Home | History | Annotate | Line # | Download | only in stdlib
merge.c revision 1.8
      1 /*	$NetBSD: merge.c,v 1.8 1999/09/16 11:45:35 lukem Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1992, 1993
      5  *	The Regents of the University of California.  All rights reserved.
      6  *
      7  * This code is derived from software contributed to Berkeley by
      8  * Peter McIlroy.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  * 3. All advertising materials mentioning features or use of this software
     19  *    must display the following acknowledgement:
     20  *	This product includes software developed by the University of
     21  *	California, Berkeley and its contributors.
     22  * 4. Neither the name of the University nor the names of its contributors
     23  *    may be used to endorse or promote products derived from this software
     24  *    without specific prior written permission.
     25  *
     26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     36  * SUCH DAMAGE.
     37  */
     38 
     39 #include <sys/cdefs.h>
     40 #if defined(LIBC_SCCS) && !defined(lint)
     41 #if 0
     42 static char sccsid[] = "from: @(#)merge.c	8.2 (Berkeley) 2/14/94";
     43 #else
     44 __RCSID("$NetBSD: merge.c,v 1.8 1999/09/16 11:45:35 lukem Exp $");
     45 #endif
     46 #endif /* LIBC_SCCS and not lint */
     47 
     48 /*
     49  * Hybrid exponential search/linear search merge sort with hybrid
     50  * natural/pairwise first pass.  Requires about .3% more comparisons
     51  * for random data than LSMS with pairwise first pass alone.
     52  * It works for objects as small as two bytes.
     53  */
     54 
     55 #define NATURAL
     56 #define THRESHOLD 16	/* Best choice for natural merge cut-off. */
     57 
     58 /* #define NATURAL to get hybrid natural merge.
     59  * (The default is pairwise merging.)
     60  */
     61 
     62 #include "namespace.h"
     63 #include <sys/types.h>
     64 
     65 #include <assert.h>
     66 #include <errno.h>
     67 #include <stdlib.h>
     68 #include <string.h>
     69 
     70 #ifdef __weak_alias
     71 __weak_alias(mergesort,_mergesort);
     72 #endif
     73 
     74 static void setup __P((u_char *, u_char *, size_t, size_t,
     75     int (*)(const void *, const void *)));
     76 static void insertionsort __P((u_char *, size_t, size_t,
     77     int (*)(const void *, const void *)));
     78 
     79 #define ISIZE sizeof(int)
     80 #define PSIZE sizeof(u_char *)
     81 #define ICOPY_LIST(src, dst, last)				\
     82 	do							\
     83 	*(int*)(void *)dst = *(int*)(void *)src,		\
     84 		src += ISIZE, dst += ISIZE;			\
     85 	while(src < last)
     86 #define ICOPY_ELT(src, dst, i)					\
     87 	do							\
     88 	*(int*)(void *)dst = *(int*)(void *)src,		\
     89 	src += ISIZE, dst += ISIZE;				\
     90 	while (i -= ISIZE)
     91 
     92 #define CCOPY_LIST(src, dst, last)		\
     93 	do					\
     94 		*dst++ = *src++;		\
     95 	while (src < last)
     96 #define CCOPY_ELT(src, dst, i)			\
     97 	do					\
     98 		*dst++ = *src++;		\
     99 	while (i -= 1)
    100 
    101 /*
    102  * Find the next possible pointer head.  (Trickery for forcing an array
    103  * to do double duty as a linked list when objects do not align with word
    104  * boundaries.
    105  */
    106 /* Assumption: PSIZE is a power of 2. */
    107 #define EVAL(p) ((u_char **)(void *)					\
    108     ((u_char *)0 +							\
    109     (((u_char *)(void *)(p) + PSIZE - 1 - (u_char *) 0) & ~(PSIZE - 1))))
    110 
    111 /*
    112  * Arguments are as for qsort.
    113  */
    114 int
    115 mergesort(base, nmemb, size, cmp)
    116 	void *base;
    117 	size_t nmemb;
    118 	size_t size;
    119 	int (*cmp) __P((const void *, const void *));
    120 {
    121 	int i, sense;
    122 	int big, iflag;
    123 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
    124 	u_char *list2, *list1, *p2, *p, *last, **p1;
    125 
    126 	_DIAGASSERT(base != NULL);
    127 	_DIAGASSERT(cmp != NULL);
    128 #ifdef _DIAGNOSTIC
    129 	if (base == NULL || cmp == NULL) {
    130 		errno = EFAULT;
    131 		return (-1);
    132 	}
    133 #endif
    134 
    135 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
    136 		errno = EINVAL;
    137 		return (-1);
    138 	}
    139 
    140 	/*
    141 	 * XXX
    142 	 * Stupid subtraction for the Cray.
