Boost C++ Libraries

...one of the most highly regarded and expertly designed C++ library projects in the world. Herb Sutter and Andrei Alexandrescu, C++ Coding Standards

Boost Pointer Container Library

Tutorial

The tutorial shows you the most simple usage of the library. It is assumed that the reader is familiar with the use of standard containers. Although the tutorial is divided into sections, it is recommended that you read it all from top to bottom.

Basic usage

The most important aspect of a pointer container is that it manages memory for you. This means that you in most cases do not need to worry about deleting memory.

Let us assume that we have an OO-hierarchy of animals

class animal : boost::noncopyable
{
public:
    virtual      ~animal()   {}
    virtual void eat()       = 0;
    virtual int  age() const = 0;
    // ...
};

class mammal : public animal
{
    // ...
};

class bird : public animal
{
    // ...
};

Then the managing of the animals is straight-forward. Imagine a Zoo:

class zoo
{
    boost::ptr_vector<animal> the_animals;
public:

    void add_animal( animal* a )
    {
        the_animals.push_back( a );
    }
};

Notice how we just pass the class name to the container; there is no * to indicate it is a pointer. With this declaration we can now say:

zoo the_zoo;
the_zoo.add_animal( new mammal("joe") );
the_zoo.add_animal( new bird("dodo") );

Thus we heap-allocate all elements of the container and never rely on copy-semantics.

Indirected interface

A particular feature of the pointer containers is that the query interface is indirected. For example,

boost::ptr_vector<animal> vec;
vec.push_back( new animal ); // you add it as pointer ...
vec[0].eat();                // but get a reference back

This indirection also happens to iterators, so

typedef std::vector<animal*> std_vec;
std_vec vec;
...
std_vec::iterator i = vec.begin();
(*i)->eat(); // '*' needed

now becomes

typedef boost::ptr_vector<animal>  ptr_vec;
ptr_vec vec;
ptr_vec::iterator i = vec.begin();
i->eat(); // no indirection needed

Sequence containers

The sequence containers are used when you do not need to keep an ordering on your elements. You can basically expect all operations of the normal standard containers to be available. So, for example, with a ptr_deque and ptr_list object you can say:

boost::ptr_deque<animal> deq;
deq.push_front( new animal );
deq.pop_front();

because std::deque and std::list have push_front() and pop_front() members.

If the standard sequence supports random access, so does the pointer container; for example:

for( boost::ptr_deque<animal>::size_type i = 0u;
     i != deq.size(); ++i )
     deq[i].eat();

The ptr_vector also allows you to specify the size of the buffer to allocate; for example

boost::ptr_vector<animal> animals( 10u );

will reserve room for 10 animals.

Associative containers

To keep an ordering on our animals, we could use a ptr_set:

boost::ptr_set<animal> set;
set.insert( new monkey("bobo") );
set.insert( new whale("anna") );
...

This requires that operator<() is defined for animals. One way to do this could be

inline bool operator<( const animal& l, const animal& r )
{
    return l.name() < r.name();
}

if we wanted to keep the animals sorted by name.

Maybe you want to keep all the animals in zoo ordered wrt. their name, but it so happens that many animals have the same name. We can then use a ptr_multimap:

typedef boost::ptr_multimap<std::string,animal> zoo_type;
zoo_type zoo;
std::string bobo = "bobo",
            anna = "anna";
zoo.insert( bobo, new monkey(bobo) );
zoo.insert( bobo, new elephant(bobo) );
zoo.insert( anna, new whale(anna) );
zoo.insert( anna, new emu(anna) );

Note that must create the key as an lvalue (due to exception-safety issues); the following would not have compiled

zoo.insert( "bobo", // this is bad, but you get compile error
            new monkey("bobo") );

If a multimap is not needed, we can use operator[]() to avoid the clumsiness:

boost::ptr_map<std::string,animal> animals;
animals["bobo"].set_name("bobo");

This requires a default constructor for animals and a function to do the initialization, in this case set_name().

A better alternative is to use Boost.Assign to help you out. In particular, consider

For example, the above insertion may now be written

boost::ptr_multimap<std::string,animal> animals;

using namespace boost::assign;
ptr_map_insert<monkey>( animals )( "bobo", "bobo" );
ptr_map_insert<elephant>( animals )( "bobo", "bobo" );
ptr_map_insert<whale>( animals )( "anna", "anna" );
ptr_map_insert<emu>( animals )( "anna", "anna" );

Null values

By default, if you try to insert null into a container, an exception is thrown. If you want to allow nulls, then you must say so explicitly when declaring the container variable

boost::ptr_vector< boost::nullable<animal> > animals_type;
animals_type animals;
...
animals.insert( animals.end(), new dodo("fido") );
animals.insert( animals.begin(), 0 ) // ok

Once you have inserted a null into the container, you must always check if the value is null before accessing the object

for( animals_type::iterator i = animals.begin();
     i != animals.end(); ++i )
{
    if( !boost::is_null(i) ) // always check for validity
        i->eat();
}

If the container support random access, you may also check this as

for( animals_type::size_type i = 0u;
     i != animals.size(); ++i )
{
    if( !animals.is_null(i) )
         animals[i].eat();
}

Note that it is meaningless to insert null into ptr_set and ptr_multiset.

