How to sort in-place using the merge sort algorithm?

Question

Knuth left this as an exercise (Vol 3, 5.2.5). There do exist in-place merge sorts. They must be implemented carefully.

First, naive in-place merge such as described here isn't the right solution. It downgrades the performance to O(N²).

The idea is to sort part of the array while using the rest as working area for merging.

For example like the following merge function.

void wmerge(Key* xs, int i, int m, int j, int n, int w) {
    while (i < m && j < n)
        swap(xs, w++, xs[i] < xs[j] ? i++ : j++);
    while (i < m)
        swap(xs, w++, i++);
    while (j < n)
        swap(xs, w++, j++);
}  

It takes the array xs, the two sorted sub-arrays are represented as ranges [i, m) and [j, n) respectively. The working area starts from w. Compare with the standard merge algorithm given in most textbooks, this one exchanges the contents between the sorted sub-array and the working area. As the result, the previous working area contains the merged sorted elements, while the previous elements stored in the working area are moved to the two sub-arrays.

However, there are two constraints that must be satisfied:

1. The work area should be within the bounds of the array. In other words, it should be big enough to hold elements exchanged in without causing any out-of-bound error.
2. The work area can be overlapped with either of the two sorted arrays; however, it must ensure that none of the unmerged elements are overwritten.

With this merging algorithm defined, it's easy to imagine a solution, which can sort half of the array; The next question is, how to deal with the rest of the unsorted part stored in work area as shown below:

... unsorted 1/2 array ... | ... sorted 1/2 array ...

One intuitive idea is to recursive sort another half of the working area, thus there are only 1/4 elements haven't been sorted yet.

... unsorted 1/4 array ... | sorted 1/4 array B | sorted 1/2 array A ...

The key point at this stage is that we must merge the sorted 1/4 elements B with the sorted 1/2 elements A sooner or later.

Is the working area left, which only holds 1/4 elements, big enough to merge A and B? Unfortunately, it isn't.

However, the second constraint mentioned above gives us a hint, that we can exploit it by arranging the working area to overlap with either sub-array if we can ensure the merging sequence that the unmerged elements won't be overwritten.

Actually, instead of sorting the second half of the working area, we can sort the first half, and put the working area between the two sorted arrays like this:

... sorted 1/4 array B | unsorted work area | ... sorted 1/2 array A ...

This setup effectively arranges the work area overlap with the sub-array A. This idea is proposed in [Jyrki Katajainen, Tomi Pasanen, Jukka Teuhola. ``Practical in-place mergesort''. Nordic Journal of Computing, 1996].

So the only thing left is to repeat the above step, which reduces the working area from 1/2, 1/4, 1/8, … When the working area becomes small enough (for example, only two elements left), we can switch to a trivial insertion sort to end this algorithm.

Here is the implementation in ANSI C based on this paper.

void imsort(Key* xs, int l, int u);

void swap(Key* xs, int i, int j) {
    Key tmp = xs[i]; xs[i] = xs[j]; xs[j] = tmp;
}

/* 
 * sort xs[l, u), and put result to working area w. 
 * constraint, len(w) == u - l
 */
void wsort(Key* xs, int l, int u, int w) {
    int m;
    if (u - l > 1) {
        m = l + (u - l) / 2;
        imsort(xs, l, m);
        imsort(xs, m, u);
        wmerge(xs, l, m, m, u, w);
    }
    else
        while (l < u)
            swap(xs, l++, w++);
}

void imsort(Key* xs, int l, int u) {
    int m, n, w;
    if (u - l > 1) {
        m = l + (u - l) / 2;
        w = l + u - m;
        wsort(xs, l, m, w); /* the last half contains sorted elements */
        while (w - l > 2) {
            n = w;
            w = l + (n - l + 1) / 2;
            wsort(xs, w, n, l);  /* the first half of the previous working area contains sorted elements */
            wmerge(xs, l, l + n - w, n, u, w);
        }
        for (n = w; n > l; --n) /*switch to insertion sort*/
            for (m = n; m < u && xs[m] < xs[m-1]; ++m)
                swap(xs, m, m - 1);
    }
}

Where wmerge is defined previously.

