Question

I would like to understand better the mechanics and the issues behind creating library and I have decided to start from the std::auto_ptr.
I am familiar with the syntax and the usage, however I am trying to understand the implementation details. I have implemented my version of the auto_ptr and it looks like this:

#ifndef _AUTO_PTR_H
#define _AUTO_PTR_H

namespace test
{    
    template<typename T>
    class auto_ptr
    {
        public:
            //ctor, dtor, copy ctor and assignment
            explicit auto_ptr(T* ptr = 0) throw() 
                : ptr_(ptr) {}

            auto_ptr(auto_ptr& other) throw()
                :ptr_(other.release()) {}

            auto_ptr<T>& operator=(auto_ptr<T>& other)
            {
                reset(other.release());
                return *this;
            }

            template<typename U>
            auto_ptr<T>& operator=(auto_ptr<U>& other) throw()
            {
                reset(other.release());
                return *this;
            }

            template<typename U>
            auto_ptr(auto_ptr<U>& other) throw() 
                : ptr_(other.release()) {}

            //member
            ~auto_ptr() throw() { delete ptr_; }

            T& operator*() const throw() { return *ptr_; }

            T* operator->() const throw() { return ptr_; }

            T* get() const throw() { return ptr_; }         

            T* release() throw()
            { 
                T* temp = ptr_;
                ptr_ = 0;
                return temp;
            }

            void reset(T* ptr = 0) throw()
            {
                if (ptr_ != ptr)
                {
                    delete ptr_;
                    ptr_ = ptr;
                }
            }

        private:
            T* ptr_;
    };
}

#endif

My class is doing pretty much the job then I have read this question and it is clear why auto_ptr_ref is there for. However I would like to have an actual example when the auto_ptr behaves differently without add auto_ptr_ref.

My implementation works correctly with this functions:

template <typename T>
test::auto_ptr<T> source()
{
    return test::auto_ptr<T>(new T());
}

template <typename T>
void sink(test::auto_ptr<T> bye)
{
}

template <typename T>
void useButNotSink(const test::auto_ptr<T> seeYouLater)
{
    std::cout << *seeYouLater << std::endl;
}

With this test program:

test::auto_ptr<double> copyConstructed(source<double>());

std::cout << *copyConstructed << std::endl;
std::cout << copyConstructed.get() << std::endl;

test::auto_ptr<double> assigned(new double(10));
assigned = source<double>();

std::cout << *assigned << std::endl;
std::cout << assigned.get() << std::endl;
*assigned = 2044;

std::cout << *assigned << std::endl;
useButNotSink(assigned);
sink(assigned);
std::cout << assigned.get() << std::endl;

I am sure I am missing something,can you give me an example where the auto_ptr_ref struct and the conversion operators are mandatory to get the correct behaviour?

Thanks

Answer 1

Your very first line

test::auto_ptr<double> copyConstructed(source<double>());

is already non-compilable by any standard-compliant compiler. MSVC will allow it when the C++ language extensions are enabled. Disable the extensions and you will immediately realize the need for auto_ptr_ref.

The mechanics behind auto_ptr_ref was too complicated for MSVC 6.0, so in order to support some resemblance of full std::auto_ptr functionality that compiler had to rely on this compiler extension, i.e. allowing non-const references to be bound to temporary objects. The problem was fixed in later versions of the compiler, but this compiler extension remains enabled by default to this day.