xref: /illumos-gate/usr/src/common/util/qsort.c (revision 7c478bd9)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License, Version 1.0 only
 * (the "License").  You may not use this file except in compliance
 * with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident	"%Z%%M%	%I%	%E% SMI"


#if !defined(_KERNEL) && !defined(_KMDB)
#include "synonyms.h"
#endif /* !_KERNEL && !_KMDB */

#include <sys/types.h>

#if !defined(_KERNEL) && !defined(_KMDB)
#include <stdlib.h>
#include <mtlib.h>
#include <synch.h>
#endif /* !_KERNEL && !_KMDB */

#include "qsort.h"

static void swapp32(uint32_t *r1, uint32_t *r2, size_t cnt);
static void swapp64(uint64_t *r1, uint64_t *r2, size_t cnt);
static void swapi(uint32_t *r1, uint32_t *r2, size_t cnt);
static void swapb(char *r1, char *r2, size_t cnt);

/*
 * choose a median of 3 values
 *
 * note: cstyle specifically prohibits nested conditional operators
 * but this is the only way to do the median of 3 function in-line
 */
#define	med3(a, b, c) (cmp((a), (b)) < 0) \
	? ((cmp((b), (c)) < 0) ? (b) : (cmp((a), (c)) < 0) ? (c) : (a)) \
	: ((cmp((b), (c)) > 0) ? (b) : (cmp((a), (c)) > 0) ? (c) : (a))

#define	THRESH_L	5	/* threshold for insertion sort */
#define	THRESH_M3	20	/* threshold for median of 3 */
#define	THRESH_M9	50	/* threshold for median of 9 */

typedef struct {
	char	*b_lim;
	size_t	nrec;
} stk_t;

/*
 * qsort() is a general purpose, in-place sorting routine using a
 * user provided call back function for comparisons.  This implementation
 * utilizes a ternary quicksort algorithm, and cuts over to an
 * insertion sort for partitions involving fewer than THRESH_L records.
 *
 * Potential User Errors
 *   There is no return value from qsort, this function has no method
 *   of alerting the user that a sort did not work or could not work.
 *   We do not print an error message or exit the process or thread,
 *   Even if we can detect an error, We CANNOT silently return without
 *   sorting the data, if we did so the user could never be sure the
 *   sort completed successfully.
 *   It is possible we could change the return value of sort from void
 *   to int and return success or some error codes, but this gets into
 *   standards  and compatibility issues.
 *
 *   Examples of qsort parameter errors might be
 *   1) record size (rsiz) equal to 0
 *      qsort will loop and never return.
 *   2) record size (rsiz) less than 0
 *      rsiz is unsigned, so a negative value is insanely large
 *   3) number of records (nrec) is 0
 *      This is legal - qsort will return without examining any records
 *   4) number of records (nrec) is less than 0
 *      nrec is unsigned, so a negative value is insanely large.
 *   5) nrec * rsiz > memory allocation for sort array
 *      a segment violation may occur
 *      corruption of other memory may occur
 *   6) The base address of the sort array is invalid
 *      a segment violation may occur
 *      corruption of other memory may occur
 *   7) The user call back function is invalid
 *      we may get alignment errors or segment violations
 *      we may jump into never-never land
 *
 *   Some less obvious errors might be
 *   8) The user compare function is not comparing correctly
 *   9) The user compare function modifies the data records
 */

void
qsort(
	void		*basep,
	size_t		nrec,
	size_t		rsiz,
	int		(*cmp)(const void *, const void *))
{
	size_t		i;		/* temporary variable */

	/* variables used by swap */
	void		(*swapf)(char *, char *, size_t);
	size_t		loops;

	/* variables used by sort */
	stk_t		stack[8 * sizeof (nrec) + 1];
	stk_t		*sp;
	char		*b_lim;		/* bottom limit */
	char		*b_dup;		/* bottom duplicate */
	char		*b_par;		/* bottom partition */
	char		*t_lim;		/* top limit */
	char		*t_dup;		/* top duplicate */
	char		*t_par;		/* top partition */
	char		*m1, *m2, *m3;	/* median pointers */
	uintptr_t	d_bytelength;	/* byte length of duplicate records */
	int		b_nrec;
	int		t_nrec;
	int		cv;		/* results of compare (bottom / top) */

