#include <c_QuickSort.h>


/**
 * Internal macro helper to swap two arbitrary blocks of memory of given size.
 */
C_STATIC_FORCE_INLINE
void c_SwapInternal(char* a, char* b, c_size_t size, void* temp) {
    if (a == b) return;
    memcpy(temp, a, size);
    memcpy(a, b, size);
    memcpy(b, temp, size);
}

/**
 * Chooses the median of left, center, and right elements as the pivot,
 * hides it at (right - 1), and returns a pointer to it.
 */
C_STATIC_FORCE_INLINE
void* c_MedianOfThree(char* arr, long long left, long long right, c_size_t size,
                                    void* temp, int (*compar)(const void*, const void*)) {
    long long center = left + (right - left) / 2;

    // Order left, center, right
    if (compar(arr + (left * size), arr + (center * size)) > 0) {
        c_SwapInternal(arr + (left * size), arr + (center * size), size, temp);
    }
    if (compar(arr + (left * size), arr + (right * size)) > 0) {
        c_SwapInternal(arr + (left * size), arr + (right * size), size, temp);
    }
    if (compar(arr + (center * size), arr + (right * size)) > 0) {
        c_SwapInternal(arr + (center * size), arr + (right * size), size, temp);
    }

    // Place pivot at position (right - 1)
    c_SwapInternal(arr + (center * size), arr + ((right - 1) * size), size, temp);
    return arr + ((right - 1) * size);
}


/* ------------------------------------------------------------------------------------------------------------------ */
/*  */

void c_QuickSortSub(char* arr, long long left, long long right, c_size_t size, void* temp,
                           int (*compar)(const void*, const void*)) {
    // Optimization: Fallback to manual insertion sort for tiny arrays to prune deep recursion
    if (left + 10 > right) {
        // Basic Inline Insertion Sort boundary loop
        for (long long i = left + 1; i <= right; i++) {
            memcpy(temp, arr + (i * size), size);
            long long j = i;
            while (j > left && compar(arr + ((j - 1) * size), temp) > 0) {
                memcpy(arr + (j * size), arr + ((j - 1) * size), size);
                j--;
            }
            memcpy(arr + (j * size), temp, size);
        }
        return;
    }

    // Retrieve Median Pivot
    c_MedianOfThree(arr, left, right, size, temp, compar);

    long long i = left;
    long long j = right - 1;

    // Hoare Partitioning Loop
    while (1) {
        while (compar(arr + ((++i) * size), arr + ((right - 1) * size)) < 0);
        while (compar(arr + ((--j) * size), arr + ((right - 1) * size)) > 0);
        if (i < j) {
            c_SwapInternal(arr + (i * size), arr + (j * size), size, temp);
        } else {
            break;
        }
    }

    // Restore pivot to its correct final slot
    c_SwapInternal(arr + (i * size), arr + ((right - 1) * size), size, temp);

    // Recursively execute left and right segments
    c_QuickSortSub(arr, left, i - 1, size, temp, compar);
    c_QuickSortSub(arr, i + 1, right, size, temp, compar);
}