Implicit typecasting for array objects in C++

Question

I am almost sure that this cannot be done, but I will ask anyway. I have to use a C based library, which defines a numeric vector as array of floats, and lots of arithmetic functions to use them. I want to create a trivial class that can be easily casted to that type, with the addition of useful operators. Let's see a MWE:

#include <iostream>

using vector_type = float[3];

class NewType
{
public:
    float& operator [](std::size_t i) { return v[i]; }
    const float& operator [](std::size_t i) const { return v[i]; }

    operator vector_type& () { return v; }
    vector_type* operator & () { return &v; }

private:
    vector_type v;
};

int main()
{
    NewType t;
    t[0] = 0.f; t[1] = 1.f; t[2] = 2.f;

    const vector_type& v = t;
    std::cout << "v(" << v[0] << "," << v[1] << "," << v[2] << ")" << std::endl;

    return 0;
}

This works flawlessly. The problem arises when we start using arrays. Let's write a new main function:

int main()
{
    constexpr std::size_t size = 10;
    vector_type v1[size];                   // OK
    NewType v2[size];                       // OK
    vector_type* v3 = v2;                   // No way, NewType* cannot be
                                            // converted to float (*)[3]
    vector_type* v4 =
        reinterpret_cast<vector_type*>(v2); // OK

    return 0;
}

The reinterpret_cast works, but it makes the code less readable and the conversion between vector_type and NewType not transparent. As far as I know, it is not possible, according to C++11 and C++14 standards, to make the NewType class implicitly castable when using arrays. Is it completely true? Are there any sort of caveats that allow this convertion?

P.s.: Please, do not start commenting about the risks of using reinterpret_cast and so on. I am aware of the risks, I know that the compiler could add some padding, and I already have some static_assert checks to avoid memory problems.

[Edit] I want to make the problem easier to understand. Let's make a different example:

struct original_vector
{
    float x;
    float y;
    float z;
};

class NewType : public original_vector
{
public:
    /* Useful functions here */
};

If this would be my case, everything would be easy! The type used in the C library would be original_vector, and I could create a derived class and I could add any sort of method.

The problem is that, in my real case, the original_vector is not a class/struct, but a raw array! And obviously, I cannot inherit it. Maybe now it is more clear the reason I am asking this question. ;)

Answer 1

I think that it is not the best solution, but it is the best I can think using C++14 capabilities. Maybe, if runtime-sized member allocation is introduced in future standards (the proposal for C++14 has been rejected), a better solution will be possible. But for now...

#include <iostream>
#include <memory>
#include <cassert>

using vector_type = float[3];

class NewType
{
public:
    float& operator [](std::size_t i) { return v[i]; }
    const float& operator [](std::size_t i) const { return v[i]; }

    operator vector_type&  () { return v; }
    vector_type* operator & () { return &v; }

private:
    vector_type v;
};

class NewTypeArray
{
public:
    NewTypeArray() : size(0), newType(nullptr) {}
    NewTypeArray(std::size_t size) : size(size) { assert(size > 0); newType = new NewType[size]; }
    ~NewTypeArray() { if(size > 0) delete[] newType; }

    NewType& operator[](std::size_t i) { return newType[i]; } 
    operator vector_type* () { return static_cast<vector_type*>(&newType[0]); }

private:
    std::size_t size;
    NewType* newType;
};

static_assert(sizeof(NewType) == sizeof(vector_type) and sizeof(NewType[7]) == sizeof(vector_type[7]),
    "NewType and vector_type have different memory layouts");

Obviously, the NewTypeArray could be modified, implementing vector-oriented methods, move constructor and assignment (like in my real-case code). Instances of NewTypeArray could be directly passed to functions which takes vector_type* as argument and, thanks to the static_assert, there should not be any sort of problems with memory management.