while debugging, you can get it by watching the stong_ref of the shared_ptr.

Destruction:

shared_ptr releases the associated resource by calling delete by default. If the user needs a different destruction policy, he/she is free to specify the same while constructing the shared_ptr. The following code is a source of trouble due to the default destruction policy:

Hide   Copy Code

class Test
{
public:
 Test(int a = 0 ) : m_a(a)
 {
 }
 ~Test( )
 {
  cout<<"Calling destructor"<<endl;
 }
public:
         int m_a;
};
void main( )
{
 shared_ptr<Test> sptr1( new Test[5] );
}

In this scenario shared_ptr owns array of objects and the default destruction policy calls "delete" to release the associated memory when it goes out of scope. Actually, delete[ ] should have been called to destroy the array. The user can specify the custom deallocator by a callable object, i.e., a function, lambda expression, function object.

Hide   Copy Code

void main( )
{
 shared_ptr<Test> sptr1( new Test[5], 
        [ ](Test* p) { delete[ ] p; } );
}

The above code works fine as we have specified the destruction should happen via delete[].

Interface

shared_ptr provides dereferencing operators *, -> like a normal pointer provides. Apart from that it provides some more important interfaces like:

• get( ) : To get the resource associated with the shared_ptr.
• reset( ) : To yield the ownership of the associated memory block. If this is the last shared_ptr owning the resource, then the resource is released automatically.
• unique: To know whether the resource is managed by only this shared_ptr instance.
• operator bool: To check whether the shared_ptr owns a memory block or not. Can be used with anif condition.

OK, that is all about shared_ptrs. But shared_ptrs too have a few issues:.

Issues:

Hide   Copy Code

void main( )
{
 shared_ptr<int> sptr1( new int );
 shared_ptr<int> sptr2 = sptr1;
 shared_ptr<int> sptr3;
 sptr3 = sptr2;
}

The below table gives you the reference count values for the above code.

All shared_ptrs share the same reference count hence belonging to the same group. The above code is fine. Let's see another piece of code.

Hide   Copy Code

void main( )
{
 int* p = new int;
 shared_ptr<int> sptr1( p);
 shared_ptr<int> sptr2( p );
}

The above piece of code is going to cause an error because two shared_ptrs from different groups share a single resource. The below table gives you a picture of the root cause.

To avoid this, better not create the shared pointers from a naked pointer.

Hide   Shrink    Copy Code

class B;
class A
{
public:
 A(  ) : m_sptrB(nullptr) { };
 ~A( )
 {
  cout<<" A is destroyed"<<endl;
 }
 shared_ptr<B> m_sptrB;
};
class B
{
public:
 B(  ) : m_sptrA(nullptr) { };
 ~B( )
 {
  cout<<" B is destroyed"<<endl;
 }
 shared_ptr<A> m_sptrA;
};
//***********************************************************
void main( )
{
 shared_ptr<B> sptrB( new B );
 shared_ptr<A> sptrA( new A );
 sptrB->m_sptrA = sptrA;
 sptrA->m_sptrB = sptrB;
}

The above code has cyclic reference. I mean class A holds a shared pointer to B and class B holds a shared pointer to A. In this case, the resource associated with both sptrA and sptrB are not released. Refer to the below table.

Reference counts for both sptrA and sptrB go down to 1 when they go out of scope and hence the resources are not released!!!!!

1. If a memory is block is associated with shared_ptrs belonging to a different group, then there is an error. All shared_ptrs sharing the same reference count belong to a group. Let's see an example.
2. There is another issue involved with creating a shared pointer from a naked pointer. In the above code, consider that only one shared pointer is created using p and the code works fine. Consider by mistake if a programmer deletes the naked pointer p before the scope of the shared pointer ends. Oooppss!!! Yet another crash..
3. Cyclic Reference: Resources are not released properly if a cyclic reference of shared pointers are involved. Consider the following piece of code.

To resolve the cyclic reference, C++ provides another smart pointer class called weak_ptr.

Weak_Ptr

A weak pointer provides sharing semantics and not owning semantics. This means a weak pointer can share a resource held by a shared_ptr. So to create a weak pointer, some body should already own the resource which is nothing but a shared pointer.

A weak pointer does not allow normal interfaces supported by a pointer, like calling *, ->. Because it is not the owner of the resource and hence it does not give any chance for the programmer to mishandle it. Then how do we make use of a weak pointer?