The full source code can be found here and the detailed explanation can be found here

By the way, this version isn't the fastest merge sort because it needs more swap operations. According to my test, it's faster than the standard version, which allocates extra spaces in every recursion. But it's slower than the optimized version, which doubles the original array in advance and uses it for further merging.

Including its "big result", this paper describes a couple of variants of in-place merge sort (PDF):

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.22.5514&rep=rep1&type=pdf

In-place sorting with fewer moves

Jyrki Katajainen, Tomi A. Pasanen

It is shown that an array of n elements can be sorted using O(1) extra space, O(n log n / log log n) element moves, and n log₂n + O(n log log n) comparisons. This is the first in-place sorting algorithm requiring o(n log n) moves in the worst case while guaranteeing O(n log n) comparisons, but due to the constant factors involved the algorithm is predominantly of theoretical interest.

I think this is relevant too. I have a printout of it lying around, passed on to me by a colleague, but I haven't read it. It seems to cover basic theory, but I'm not familiar enough with the topic to judge how comprehensively:

http://comjnl.oxfordjournals.org/cgi/content/abstract/38/8/681

Optimal Stable Merging

Antonios Symvonis

This paper shows how to stably merge two sequences A and B of sizes m and n, m ≤ n, respectively, with O(m+n) assignments, O(mlog(n/m+1)) comparisons and using only a constant amount of additional space. This result matches all known lower bounds...

It really isn't easy or efficient, and I suggest you don't do it unless you really have to (and you probably don't have to unless this is homework since the applications of inplace merging are mostly theoretical). Can't you use quicksort instead? Quicksort will be faster anyway with a few simpler optimizations and its extra memory is O(log N).

Anyway, if you must do it then you must. Here's what I found: one and two. I'm not familiar with the inplace merge sort, but it seems like the basic idea is to use rotations to facilitate merging two arrays without using extra memory.

Note that this is slower even than the classic merge sort that's not inplace.

The critical step is getting the merge itself to be in-place. It's not as difficult as those sources make out, but you lose something when you try.

Looking at one step of the merge:

[...list-sorted...|x...list-A...|y...list-B...]

We know that the sorted sequence is less than everything else, that x is less than everything else in A, and that y is less than everything else in B. In the case where x is less than or equal to y, you just move your pointer to the start of A on one. In the case where y is less than x, you've got to shuffle y past the whole of A to sorted. That last step is what makes this expensive (except in degenerate cases).

It's generally cheaper (especially when the arrays only actually contain single words per element, e.g., a pointer to a string or structure) to trade off some space for time and have a separate temporary array that you sort back and forth between.

An example of bufferless mergesort in C.

#define SWAP(type, a, b) \
    do { type t=(a);(a)=(b);(b)=t; } while (0)

static void reverse_(int* a, int* b)
{
    for ( --b; a < b; a++, b-- )
       SWAP(int, *a, *b);
}
static int* rotate_(int* a, int* b, int* c)
/* swap the sequence [a,b) with [b,c). */
{
    if (a != b && b != c)
     {
       reverse_(a, b);
       reverse_(b, c);
       reverse_(a, c);
     }
    return a + (c - b);
}

static int* lower_bound_(int* a, int* b, const int key)
/* find first element not less than @p key in sorted sequence or end of
 * sequence (@p b) if not found. */
{
    int i;
    for ( i = b-a; i != 0; i /= 2 )
     {
       int* mid = a + i/2;
       if (*mid < key)
          a = mid + 1, i--;
     }
    return a;
}
static int* upper_bound_(int* a, int* b, const int key)
/* find first element greater than @p key in sorted sequence or end of
 * sequence (@p b) if not found. */
{
    int i;
    for ( i = b-a; i != 0; i /= 2 )
     {
       int* mid = a + i/2;
       if (*mid <= key)
          a = mid + 1, i--;
     }
    return a;
}

