From: Christian Engström (christian.engstrom_at_[hidden])
Date: 2003-04-24 09:57:46
HEATH DAVIS wrote:
> I have experienced some difficulty with sorting lists of type:
>
> list < shared_ptr < foo > >
>
For a while I have been toying with the idea of a smart pointer called
sortable_ptr, which would behave in the way that [I believe that] the
original poster erroneously presumed that boost::shared_ptr already did.
I was originally planning to do some more practical testing of the
concept before possibly submitting it as a proposal to Boost, but since
the topic has now been raised, I will take the opportunity to get some
feedback as to whether the idea is sound, or if I have missed something
crucial.
The following is an extract from the draft documentation I have been
writing.
- - - -
The class sortable_ptr is publicly derived from boost::shared_ptr, but
defines the comparison operator < (and its siblings >, <=, and >= ) to
make a comparison of the actual objects. This means that if the class
myclass has operator< defined, we can write
sortable_ptr<myclass> aptr(new myclass(/* Something */));
sortable_ptr<myclass> bptr(new myclass(/* Something else*/));
assert((aptr < bptr) == (*aptr < *bptr));
This is different from boost::shared_ptr, which does only define the <
operator to do a direct comparison of the pointers themselves (or
something to that effect).
The == and != operators retain the semantics that they inherited from
boost::shared_ptr, which is to test the pointers themselves for
identity. This means that even if neither aptr < bptr nor aptr > bptr
is true for two sortable_ptr pointers, the expression aptr == bptr may
or may not be true. If it is true this will mean that both aptr and
bptr refer to the same myclass object, which of course is neither less
than or greater than itself. If the expression aptr == bptr is false,
it will mean that aptr and bptr point at different objects that are
equivalent as to their sort order.
What is the advantage of this?
Before answering that question, we note that we have at leas not done
anything illegal by introducing these definitions. Neither
boost::shared_ptr nor sortable_ptr support the notion of pointer
arithmetic, so we do not need the < operator to be defined in any
particular way for that purpose.
We can still store sortable_ptr pointers in any of the sequence
containers in the Standard Template Library, since these only demand
that the == operator tests for proper identity between the objects that
are being stored in the list or deque or vector, which it does.
The sort algorithms in STL use the < operator of the objects in the
container by default, so if we sort a container of sortable_ptr
pointers, they will appear in the same order as the objects themselves
would have been sorted. For vectors and deques it can be considerably
faster to sort pointers rather than the objects themselves, since the
generic sort algorithms work by swapping the items in the container, and
it can be much faster to swap two small pointers than to swap two big
and complex objects.
For the associative containers set, multiset, map, and multimap, the
default is to use the < operator of the objects that are stored in the
container to define equivalence and sort order, which is exactly what we
want. When we store sortable_ptr pointers in an associative container,
the pointers will be sorted in the same order as the actual objects
would have been if we had stored them directly in a corresponding
associative container for objects. The definition of operator< ensures
that any two sortable_ptr pointers are equivalent if, and only if, the
two underlying objects are equivalent, so for sets and maps, which
cannot hold more than one object with equivalent keys, it will be the
same objects that are represented in the container whether the objects
are stored directly or via sortable_ptr pointers.
- - - -
The draft implementation looks like this at the moment:
//======================================
// 'sortable_ptr'
//--------------------------------------
template <class T, class Compare = std::less<T> >
class sortable_ptr : public boost::shared_ptr<T>
{
private:
typedef sortable_ptr self;
typedef boost::shared_ptr<T> base;
public:
// From 'boost::shared_ptr'
// typedef T element_type;
// typedef T value_type;
// typedef T * pointer;
// typedef typename detail::shared_ptr_traits<T>::reference reference;
//
typedef Compare element_compare;
element_compare element_comp() const {return element_compare();}
bool operator< (const self& other) const {return
element_compare()(**this, *other);}
bool operator> (const self& other) const {return other < *this;}
bool operator<= (const self& other) const {return !(other < *this);}
bool operator>= (const self& other) const {return !(other > *this);}
// Constructors
sortable_ptr() {}
template<class Y>
sortable_ptr(const sortable_ptr<Y>& other) : base(other) {}
template<class Y>
explicit sortable_ptr(const boost::shared_ptr<Y>& other) : base(other) {}
template<class Y>
explicit sortable_ptr(const boost::weak_ptr<Y>& other) : base(other) {}
template<class Y>
explicit sortable_ptr(std::auto_ptr<Y>& other) : base(other) {}
template<class Y>
explicit sortable_ptr(Y* other) : base(other) {}
// Assignment
};
- - - -
Does this represent a sensible idea, or am I barking up a totally
pointless tree?
Christian Engström