Is there an STL container that stores an array of elements in contiguous memory where the element size is specified at runtime?

Question

I\'m trying to create container that looks close to how my file spec works. It\'s like a vector but the type of the elements is defined by a hashtable.

If I knew the typ

Answer 1

Is there an STL container that stores an array of elements in contiguous memory where the element size is specified at runtime?

No.

What you're asking for looks like a specific implementation of a memory pool. Maybe the Boost.Pool library or other implementations would be of use to you? Writing your own one shouldn't be hard if you're used to work with raw memory and C++-specific creation/destruction of objects.

To answer your questions:

Does anything do this already?

To me the idea looks like a memory pool. There are tons of kind of memory pools so the implementation you'll want totally depends on your specific needs.

Is this a bad idea? Should I just create a vector and new up each element? I expect to have millions of elements. The range of sizes of foo will be 20 to 200 bytes.

It's not a bad idea if you're wanting to limit memory fragmentation. Pools are often used to fix this and other memory-organisation related prolems.

For example, it's very frequent in video games to do this, mostly on console but also on PC if you need high performance or a lot of memory.

I just wouldn't recommand bothering if you're making a prototype or if you don't have tons of data to allocate. If you do, then maybe first implementing the simplest allocation (with vectors and new) hidden behind a factory would be nice and would allow you to replace the factory implementation by using a pool. That way you first check that everything work, then optimize.

Answer 2

Given this code:

typedef Toffset uint; //byte offset;
typedef Ttype   uint; //enum of types
typedef std::pair<Toffset,Ttype> member;
typedef std::unordered_map<std::string, member> memberdefs;

memberdefs itemKey; 
itemKey["a"] = member(0, 0);
itemKey["b"] = member(4, 1);
itemKey["c"] = member(8, 2);
itemKey["d"] = member(12,1);
itemKey["e"] = member(16,3);
itemKey["f"] = member(17,2);

You could read into a char* buffer, and use a simple wrapper class. Still bug-prone and highly confusing. This demo has no iterator (though that would be simple), and requires an external buffer to stay in scope at least as long as the class does.

class interleaved_vector {
    const char* buffer;
    size_t count;
    size_t size;
    std::shared_ptr<memberdefs> members;
public: 
    class dynamic_object {
        const char* buffer;
        std::shared_ptr<memberdefs> members;
        friend interleaved_vector;
        dynamic_object(const char* buffer_, std::shared_ptr<memberdefs> members_)
        :buffer(buffer_), members(members_)
        {}
        dynamic_object& operator=(const dynamic_object& b) = delete;
    public:
        dynamic_object(const dynamic_object& b) 
        :buffer(b.buffer), members(b.members)
        {}
        template <class T>
        T get(const std::string& member) const {
            assert((*members)[member].second > 0); //no new members, requires pos sizes
            assert((*members)[member].second == sizeof(T));
            return *reinterpret_cast<T*>(buffer+(*members)[member].first); //technically undefined I think
        };
        template <>
        T* get<T*>(const std::string& member) const {
            assert((*members)[member].second > 0); //no new members, requires pos sizes
            assert((*members)[member].second == sizeof(T));
            return reinterpret_cast<T*>(buffer+(*members)[member].first); //technically undefined I think
        };
        void* operator[](const std::string& member) const {
            assert((*members)[member].second > 0); //no new members, requires pos sizes
            assert((*members)[member].second == sizeof(T));
            return reinterpret_cast<void*>(buffer+(*members)[member].first); //technically undefined I think
        };
    };
    interleaved_vector(const char* buffer_, size_t count_, size_t size_, const memberdefs& members_)
    :buffer(buffer_), count(count_), size(size_), members(members_) 
    {}
    dynamic_object get(size_t index) const { 
        assert(index<count);
        return dynamic_object(buffer + size*index, members);
    }
    dynamic_object operator[](size_t index) const { 
        assert(index<count);
        return dynamic_object(buffer + size*index, members);
    }
    size_t size() {
        return count;
    }
};

This would allow code such as:

size_t element_size = 32;
size_t num_elements = 1000000
char * buffer = new char[num_elements*element_size];
/*read into buffer*/
interleaved_vector iv(buffer, num_elements, element_size , members);
/*interleaved_vector DOES NOT COPY BUFFER. BUFFER MUST REMAIN IN SCOPE*/
for(int i=0; i<iv.size(); ++i) {
    for(memberdefs::const_iterator j=members.begin(); j!=members.end(); ++j) {
        dynamic_object ob = iv[i];
        std::cout << "data[" << i << "]." << j.first << " = ";
        std::cout << ob.get<int>(j.first) << '\n';
    }
}

This demo code assumes all members are ints (get) but hopefully you can see what's intended.

Answer 3

You could write your own class, but it'd be a serious pain. Better just go with vector (or boost::ptr_vector), which takes no effort on your part, and is easily readable to every programmer who comes along.