	/*
	 * choose a swap function based on alignment and size
	 *
	 * The qsort function sorts an array of fixed length records.
	 * We have very limited knowledge about the data record itself.
	 * It may be that the data record is in the array we are sorting
	 * or it may be that the array contains pointers or indexes to
	 * the actual data record and all that we are sorting is the indexes.
	 *
	 * The following decision will choose an optimal swap function
	 * based on the size and alignment of the data records
	 *   swapp64	will swap 64 bit pointers
	 *   swapp32	will swap 32 bit pointers
	 *   swapi	will swap an array of 32 bit integers
	 *   swapb	will swap an array of 8 bit characters
	 *
	 * swapi and swapb will also require the variable loops to be set
	 * to control the length of the array being swapped
	 */
	if ((((uintptr_t)basep & (sizeof (uint64_t) - 1)) == 0) &&
	    (rsiz == sizeof (uint64_t))) {
		loops = 1;
		swapf = (void (*)(char *, char *, size_t))swapp64;
	} else if ((((uintptr_t)basep & (sizeof (uint32_t) - 1)) == 0) &&
	    (rsiz == sizeof (uint32_t))) {
		loops = 1;
		swapf = (void (*)(char *, char *, size_t))swapp32;
	} else if ((((uintptr_t)basep & (sizeof (uint32_t) - 1)) == 0) &&
	    ((rsiz & (sizeof (uint32_t) - 1)) == 0)) {
		loops = rsiz / sizeof (int);
		swapf = (void (*)(char *, char *, size_t))swapi;
	} else {
		loops = rsiz;
		swapf = swapb;
	}

	/*
	 * qsort is a partitioning sort
	 *
	 * the stack is the bookkeeping mechanism to keep track of all
	 * the partitions.
	 *
	 * each sort pass takes one partition and sorts it into two partitions.
	 * at the top of the loop we simply take the partition on the top
	 * of the stack and sort it. See the comments at the bottom
	 * of the loop regarding which partitions to add in what order.
	 *
	 * initially put the whole partition on the stack
	 */
	sp = stack;
	sp->b_lim = (char *)basep;
	sp->nrec = nrec;
	sp++;
	while (sp > stack) {
		sp--;
		b_lim = sp->b_lim;
		nrec = sp->nrec;

		/*
		 * a linear insertion sort i faster than a qsort for
		 * very small number of records (THRESH_L)
		 *
		 * if number records < threshold use linear insertion sort
		 *
		 * this also handles the special case where the partition
		 * 0 or 1 records length.
		 */
		if (nrec < THRESH_L) {
			/*
			 * Linear insertion sort
			 */
			t_par = b_lim;
			for (i = 1; i < nrec; i++) {
				t_par += rsiz;
				b_par = t_par;
				while (b_par > b_lim) {
					b_par -= rsiz;
					if ((*cmp)(b_par, b_par + rsiz) <= 0) {
						break;
					}
					(*swapf)(b_par, b_par + rsiz, loops);
				}
			}

			/*
			 * a linear insertion sort will put all records
			 * in their final position and will not create
			 * subpartitions.
			 *
			 * therefore when the insertion sort is complete
			 * just go to the top of the loop and get the
			 * next partition to sort.
			 */
			continue;
		}