    143 	 */
    144 	iflag = 0;
    145 	if (!(size % ISIZE) && !(((char *)base - (char *)0) % ISIZE))
    146 		iflag = 1;
    147 
    148 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
    149 		return (-1);
    150 
    151 	list1 = base;
    152 	setup(list1, list2, nmemb, size, cmp);
    153 	last = list2 + nmemb * size;
    154 	i = big = 0;
    155 	while (*EVAL(list2) != last) {
    156 	    l2 = list1;
    157 	    p1 = EVAL(list1);
    158 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
    159 	    	p2 = *EVAL(p2);
    160 	    	f1 = l2;
    161 	    	f2 = l1 = list1 + (p2 - list2);
    162 	    	if (p2 != last)
    163 	    		p2 = *EVAL(p2);
    164 	    	l2 = list1 + (p2 - list2);
    165 	    	while (f1 < l1 && f2 < l2) {
    166 	    		if ((*cmp)(f1, f2) <= 0) {
    167 	    			q = f2;
    168 	    			b = f1, t = l1;
    169 	    			sense = -1;
    170 	    		} else {
    171 	    			q = f1;
    172 	    			b = f2, t = l2;
    173 	    			sense = 0;
    174 	    		}
    175 	    		if (!big) {	/* here i = 0 */
    176 #if 0
    177 LINEAR:
    178 #endif
    179 				while ((b += size) < t && cmp(q, b) >sense)
    180 	    				if (++i == 6) {
    181 	    					big = 1;
    182 	    					goto EXPONENTIAL;
    183 	    				}
    184 	    		} else {
    185 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
    186 	    				if ((p = (b + i)) >= t) {
    187 	    					if ((p = t - size) > b &&
    188 						    (*cmp)(q, p) <= sense)
    189 	    						t = p;
    190 	    					else
    191 	    						b = p;
    192 	    					break;
    193 	    				} else if ((*cmp)(q, p) <= sense) {
    194 	    					t = p;
    195 	    					if (i == size)
    196 	    						big = 0;
    197 	    					goto FASTCASE;
    198 	    				} else
    199 	    					b = p;
    200 #if 0
    201 SLOWCASE:
    202 #endif
    203 				while (t > b+size) {
    204 	    				i = (((t - b) / size) >> 1) * size;
    205 	    				if ((*cmp)(q, p = b + i) <= sense)
    206 	    					t = p;
    207 	    				else
    208 	    					b = p;
    209 	    			}
    210 	    			goto COPY;
    211 FASTCASE:	    		while (i > size)
    212 	    				if ((*cmp)(q,
    213 	    					p = b +
    214 						    (i = (unsigned int) i >> 1)) <= sense)
    215 	    					t = p;
    216 	    				else
    217 	    					b = p;
    218 COPY:	    			b = t;
    219 	    		}
    220 	    		i = size;
    221 	    		if (q == f1) {
    222 	    			if (iflag) {
    223 	    				ICOPY_LIST(f2, tp2, b);
    224 	    				ICOPY_ELT(f1, tp2, i);
    225 	    			} else {
    226 	    				CCOPY_LIST(f2, tp2, b);
    227 	    				CCOPY_ELT(f1, tp2, i);
    228 	    			}
    229 	    		} else {
    230 	    			if (iflag) {
    231 	    				ICOPY_LIST(f1, tp2, b);
    232 	    				ICOPY_ELT(f2, tp2, i);
    233 	    			} else {
    234 	    				CCOPY_LIST(f1, tp2, b);
    235 	    				CCOPY_ELT(f2, tp2, i);
    236 	    			}
    237 	    		}
    238 	    	}
    239 	    	if (f2 < l2) {
    240 	    		if (iflag)
    241 	    			ICOPY_LIST(f2, tp2, l2);
    242 	    		else
    243 	    			CCOPY_LIST(f2, tp2, l2);
    244 	    	} else if (f1 < l1) {
    245 	    		if (iflag)
    246 	    			ICOPY_LIST(f1, tp2, l1);
    247 	    		else
    248 	    			CCOPY_LIST(f1, tp2, l1);
    249 	    	}
    250 	    	*p1 = l2;
    251 	    }
    252 	    tp2 = list1;	/* swap list1, list2 */
    253 	    list1 = list2;
    254 	    list2 = tp2;
    255 	    last = list2 + nmemb*size;
    256 	}
    257 	if (base == list2) {
    258 		memmove(list2, list1, nmemb*size);
    259 		list2 = list1;
    260 	}
    261 	free(list2);
    262 	return (0);
    263 }
    264 
    265 #define	swap(a, b) {					\
    266 		s = b;					\
    267 		i = size;				\
    268 		do {					\
    269 			tmp = *a; *a++ = *s; *s++ = tmp; \
    270 		} while (--i);				\
    271 		a -= size;				\
    272 	}
    273 #define reverse(bot, top) {				\
    274 	s = top;					\
    275 	do {						\
    276 		i = size;				\
    277 		do {					\
    278 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
    279 		} while (--i);				\
    280 		s -= size2;				\
    281 	} while(bot < s);				\
    282 }
    283 
    284 /*
    285  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
    286  * increasing order, list2 in a corresponding linked list.  Checks for runs
    287  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
    288  * is defined.  Otherwise simple pairwise merging is used.)
