#ifndef TURBO_LIST_HPP
#define TURBO_LIST_HPP

#include<cstdint>
#include<cstdlib>
#include<cassert>

#ifndef TL_NOINLINE
#define TL_NOINLINE __attribute__((noinline))
#endif /* TL_NOINLINE */

#ifndef TL_LIKELY
#define TL_LIKELY(x) __builtin_expect(!!(x), 1)
#endif /* TL_LIKELY */
#ifndef TL_UNLIKELY
#define TL_UNLIKELY(x) __builtin_expect(!!(x), 0)
#endif /* TL_UNLIKELY */

typedef void*(MallocLike)(size_t);
typedef void(FreeLike)(void*);

template<typename T, MallocLike MALLOC = malloc, FreeLike FREE = free>
class TurboList {
	T *old;
	T *nex;

	uint32_t mid; // non-inclusive . . . m
	uint32_t end; // non-inclusive     e . . . .

	uint32_t capacity;

	TL_NOINLINE T& grow_and_insert(T elem) noexcept {
			// assert(mid == 0);
			if(old) FREE(old);
			old = nex;
			mid = end;
			capacity *= 2;
			nex = (T *) MALLOC(this->capacity * sizeof(T));

			// Will go into the INSERT code path here
			return insert(elem);
	}

public:
	inline TurboList(uint32_t initial_size = 0, uint32_t initial_cap = 16) noexcept :
			old(nullptr),
			mid(0),
			end(initial_size),
			capacity(initial_cap) {

		nex = (T *) MALLOC(this->capacity * sizeof(T));
	}

	inline ~TurboList() noexcept {
		if(nex) FREE(nex);
		if(old) FREE(old);
	}

	inline T& operator[](uint32_t i) noexcept {
		// This seem to be more often compiled to cmov
		// branchless conditional codes this way..
		T *base = (i < mid) ? old : nex;
		return base[i];
	}

	/** This is much faster than operator[] if you do small amounts of work per access */
	inline void iterate(void(callback)(T&)) noexcept {
		// old
		for(uint32_t i = 0; i < mid; ++i) {
			callback(old[i]);
		}
		// nex
		for(uint32_t i = mid; i < end; ++i) {
			callback(nex[i]);
		}
	}

	inline T& insert(T elem) noexcept {
		if(TL_LIKELY(end < capacity)) {

			// INSERT

			/* Same as this:
			if(mid > 0) {
				nex[mid - 1] = old[mid - 1];
				--mid;
			}
			*/
			bool hasmid = (mid > 0);
			mid -= hasmid;
			nex[mid] = hasmid ? old[mid] : nex[mid];

			return (nex[end++] = elem);
		} else {
			// GROW
			return grow_and_insert(elem);
		}
	}

	template<typename... Args>
	inline T& emplace(Args&&... args) {
		if(TL_LIKELY(end < capacity)) {

			// INSERT

			/* Same as this:
			if(mid > 0) {
				nex[mid - 1] = old[mid - 1];
				--mid;
			}
			*/
			bool hasmid = (mid > 0);
			mid -= hasmid;
			nex[mid] = hasmid ? old[mid] : nex[mid];

			// Placement new
			return *new (nex + end++) T(std::forward<Args>(args)...);
		} else {
			// GROW
			//
			// Rem.: I just chose this to be less optimized than
			//       it is possible by making a copy and reusing
			//       the existing grow and insert code instead of
			//       writing a new "grow_and_emplace" again.
			//
			//       This happens rarely so its probably fine and
			//       makes less template instantiations, smaller binary.
			return grow_and_insert(T(std::forward<Args>(args)...));
		}
	}

	// TODO: finalize() call which memcpy remaining elements of old into nex and then frees old and sets nullptr + mid = 0;

	inline uint32_t size() noexcept {
		return end;
	}
};

#endif /* TURBO_LIST_HPP */