		/* quicksort */

		/*
		 * choose a pivot record
		 *
		 * Ideally the pivot record will divide the partition
		 * into two equal parts. however we have to balance the
		 * work involved in selecting the pivot record with the
		 * expected benefit.
		 *
		 * The choice of pivot record depends on the number of
		 * records in the partition
		 *
		 * for small partitions (nrec < THRESH_M3)
		 *   we just select the record in the middle of the partition
		 *
		 * if (nrec >= THRESH_M3 && nrec < THRESH_M9)
		 *   we select three values and choose the median of 3
		 *
		 * if (nrec >= THRESH_M9)
		 *   then we use an approximate median of 9
		 *   9 records are selected and grouped in 3 groups of 3
		 *   the median of each of these 3 groups is fed into another
		 *   median of 3 decision.
		 *
		 * Each median of 3 decision is 2 or 3 compares,
		 * so median of 9 costs between 8 and 12 compares.
		 *
		 * i is byte distance between two consecutive samples
		 * m2 will point to the pivot record
		 */
		if (nrec < THRESH_M3) {
			m2 = b_lim + (nrec / 2) * rsiz;
		} else if (nrec < THRESH_M9) {
			/* use median of 3 */
			i = ((nrec - 1) / 2) * rsiz;
			m2 = med3(b_lim, b_lim + i, b_lim + 2 * i);
		} else {
			/* approx median of 9 */
			i = ((nrec - 1) / 8) * rsiz;
			m1 = med3(b_lim, b_lim +  i, b_lim + 2 * i);
			m2 = med3(b_lim + 3 * i, b_lim + 4 * i, b_lim + 5 * i);
			m3 = med3(b_lim + 6 * i, b_lim + 7 * i, b_lim + 8 * i);
			m2 = med3(m1, m2, m3);
		}

		/*
		 * quick sort partitioning
		 *
		 * The partition limits are defined by bottom and top pointers
		 * b_lim and t_lim.
		 *
		 * qsort uses a fairly standard method of moving the
		 * partitioning pointers, b_par and t_par, to the middle of
		 * the partition and exchanging records that are in the
		 * wrong part of the partition.
		 *
		 * Two enhancements have been made to the basic algorithm.
		 * One for handling duplicate records and one to minimize
		 * the number of swaps.
		 *
		 * Two duplicate records pointers are (b_dup and t_dup) are
		 * initially set to b_lim and t_lim.  Each time a record
		 * whose sort key value is equal to the pivot record is found
		 * it will be swapped with the record pointed to by
		 * b_dup or t_dup and the duplicate pointer will be
		 * incremented toward the center.
		 * When partitioning is complete, all the duplicate records
		 * will have been collected at the upper and lower limits of
		 * the partition and can easily be moved adjacent to the
		 * pivot record.
		 *
		 * The second optimization is to minimize the number of swaps.
		 * The pointer m2 points to the pivot record.
		 * During partitioning, if m2 is ever equal to the partitioning
		 * pointers, b_par or t_par, then b_par or t_par just moves
		 * onto the next record without doing a compare.
		 * If as a result of duplicate record detection,
		 * b_dup or t_dup is ever equal to m2,
		 * then m2 is changed to point to the duplicate record and
		 * b_dup or t_dup is incremented with out swapping records.
		 *
		 * When partitioning is done, we may not have the same pivot
		 * record that we started with, but we will have one with
		 * an equal sort key.
		 */
		b_dup = b_par		= b_lim;
		t_dup = t_par = t_lim	= b_lim + rsiz * (nrec - 1);
		for (;;) {

			/* move bottom pointer up */
			for (; b_par <= t_par; b_par += rsiz) {
				if (b_par == m2) {
					continue;
				}
				cv = cmp(b_par, m2);
				if (cv > 0) {
					break;
				}
				if (cv == 0) {
					if (b_dup == m2) {
						m2 = b_par;
					} else if (b_dup != b_par) {
						(*swapf)(b_dup, b_par, loops);
					}
					b_dup += rsiz;
				}
			}

			/* move top pointer down */
			for (; b_par < t_par; t_par -= rsiz) {
				if (t_par == m2) {
					continue;
				}
				cv = cmp(t_par, m2);
				if (cv < 0) {
					break;
				}
				if (cv == 0) {
					if (t_dup == m2) {
						m2 = t_par;
					} else if (t_dup != t_par) {
						(*swapf)(t_dup, t_par, loops);
					}
					t_dup -= rsiz;
				}
			}

			/* break if we are done partitioning */
			if (b_par >= t_par) {
				break;
			}

			/* exchange records at upper and lower break points */
			(*swapf)(b_par, t_par, loops);
			b_par += rsiz;
			t_par -= rsiz;
		}