The answer is to create a shared_ptr out of a weak _ptr and use it. Because this makes sure that the resource won't be destroyed while using by incrementing the strong reference count. As the reference count is incremented, it is sure that the count will be at least 1 till you complete using the shared_ptr created out of theweak_ptr. Otherwise what may happen is while using the weak_ptr, the resource held by the shared_ptrgoes out of scope and the memory is released which creates chaos.

Creation

A weak pointer constructor takes a shared pointer as one of its parameters. Creating a weak pointer out of a shared pointer increases the weak reference counter of the shared pointer. This means that the shared pointer shares it resource with another pointer. But this counter is not considered to release the resource when the shared pointer goes out of scope. I mean if the strong reference of the shared pointer goes to 0, then the resource is released irrespective of the weak reference value.

Hide   Copy Code

void main( )
{
 shared_ptr<Test> sptr( new Test );
 weak_ptr<Test> wptr( sptr );
 weak_ptr<Test> wptr1 = wptr;
}

We can watch the reference counters of the shared/weak pointer.

Assigning a weak pointer to another weak pointer increases the weak reference count.

So what happens when a weak pointer points to a resource held by the shared pointer and the shared pointer destroys the associated resource when it goes out of scope? The weak pointer gets expired.

How to check whether the weak pointer is pointing to a valid resource? There are two ways:

1. Call the use_count( ) method to know the count. Note that this method returns the strong reference count and not the weak reference.
2. Call the expired( ) method. This is faster than calling use_count( ).

To get a shared_ptr from a weak_ptr call lock( ) or directly casting the weak_ptr to shared_ptr.

Hide   Copy Code

void main( )
{
 shared_ptr<Test> sptr( new Test );
 weak_ptr<Test> wptr( sptr );
 shared_ptr<Test> sptr2 = wptr.lock( );
}

Getting the shared_ptr from the weak_ptr increases the strong reference as said earlier.

Now let's see how the cyclic reference issue is resolved using the weak_ptr.

Hide   Shrink    Copy Code

class B;
class A
{
public:
 A(  ) : m_a(5)  { };
 ~A( )
 {
  cout<<" A is destroyed"<<endl;
 }
 void PrintSpB( );
 weak_ptr<B> m_sptrB;
 int m_a;
};
class B
{
public:
 B(  ) : m_b(10) { };
 ~B( )
 {
  cout<<" B is destroyed"<<endl;
 }
 weak_ptr<A> m_sptrA;
 int m_b;
};

void A::PrintSpB( )
{
 if( !m_sptrB.expired() )
 {  
  cout<< m_sptrB.lock( )->m_b<<endl;
 }
}

void main( )
{
 shared_ptr<B> sptrB( new B );
 shared_ptr<A> sptrA( new A );
 sptrB->m_sptrA = sptrA;
 sptrA->m_sptrB = sptrB;
 sptrA->PrintSpB( ); 
}

Unique_ptr

This is almost a kind of replacement to the error prone auto_ptr. unique_ptr follows the exclusive ownership semantics, i.e., at any point of time, the resource is owned by only one unique_ptr. When unique_ptr goes out of scope, the resource is released. If the resource is overwritten by some other resource, the previously owned resource is released. So it guarantees that the associated resource is released always.

Creation

unique_ptr is created in the same way as shared_ptr except it has an additional facility for an array of objects.

Hide   Copy Code

unique_ptr<int> uptr( new int );

The unique_ptr class provides the specialization to create an array of objects which calls delete[ ] instead ofdelete when the pointer goes out of scope. The array of objects can be specified as a part of the template parameter while creating the unique_ptr. In this way, the programmer does not have to provide a custom deallocator, as unique_ptr does it.

Hide   Copy Code

unique_ptr<int[ ]> uptr( new int[5] );

 

 

Ownership of the resource can be transferred from one unique_ptr to another by assigning it.

Keep in mind that unique_ptr does not provide you copy semantics [copy assignment and copy construction is not possible] but move semantics.

In the above case, if upt3 and uptr5 owns some resource already, then it will be destroyed properly before owning a new resource.

Interface

The interface that unique_ptr provides is very similar to the ordinary pointer but no pointer arithmetic is allowed.

unique_ptr provides a function called release which yields the ownership. The difference betweenrelease() and reset( ), is release just yields the ownership and does not destroy the resource whereasreset destroys the resource.

Which one to use?

It purely depends upon how you want to own a resource. If shared ownership is needed then go forshared_ptr, otherwise unique_ptr.

Apart from that, shared_ptr is a bit heavier than unique_ptr because internally it allocates memory to do a lot of book keeping like strong reference, weak reference, etc. But unique_ptr does not need these counters as it is the only owner for the resource.