Cloneability

In OO programming it is typical to prohibit copying of objects; the objects may sometimes be allowed to be Cloneable; for example,:

animal* animal::clone() const
{
    return do_clone(); // implemented by private virtual function
}

If the OO hierarchy thus allows cloning, we need to tell the pointer containers how cloning is to be done. This is simply done by defining a free-standing function, new_clone(), in the same namespace as the object hierarchy:

inline animal* new_clone( const animal& a )
{
    return a.clone();
}

That is all, now a lot of functions in a pointer container can exploit the cloneability of the animal objects. For example

typedef boost::ptr_list<animal> zoo_type;
zoo_type zoo, another_zoo;
...
another_zoo.assign( zoo.begin(), zoo.end() );

will fill another zoo with clones of the first zoo. Similarly, insert() can now insert clones into your pointer container

another_zoo.insert( another_zoo.begin(), zoo.begin(), zoo.end() );

The whole container can now also be cloned

zoo_type yet_another_zoo = zoo.clone();

Copying or assigning the container has the same effect as cloning (though it is slightly cheaper):

zoo_type yet_another_zoo = zoo;

Copying also support derived-to-base class conversions:

boost::ptr_vector<monkey> monkeys = boost::assign::ptr_list_of<monkey>( "bobo" )( "bebe")( "uhuh" );
boost::ptr_vector<animal> animals = monkeys;

This also works for maps:

boost::ptr_map<std::string,monkey> monkeys = ...;
boost::ptr_map<std::string,animal> animals = monkeys;

New functions

Given that we know we are working with pointers, a few new functions make sense. For example, say you want to remove an animal from the zoo

zoo_type::auto_type the_animal = zoo.release( zoo.begin() );
the_animal->eat();
animal* the_animal_ptr = the_animal.release(); // now this is not deleted
zoo.release(2); // for random access containers

You can think of auto_type as a non-copyable form of std::auto_ptr. Notice that when you release an object, the pointer is removed from the container and the containers size shrinks. For containers that store nulls, we can exploit that auto_type is convertible to bool:

if( ptr_vector< nullable<T> >::auto_type r = vec.pop_back() )
{
  ...
}

You can also release the entire container if you want to return it from a function

compatible-smart-ptr< boost::ptr_deque<animal> > get_zoo()
{
    boost::ptr_deque<animal>  result;
    ...
    return result.release(); // give up ownership
}
...
boost::ptr_deque<animal> animals = get_zoo();

Let us assume we want to move an animal object from one zoo to another. In other words, we want to move the animal and the responsibility of it to another zoo

another_zoo.transfer( another_zoo.end(), // insert before end
                      zoo.begin(),       // insert this animal ...
                      zoo );             // from this container

This kind of "move-semantics" is different from normal value-based containers. You can think of transfer() as the same as splice() on std::list.

If you want to replace an element, you can easily do so

zoo_type::auto_type old_animal = zoo.replace( zoo.begin(), new monkey("bibi") );
zoo.replace( 2, old_animal.release() ); // for random access containers

A map is slightly different to iterate over than standard maps. Now we say

typedef boost::ptr_map<std::string, boost::nullable<animal> > animal_map;
animal_map map;
...
for( animal_map::const_iterator i = map.begin(), e = map.end(); i != e; ++i )
{
    std::cout << "\n key: " << i->first;
    std::cout << "\n age: ";

    if( boost::is_null(i) )
        std::cout << "unknown";
    else
        std::cout << i->second->age();
 }

Except for the check for null, this looks like it would with a normal map. But if age() had not been a const member function, it would not have compiled.

Maps can also be indexed with bounds-checking

try
{
    animal& bobo = map.at("bobo");
}
catch( boost::bad_ptr_container_operation& e )
{
    // "bobo" not found
}

Compatible smart pointer overloads

Every time there is a function that takes a T* parameter, there is also a function overload (or two) taking a compatible-smart-ptr<U> parameter. This is of course done to make the library intregrate seamlessly with std::auto_ptr or std::unique_ptr. For example, consider a statement like

std::ptr_vector<Base> vec;
vec.push_back( new Base );

If the compiler supports std::auto_ptr, this is complemented by

std::auto_ptr<Derived> p( new Derived );
vec.push_back( p );

Similarly if std::unique_ptr is available, we can write

std::unique_ptr<Derived> p( new Derived );
vec.push_back( std::move( p ) );

Notice that the template argument for compatible-smart-ptr does not need to follow the template argument for ptr_vector as long as Derived* can be implicitly converted to Base*.

Algorithms

Unfortunately it is not possible to use pointer containers with mutating algorithms from the standard library. However, the most useful ones are instead provided as member functions:

boost::ptr_vector<animal> zoo;
...
zoo.sort();                               // assume 'bool operator<( const animal&, const animal& )'
zoo.sort( std::less<animal>() );          // the same, notice no '*' is present
zoo.sort( zoo.begin(), zoo.begin() + 5 ); // sort selected range

Notice that predicates are automatically wrapped in an indirect_fun object.

You can remove equal and adjacent elements using unique():

zoo.unique();                             // assume 'bool operator==( const animal&, const animal& )'
zoo.unique( zoo.begin(), zoo.begin() + 5, my_comparison_predicate() );

If you just want to remove certain elements, use erase_if:

zoo.erase_if( my_predicate() );

Finally you may want to merge two sorted containers:

boost::ptr_vector<animal> another_zoo = ...;
another_zoo.sort();                      // sorted wrt. to same order as 'zoo'
zoo.merge( another_zoo );
BOOST_ASSERT( another_zoo.empty() );

That is all; now you have learned all the basics!


See also

Navigate


Copyright:Thorsten Ottosen 2004-2006. Use, modification and distribution is subject to the Boost Software License, Version 1.0 (see LICENSE_1_0.txt).