#pragma once

#include <vector>
#include <algorithm>
#include <thread>
#include <mutex>

namespace algo {

    class MergeSort_v1 {
    private:
        template<typename Iterator, typename Comparator>
        static auto
        merge(Iterator start, Iterator middle, Iterator end, Comparator cmp, Iterator output_start) -> void {
            Iterator start_m = start;
            Iterator begin = output_start;
            Iterator start2 = middle + 1;

            //merge from input until one half completes
            while (start <= middle && start2 <= end) {
                if (cmp(*start, *start2)) {
                    *output_start = *start;
                    start++;
                } else {
                    *output_start = *start2;
                    start2++;
                }
                output_start++;
            }

            //try to finish first half
            while (start <= middle) {
                *output_start = *start;
                start++;
                output_start++;
            }

            while (start2 <= end) {
                *output_start = *start2;
                start2++;
                output_start++;
            }

            const auto size = std::distance(start_m, end);
            for (auto i = 0; i <= size; i++, start_m++, begin++) {
                *start_m = *begin;
            }

        }

        template<typename Container, typename Iterator, typename Comparator>
        static auto
        ms_split(Container &output_vec, Iterator start, Iterator end, Comparator cmp, Iterator output_start) -> void {
            Iterator mid = start;
            Iterator begin = output_start;

            if (std::distance(start, end) < 1) {
                return;
            } else {
                //move mid iterator litterally to the mid
                std::advance(mid, std::distance(start, end) / 2);
                //sort the first half within an recursion
                ms_split(output_vec, start, mid, cmp, output_start);

                //move output iterator
                std::advance(output_start, std::distance(start, mid + 1));
                //sort the second half within a recursion
                ms_split(output_vec, mid + 1, end, cmp, output_start);

                //merge everything together starting from the complete beginning
                merge(start, mid, end, cmp, begin);
            }
        }

    public:
        template<typename Iterator, typename Comparator>
        static auto sort(Iterator start, Iterator end, Comparator cmp) -> void {
            using valtype = typename std::iterator_traits<Iterator>::value_type;
            std::vector<valtype> temporary_dataset(std::distance(start, end));
            ms_split(temporary_dataset, start, end - 1, cmp, temporary_dataset.begin());
        }
    };

    class MergeSort_v2 {
    private:
        template<typename Iterator, typename Comparator>
        static auto
        merge(Iterator start, Iterator middle, Iterator end, Comparator cmp, Iterator output_start, std::recursive_mutex &dataset_guard) -> void {
            Iterator start_m = start;
            Iterator begin = output_start;
            Iterator start2 = middle + 1;

            //merge from input until one half completes
            while (start <= middle && start2 <= end) {
                if (cmp(*start, *start2)) {
                    *output_start = *start;
                    start++;
                } else {
                    *output_start = *start2;
                    start2++;
                }
                output_start++;
            }

            //try to finish first half
            while (start <= middle) {
                *output_start = *start;
                start++;
                output_start++;
            }

            while (start2 <= end) {
                *output_start = *start2;
                start2++;
                output_start++;
            }

            dataset_guard.lock();
            const auto size = std::distance(start_m, end);
            for (auto i = 0; i <= size; i++, start_m++, begin++) {
                *start_m = *begin;
            }
            dataset_guard.unlock();

        }

        template<typename Container, typename Iterator, typename Comparator>
        static auto ms_split(Container &output_vec, Iterator start, Iterator end, Comparator cmp, Iterator output_start,
                             int &nthreads, std::recursive_mutex &dataset_guard, std::mutex &depth_guard) -> void {
            Iterator mid = start;
            Iterator begin = output_start;

            if (std::distance(start, end) < 1) {
                //Quit on smalles list size (one element is always sorted)
                return;
            } else {
                if (nthreads > 1) {
                    depth_guard.lock();
                    nthreads -= 2;
                    depth_guard.unlock();
                    //move mid iterator litterally to the mid
                    std::advance(mid, std::distance(start, end) / 2);
                    //sort the first half within an recursion

                    std::thread t1([&]() { ms_split(output_vec, start, mid, cmp, output_start, nthreads, dataset_guard, depth_guard); });

                    //move output iteratoroutput_vec, start, mid, cmp, output_start
                    std::advance(output_start, std::distance(start, mid + 1));
                    //sort the second half within a recursion
                    std::thread t2([&]() { ms_split(output_vec, mid + 1, end, cmp, output_start, nthreads, dataset_guard, depth_guard); });


                    //merge everything together starting from the complete beginning
                    t1.join();
                    t2.join();
                    merge(start, mid, end, cmp, begin, dataset_guard);
                } else {
                    //move mid iterator litterally to the mid
                    std::advance(mid, std::distance(start, end) / 2);
                    //sort the first half within an recursion
                    ms_split(output_vec, start, mid, cmp, output_start, nthreads, dataset_guard, depth_guard);

                    //move output iterator
                    std::advance(output_start, std::distance(start, mid + 1));
                    //sort the second half within a recursion
                    ms_split(output_vec, mid + 1, end, cmp, output_start, nthreads, dataset_guard, depth_guard);

                    //merge everything together starting from the complete beginning
                    merge(start, mid, end, cmp, begin, dataset_guard);
                }
            }
        }

    public:
        template<typename Iterator, typename Comparator>
        static auto sort(Iterator start, Iterator end, Comparator cmp, int nthreads) -> void {
            using valtype = typename std::iterator_traits<Iterator>::value_type;
            std::vector<valtype> temporary_dataset(std::distance(start, end));
            std::recursive_mutex dataset_guard;
            std::mutex depth_guard;

            ms_split(temporary_dataset, start, end - 1, cmp, temporary_dataset.begin(), nthreads, dataset_guard, depth_guard);
        }
    };

}