compact vector a small memory foot print vector - c++

Question

I maintain a old C++ application which has a lot of classes like below:
class ClassWithALotOfVectors
{
std::vector<int> _vector1;
std::vector<int> _vector2;
// a lot of other vector datamembers go here
std::vector<double> _vectorN;
};
that is - a data type where members a vectors of double or int.
The thing is - these vectors are never populated at the same time - and therefore
when we create 100000 instances of ClassWithALotOfVectors - memory usage adds up
to impressive number even though only 10% of these vector are in use.
So I decided to write a small "allocate on demand" vector class.
It is a wrapper around std::vector - where internal vector only create when
accessed for the first time (using two getters - ref() & const_ref() methods)
When I replaced std::vector in a ClassWithALotOfVectors class with a compact_vector
as below:
class ClassWithALotOfVectors2
{
compact_vector<int> _vector1;
compact_vector<int> _vector2;
compact_vector<double> _vectorN;
};
did a few tests and the results were promising - memory usage went down
dramatically, but suddenly found that application does not release memory
at the end - the memory consumptions grows in much slower rate than it
is used to be - but application does not seem to be deallocating the memory
at the end.
Can you look at my implementation of the compact_vector
and see if you can spot something wrong with a memory management.
template <class T> class compact_vector
{
public:
compact_vector<T>()
:_data(NULL)
{
}
~compact_vector<T>()
{
clear();
}
compact_vector<T>(const compact_vector<T> & rhs)
:_data(NULL)
{
if (NULL != rhs._data)
{
_data = new std::vector<T>(*(rhs._data));
}
else
{
clear();
}
}
// assignment
//
compact_vector<T> & operator=(const compact_vector<T> & rhs)
{
if (this != &rhs)
{
clear();
if (NULL != rhs._data)
{
_data = new std::vector<T>(*rhs._data);
}
}
return *this;
}
compact_vector<T> & operator=(const std::vector<T> & rhs)
{
clear();
if (!rhs.empty())
{
_data = new std::vector<T>(rhs);
}
return *this;
}
const std::vector<T> & const_ref()
{
createInternal();
return *_data;
}
std::vector<T> & ref()
{
createInternal();
return *_data;
}
void clear()
{
if (NULL != _data)
{
_data->clear();
delete _data;
_data = NULL;
}
}
private:
void createInternal()
{
if (NULL == _data)
{
_data = new std::vector<T>();
}
}
private:
compact_vector<T>(const std::vector<T> & rhs)
{
}
compact_vector<T>(std::vector<T> & rhs)
{
}
compact_vector<T>(std::vector<T> rhs)
{
}
std::vector<T> * _data;
};

Answer 1

Most implementations of std::vector don't acquire memory until you need it, and the size of a vector is usually just a few (3 + possibly extra debug information) pointers. That is, std::vector<int>() will not allocate space for any object. I believe that you are barking at the wrong tree here.
Why is your original code having a much higher memory usage? Are you expecting clear() to release memory?
If so you are mistaken. The clear() function destroys the contained elements but does not release the allocated memory. Consider adding a wrapper to clear the memory that uses the following idiom:
std::vector<int>().swap( _data );
What the previous line does is creating a new empty vector (usually no memory attached, not mandated, but common implementation) and swapping the contents of the two vectors. At the end of the expression, the temporary contains the data that was originally held by the _data vector and _data is empty and with no memory (implementation defined). The temporary is destructed at the end of the full expression and memory is released.
Alternatively, if your compiler supports C++11, you can use shrink_to_fit() after calling clear(). The standard does not require shrink_to_fit() to actually shrink to fit, but for an empty vector I would expect the implementation to do it. Test it by calling capacity() after the call and seeing whether it has gone down to 0 or not.

Answer 2

Use smart pointers. In this case, std::unique_ptr with hand-made copy ctor / assignment operator. Suddenly, all your problems are solved. It's like magic!
Sorry to sound so snarky, but memory leaks always have the same answer: smart pointer.

Answer 3

Since you are saying it's an old application and can't do much with it. How about having all the vectors having zero size reserved during construction.
It's tedious, but since old app and not much is expected to be modified. May be worth the effort for once.
class ClassWithALotOfVectors
{
std::vector<int> _vector1;
std::vector<int> _vector2;
// a lot of other vector datamembers go here
std::vector<double> _vectorN;
public:
ClassWithALotOfVectors() : Init() { }
void Init()
{
vector1.reserve(0);
/// Like wise other vectors.
}
};

Answer 4

The idea that only one vector at a time is ever in use sounds like the union concept. Sure enough, C++ has an element called an anonymous union that does precisely this.
class ClassWithALotOfVectors {
public:
union {
std::vector<double> _vector1;
std::vector<double> _vector2;
};
ClassWithALotOfVectors() { new(&_vector1) std::vector<double>; };
~ClassWithALotOfVectors() { _vector1.~vector<double>(); };
};
Because the union can't know which element's constructor to call, default constructors are disabled in the union and you have to manually construct and deconstruct the union's element, but I think this will accomplish what you are looking for, aliasing _vector1 and _vector2 to the same element but placing both in the ClassWithALotOfVectors namespace, then deallocating the vector when the destructor for ClassWithALotOfVectors is called.
For more on anonymous unions see:
CPP reference: Anonymous Unions