		/*
		 * partitioning is now complete
		 *
		 * there are two termination conditions from the partitioning
		 * loop above.  Either b_par or t_par have crossed or
		 * they are equal.
		 *
		 * we need to swap the pivot record to its final position
		 * m2 could be in either the upper or lower partitions
		 * or it could already be in its final position
		 */
		/*
		 * R[b_par] > R[m2]
		 * R[t_par] < R[m2]
		 */
		if (t_par < b_par) {
			if (m2 < t_par) {
				(*swapf)(m2, t_par, loops);
				m2 = b_par = t_par;
			} else if (m2 > b_par) {
				(*swapf)(m2, b_par, loops);
				m2 = t_par = b_par;
			} else {
				b_par = t_par = m2;
			}
		} else {
			if (m2 < t_par) {
				t_par = b_par = t_par - rsiz;
			}
			if (m2 != b_par) {
				(*swapf)(m2, b_par, loops);
			}
			m2 = t_par;
		}

		/*
		 * move bottom duplicates next to pivot
		 * optimized to eliminate overlap
		 */
		d_bytelength = b_dup - b_lim;
		if (b_par - b_dup < d_bytelength) {
			b_dup = b_lim + (b_par - b_dup);
		}
		while (b_dup > b_lim) {
			b_dup -= rsiz;
			b_par -= rsiz;
			(*swapf)(b_dup, b_par, loops);
		}
		b_par = m2 - d_bytelength;

		/*
		 * move top duplicates next to pivot
		 */
		d_bytelength = t_lim - t_dup;
		if (t_dup - t_par < d_bytelength) {
			t_dup = t_lim - (t_dup - t_par);
		}
		while (t_dup < t_lim) {
			t_dup += rsiz;
			t_par += rsiz;
			(*swapf)(t_dup, t_par, loops);
		}
		t_par = m2 + d_bytelength;

		/*
		 * when a qsort pass completes there are three partitions
		 * 1) the lower contains all records less than pivot
		 * 2) the upper contains all records greater than pivot
		 * 3) the pivot partition contains all record equal to pivot
		 *
		 * all records in the pivot partition are in their final
		 * position and do not need to be accounted for by the stack
		 *
		 * when adding partitions to the stack
		 * it is important to add the largest partition first
		 * to prevent stack overflow.
		 *
		 * calculate number of unsorted records in top and bottom
		 * push resulting partitions on stack
		 */
		b_nrec = (b_par - b_lim) / rsiz;
		t_nrec = (t_lim - t_par) / rsiz;
		if (b_nrec < t_nrec) {
			sp->b_lim = t_par + rsiz;
			sp->nrec = t_nrec;
			sp++;
			sp->b_lim = b_lim;
			sp->nrec = b_nrec;
			sp++;
		} else {
			sp->b_lim = b_lim;
			sp->nrec = b_nrec;
			sp++;
			sp->b_lim = t_par + rsiz;
			sp->nrec = t_nrec;
			sp++;
		}
	}
}

/*
 * The following swap functions should not create a stack frame
 * the SPARC call / return instruction will be executed
 * but the a save / restore will not be executed
 * which means we won't do a window turn with the spill / fill overhead
 * verify this by examining the assembly code
 */

/* ARGSUSED */
static void
swapp32(uint32_t *r1, uint32_t *r2, size_t cnt)
{
	uint32_t temp;

	temp = *r1;
	*r1++ = *r2;
	*r2++ = temp;
}

/* ARGSUSED */
static void
swapp64(uint64_t *r1, uint64_t *r2, size_t cnt)
{
	uint64_t temp;

	temp = *r1;
	*r1++ = *r2;
	*r2++ = temp;
}

static void
swapi(uint32_t *r1, uint32_t *r2, size_t cnt)
{
	uint32_t temp;

	/* character by character */
	while (cnt--) {
		temp = *r1;
		*r1++ = *r2;
		*r2++ = temp;
	}
}

static void
swapb(char *r1, char *r2, size_t cnt)
{
	char	temp;

	/* character by character */
	while (cnt--) {
		temp = *r1;
		*r1++ = *r2;
		*r2++ = temp;
	}
}