One way to implement a C++11 array that has its elements initialized by a function of their index calculated by the compiler and have the results stored in the data section (.rodata) of the application image is to use templates, partial specialization and constexpr as follows:
#include <iostream>
#include <array>
using namespace std;
constexpr int N = 1000000;
constexpr int f(int x) { return x*2; }
typedef array<int, N> A;
template<int... i> constexpr A fs() { return A{{ f(i)... }}; }
template<int...> struct S;
template<int... i> struct S<0,i...>
{ static constexpr A gs() { return fs<0,i...>(); } };
template<int i, int... j> struct S<i,j...>
{ static constexpr A gs() { return S<i-1,i,j...>::gs(); } };
constexpr auto X = S<N-1>::gs();
int main()
{
cout << X[3] << endl;
}
This doesn't work for large values of N:
error: constexpr evaluation depth exceeds maximum of 512
This is because of the head-tail style of recursive template evaluation, that has linear depth in terms of N.
Is there a way to do it such that the evaluation depth is logarithmic in terms of N, rather than linear? (and hence would avoid the default depth limit)
See Question&Answers more detail:os