static void ip_merge_(int* a, int* b, int* c)
/* inplace merge. */
{
    int n1 = b - a;
    int n2 = c - b;

    if (n1 == 0 || n2 == 0)
       return;
    if (n1 == 1 && n2 == 1)
     {
       if (*b < *a)
          SWAP(int, *a, *b);
     }
    else
     {
       int* p, * q;

       if (n1 <= n2)
          p = upper_bound_(a, b, *(q = b+n2/2));
       else
          q = lower_bound_(b, c, *(p = a+n1/2));
       b = rotate_(p, b, q);

       ip_merge_(a, p, b);
       ip_merge_(b, q, c);
     }
}

void mergesort(int* v, int n)
{
    if (n > 1)
     {
       int h = n/2;
       mergesort(v, h); mergesort(v+h, n-h);
       ip_merge_(v, v+h, v+n);
     }
}

An example of adaptive mergesort (optimized).

Adds support code and modifications to accelerate the merge when an auxiliary buffer of any size is available (still works without additional memory). Uses forward and backward merging, ring rotation, small sequence merging and sorting, and iterative mergesort.

#include <stdlib.h>
#include <string.h>

static int* copy_(const int* a, const int* b, int* out)
{
    int count = b - a;
    if (a != out)
       memcpy(out, a, count*sizeof(int));
    return out + count;
}
static int* copy_backward_(const int* a, const int* b, int* out)
{
    int count = b - a;
    if (b != out)
       memmove(out - count, a, count*sizeof(int));
    return out - count;
}

static int* merge_(const int* a1, const int* b1, const int* a2,
  const int* b2, int* out)
{
    while ( a1 != b1 && a2 != b2 )
       *out++ = (*a1 <= *a2) ? *a1++ : *a2++;
    return copy_(a2, b2, copy_(a1, b1, out));
}
static int* merge_backward_(const int* a1, const int* b1,
  const int* a2, const int* b2, int* out)
{
    while ( a1 != b1 && a2 != b2 )
       *--out = (*(b1-1) > *(b2-1)) ? *--b1 : *--b2;
    return copy_backward_(a1, b1, copy_backward_(a2, b2, out));
}

static unsigned int gcd_(unsigned int m, unsigned int n)
{
    while ( n != 0 )
     {
       unsigned int t = m % n;
       m = n;
       n = t;
     }
    return m;
}
static void rotate_inner_(const int length, const int stride,
  int* first, int* last)
{
    int* p, * next = first, x = *first;
    while ( 1 )
     {
       p = next;
       if ((next += stride) >= last)
          next -= length;
       if (next == first)
          break;
       *p = *next;
     }
    *p = x;
}
static int* rotate_(int* a, int* b, int* c)
/* swap the sequence [a,b) with [b,c). */
{
    if (a != b && b != c)
     {
       int n1 = c - a;
       int n2 = b - a;

       int* i = a;
       int* j = a + gcd_(n1, n2);

       for ( ; i != j; i++ )
          rotate_inner_(n1, n2, i, c);
     }
    return a + (c - b);
}

static void ip_merge_small_(int* a, int* b, int* c)
/* inplace merge.
 * @note faster for small sequences. */
{
    while ( a != b && b != c )
       if (*a <= *b)
          a++;
       else
        {
          int* p = b+1;
          while ( p != c && *p < *a )
             p++;
          rotate_(a, b, p);
          b = p;
        }
}
static void ip_merge_(int* a, int* b, int* c, int* t, const int ts)
/* inplace merge.
 * @note works with or without additional memory. */
{
    int n1 = b - a;
    int n2 = c - b;

    if (n1 <= n2 && n1 <= ts)
     {
       merge_(t, copy_(a, b, t), b, c, a);
     }
    else if (n2 <= ts)
     {
       merge_backward_(a, b, t, copy_(b, c, t), c);
     }
    /* merge without buffer. */
    else if (n1 + n2 < 48)
     {
       ip_merge_small_(a, b, c);
     }
    else
     {
       int* p, * q;

