#ifndef INCLUDED_C_QUICKSORTITERATIVE_H
#define INCLUDED_C_QUICKSORTITERATIVE_H

#ifndef INCLUDED_C_BASE_H
#include <c_Base.h>
#endif /*INCLUDED_C_BASE_H*/


#ifndef INCLUDED_C_MEMORY_H
#include <c_Memory.h>
#endif /*INCLUDED_C_MEMORY_H*/

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

/**
 * Internal macro helper to swap two arbitrary blocks of memory.
 */
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);
}

/**
 * Standard Lomuto or Hoare-style tracking partition loop.
 * Uses the rightmost element as the pivot for flat linear execution.
 */
C_STATIC_FORCE_INLINE
long long c_PartitionIterative(char* arr, long long left, long long right, c_size_t size,
                                             void* temp, int (*compar)(const void*, const void*)) {
    char* pivot = arr + (right * size);
    long long i = left - 1;

    for (long long j = left; j < right; j++) {
        if (compar(arr + (j * size), pivot) <= 0) {
            i++;
            c_SwapInternal(arr + (i * size), arr + (j * size), size, temp);
        }
    }
    c_SwapInternal(arr + ((i + 1) * size), arr + (right * size), size, temp);
    return (i + 1);
}

/**
 * Top-Level Custom Framework Entry Point for Non-Recursive Quicksort.
 * Time Complexity: O(n log n) average | Space Complexity: O(log n) explicit stack | Stable: No
 */
C_STATIC_FORCE_INLINE
void c_QuickSortIterative(void* base, c_size_t num, c_size_t size,
                          int (*compar)(const void*, const void*)) {
    if (base == NULL || num < 2 || size == 0) return;

    char* arr = (char*)base;

    // Optimization: Stack buffer allocation for pivot element swaps
#define QSORT_STACK_LIMIT 128
    char stack_buf[QSORT_STACK_LIMIT];
    void* temp = (size <= QSORT_STACK_LIMIT) ? stack_buf : C_ALLOC(size);
    if (temp == NULL) return;

    // Allocate the explicit partition boundary stack.
    // Max required stack depth for range tracking is 2 * ceil(log2(num)) + 2.
    // For a 64-bit address space, 128 slots safely handles any possible array size.
    long long range_stack[128];
    long long top = -1;

    // Push initial array boundaries onto the tracking stack
    range_stack[++top] = 0;
    range_stack[++top] = (long long)num - 1;

    // Keep processing partitions until the explicit boundary stack is empty
    while (top >= 0) {
        // Pop right and left boundaries
        long long right = range_stack[top--];
        long long left = range_stack[top--];

        // Execute linear pivot segmentation
        long long p = c_PartitionIterative(arr, left, right, size, temp, compar);

        // If there are elements on the left side of the pivot, push their range to the stack
        if (p - 1 > left) {
            range_stack[++top] = left;
            range_stack[++top] = p - 1;
        }

        // If there are elements on the right side of the pivot, push their range to the stack
        if (p + 1 < right) {
            range_stack[++top] = p + 1;
            range_stack[++top] = right;
        }
    }

    if (size > QSORT_STACK_LIMIT) {
        C_FREE(temp);
    }
#undef QSORT_STACK_LIMIT
}


#endif /*INCLUDED_C_QUICKSORTITERATIVE_H*/