    289  */
    290 
    291 /* XXX: shouldn't this function be static? - lukem 990810 */
    292 void
    293 setup(list1, list2, n, size, cmp)
    294 	size_t n, size;
    295 	int (*cmp) __P((const void *, const void *));
    296 	u_char *list1, *list2;
    297 {
    298 	int i, length, size2, tmp, sense;
    299 	u_char *f1, *f2, *s, *l2, *last, *p2;
    300 
    301 	_DIAGASSERT(cmp != NULL);
    302 	_DIAGASSERT(list1 != NULL);
    303 	_DIAGASSERT(list2 != NULL);
    304 
    305 	size2 = size*2;
    306 	if (n <= 5) {
    307 		insertionsort(list1, n, size, cmp);
    308 		*EVAL(list2) = list2 + n*size;
    309 		return;
    310 	}
    311 	/*
    312 	 * Avoid running pointers out of bounds; limit n to evens
    313 	 * for simplicity.
    314 	 */
    315 	i = 4 + (n & 1);
    316 	insertionsort(list1 + (n - i) * size, (size_t)i, size, cmp);
    317 	last = list1 + size * (n - i);
    318 	*EVAL(list2 + (last - list1)) = list2 + n * size;
    319 
    320 #ifdef NATURAL
    321 	p2 = list2;
    322 	f1 = list1;
    323 	sense = (cmp(f1, f1 + size) > 0);
    324 	for (; f1 < last; sense = !sense) {
    325 		length = 2;
    326 					/* Find pairs with same sense. */
    327 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
    328 			if ((cmp(f2, f2+ size) > 0) != sense)
    329 				break;
    330 			length += 2;
    331 		}
    332 		if (length < THRESHOLD) {		/* Pairwise merge */
    333 			do {
    334 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
    335 				if (sense > 0)
    336 					swap (f1, f1 + size);
    337 			} while ((f1 += size2) < f2);
    338 		} else {				/* Natural merge */
    339 			l2 = f2;
    340 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
    341 				if ((cmp(f2-size, f2) > 0) != sense) {
    342 					p2 = *EVAL(p2) = f2 - list1 + list2;
    343 					if (sense > 0)
    344 						reverse(f1, f2-size);
    345 					f1 = f2;
    346 				}
    347 			}
    348 			if (sense > 0)
    349 				reverse (f1, f2-size);
    350 			f1 = f2;
    351 			if (f2 < last || cmp(f2 - size, f2) > 0)
    352 				p2 = *EVAL(p2) = f2 - list1 + list2;
    353 			else
    354 				p2 = *EVAL(p2) = list2 + n*size;
    355 		}
    356 	}
    357 #else		/* pairwise merge only. */
    358 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
    359 		p2 = *EVAL(p2) = p2 + size2;
    360 		if (cmp (f1, f1 + size) > 0)
    361 			swap(f1, f1 + size);
    362 	}
    363 #endif /* NATURAL */
    364 }
    365 
    366 /*
    367  * This is to avoid out-of-bounds addresses in sorting the
    368  * last 4 elements.
    369  */
    370 static void
    371 insertionsort(a, n, size, cmp)
    372 	u_char *a;
    373 	size_t n, size;
    374 	int (*cmp) __P((const void *, const void *));
    375 {
    376 	u_char *ai, *s, *t, *u, tmp;
    377 	int i;
    378 
    379 	_DIAGASSERT(a != NULL);
    380 	_DIAGASSERT(cmp != NULL);
    381 
    382 	for (ai = a+size; --n >= 1; ai += size)
    383 		for (t = ai; t > a; t -= size) {
    384 			u = t - size;
    385 			if (cmp(u, t) <= 0)
    386 				break;
    387 			swap(u, t);
    388 		}
    389 }
    390