       if (n1 <= n2)
          p = upper_bound_(a, b, *(q = b+n2/2));
       else
          q = lower_bound_(b, c, *(p = a+n1/2));
       b = rotate_(p, b, q);

       ip_merge_(a, p, b, t, ts);
       ip_merge_(b, q, c, t, ts);
     }
}
static void ip_merge_chunk_(const int cs, int* a, int* b, int* t,
  const int ts)
{
    int* p = a + cs*2;
    for ( ; p <= b; a = p, p += cs*2 )
       ip_merge_(a, a+cs, p, t, ts);
    if (a+cs < b)
       ip_merge_(a, a+cs, b, t, ts);
}

static void smallsort_(int* a, int* b)
/* insertion sort.
 * @note any stable sort with low setup cost will do. */
{
    int* p, * q;
    for ( p = a+1; p < b; p++ )
     {
       int x = *p;
       for ( q = p; a < q && x < *(q-1); q-- )
          *q = *(q-1);
       *q = x;
     }
}
static void smallsort_chunk_(const int cs, int* a, int* b)
{
    int* p = a + cs;
    for ( ; p <= b; a = p, p += cs )
       smallsort_(a, p);
    smallsort_(a, b);
}

static void mergesort_lower_(int* v, int n, int* t, const int ts)
{
    int cs = 16;
    smallsort_chunk_(cs, v, v+n);
    for ( ; cs < n; cs *= 2 )
       ip_merge_chunk_(cs, v, v+n, t, ts);
}

static void* get_buffer_(int size, int* final)
{
    void* p = NULL;
    while ( size != 0 && (p = malloc(size)) == NULL )
       size /= 2;
    *final = size;
    return p;
}
void mergesort(int* v, int n)
{
    /* @note buffer size may be in the range [0,(n+1)/2]. */
    int request = (n+1)/2 * sizeof(int);
    int actual;
    int* t = (int*) get_buffer_(request, &actual);

    /* @note allocation failure okay. */
    int tsize = actual / sizeof(int);
    mergesort_lower_(v, n, t, tsize);
    free(t);
}

This answer has a code example, which implements the algorithm described in the paper Practical In-Place Merging by Bing-Chao Huang and Michael A. Langston. I have to admit that I do not understand the details, but the given complexity of the merge step is O(n).

From a practical perspective, there is evidence that pure in-place implementations are not performing better in real world scenarios. For example, the C++ standard defines std::inplace_merge, which is as the name implies an in-place merge operation.

Assuming that C++ libraries are typically very well optimized, it is interesting to see how it is implemented:

1) libstdc++ (part of the GCC code base): std::inplace_merge

The implementation delegates to __inplace_merge, which dodges the problem by trying to allocate a temporary buffer:

typedef _Temporary_buffer<_BidirectionalIterator, _ValueType> _TmpBuf;
_TmpBuf __buf(__first, __len1 + __len2);

if (__buf.begin() == 0)
  std::__merge_without_buffer
    (__first, __middle, __last, __len1, __len2, __comp);
else
  std::__merge_adaptive
   (__first, __middle, __last, __len1, __len2, __buf.begin(),
     _DistanceType(__buf.size()), __comp);

Otherwise, it falls back to an implementation (__merge_without_buffer), which requires no extra memory, but no longer runs in O(n) time.

2) libc++ (part of the Clang code base): std::inplace_merge

Looks similar. It delegates to a function, which also tries to allocate a buffer. Depending on whether it got enough elements, it will choose the implementation. The constant-memory fallback function is called __buffered_inplace_merge.

Maybe even the fallback is still O(n) time, but the point is that they do not use the implementation if temporary memory is available.

---

Note that the C++ standard explicitly gives implementations the freedom to choose this approach by lowering the required complexity from O(n) to O(N log N):

Complexity: Exactly N-1 comparisons if enough additional memory is available. If the memory is insufficient, O(N log N) comparisons.

Of course, this cannot be taken as a proof that constant space in-place merges in O(n) time should never be used. On the other hand, if it would be faster, the optimized C++ libraries would probably switch to that type of implementation.