xref: /dragonfly/lib/libc/stdlib/merge.c (revision 0de61e28)
1 /*-
2  * Copyright (c) 1992, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Peter McIlroy.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  * @(#)merge.c	8.2 (Berkeley) 2/14/94
33  * $FreeBSD: src/lib/libc/stdlib/merge.c,v 1.8 2007/01/09 00:28:10 imp Exp $
34  */
35 
36 /*
37  * Hybrid exponential search/linear search merge sort with hybrid
38  * natural/pairwise first pass.  Requires about .3% more comparisons
39  * for random data than LSMS with pairwise first pass alone.
40  * It works for objects as small as two bytes.
41  */
42 
43 #define NATURAL
44 #define THRESHOLD 16	/* Best choice for natural merge cut-off. */
45 
46 /* #define NATURAL to get hybrid natural merge.
47  * (The default is pairwise merging.)
48  */
49 
50 #include <sys/param.h>
51 
52 #include <errno.h>
53 #include <stdlib.h>
54 #include <string.h>
55 
56 static void setup(u_char *, u_char *, size_t, size_t,
57     int (*)(const void *, const void *));
58 static void insertionsort(u_char *, size_t, size_t,
59     int (*)(const void *, const void *));
60 
61 #define ISIZE sizeof(int)
62 #define PSIZE sizeof(u_char *)
63 #define ICOPY_LIST(src, dst, last)				\
64 	do							\
65 	*(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE;	\
66 	while(src < last)
67 #define ICOPY_ELT(src, dst, i)					\
68 	do							\
69 	*(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE;	\
70 	while (i -= ISIZE)
71 
72 #define CCOPY_LIST(src, dst, last)		\
73 	do					\
74 		*dst++ = *src++;		\
75 	while (src < last)
76 #define CCOPY_ELT(src, dst, i)			\
77 	do					\
78 		*dst++ = *src++;		\
79 	while (i -= 1)
80 
81 /*
82  * Find the next possible pointer head.  (Trickery for forcing an array
83  * to do double duty as a linked list when objects do not align with word
84  * boundaries.
85  */
86 /* Assumption: PSIZE is a power of 2. */
87 #define EVAL(p) (u_char **)						\
88 	((u_char *)0 +							\
89 	    (rounddown2((u_char *)p + PSIZE - 1 - (u_char *)0, PSIZE)))
90 
91 /*
92  * Arguments are as for qsort.
93  */
94 int
95 mergesort(void *base, size_t nmemb, size_t size,
96 	  int (*cmp)(const void *, const void *))
97 {
98 	size_t i;
99 	int sense;
100 	int big, iflag;
101 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
102 	u_char *list2, *list1, *p2, *p, *last, **p1;
103 
104 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
105 		errno = EINVAL;
106 		return (-1);
107 	}
108 
109 	if (nmemb == 0)
110 		return (0);
111 
112 	/*
113 	 * XXX
114 	 * Stupid subtraction for the Cray.
115 	 */
116 	iflag = 0;
117 	if (!(size % ISIZE) && !(((char *)base - (char *)0) % ISIZE))
118 		iflag = 1;
119 
120 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
121 		return (-1);
122 
123 	list1 = base;
124 	setup(list1, list2, nmemb, size, cmp);
125 	last = list2 + nmemb * size;
126 	i = big = 0;
127 	while (*EVAL(list2) != last) {
128 	    l2 = list1;
129 	    p1 = EVAL(list1);
130 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
131 	    	p2 = *EVAL(p2);
132 	    	f1 = l2;
133 	    	f2 = l1 = list1 + (p2 - list2);
134 	    	if (p2 != last)
135 	    		p2 = *EVAL(p2);
136 	    	l2 = list1 + (p2 - list2);
137 	    	while (f1 < l1 && f2 < l2) {
138 	    		if ((*cmp)(f1, f2) <= 0) {
139 	    			q = f2;
140 	    			b = f1, t = l1;
141 	    			sense = -1;
142 	    		} else {
143 	    			q = f1;
144 	    			b = f2, t = l2;
145 	    			sense = 0;
146 	    		}
147 	    		if (!big) {	/* here i = 0 */
148 				while ((b += size) < t && cmp(q, b) >sense)
149 	    				if (++i == 6) {
150 	    					big = 1;
151 	    					goto EXPONENTIAL;
152 	    				}
153 	    		} else {
154 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
155 	    				if ((p = (b + i)) >= t) {
156 	    					if ((p = t - size) > b &&
157 						    (*cmp)(q, p) <= sense)
158 	    						t = p;
159 	    					else
160 	    						b = p;
161 	    					break;
162 	    				} else if ((*cmp)(q, p) <= sense) {
163 	    					t = p;
164 	    					if (i == size)
165 	    						big = 0;
166 	    					goto FASTCASE;
167 	    				} else
168 	    					b = p;
169 				while (t > b+size) {
170 	    				i = (((t - b) / size) >> 1) * size;
171 	    				if ((*cmp)(q, p = b + i) <= sense)
172 	    					t = p;
173 	    				else
174 	    					b = p;
175 	    			}
176 	    			goto COPY;
177 FASTCASE:	    		while (i > size)
178 	    				if ((*cmp)(q,
179 	    					p = b + (i >>= 1)) <= sense)
180 	    					t = p;
181 	    				else
182 	    					b = p;
183 COPY:	    			b = t;
184 	    		}
185 	    		i = size;
186 	    		if (q == f1) {
187 	    			if (iflag) {
188 	    				ICOPY_LIST(f2, tp2, b);
189 	    				ICOPY_ELT(f1, tp2, i);
190 	    			} else {
191 	    				CCOPY_LIST(f2, tp2, b);
192 	    				CCOPY_ELT(f1, tp2, i);
193 	    			}
194 	    		} else {
195 	    			if (iflag) {
196 	    				ICOPY_LIST(f1, tp2, b);
197 	    				ICOPY_ELT(f2, tp2, i);
198 	    			} else {
199 	    				CCOPY_LIST(f1, tp2, b);
200 	    				CCOPY_ELT(f2, tp2, i);
201 	    			}
202 	    		}
203 	    	}
204 	    	if (f2 < l2) {
205 	    		if (iflag)
206 	    			ICOPY_LIST(f2, tp2, l2);
207 	    		else
208 	    			CCOPY_LIST(f2, tp2, l2);
209 	    	} else if (f1 < l1) {
210 	    		if (iflag)
211 	    			ICOPY_LIST(f1, tp2, l1);
212 	    		else
213 	    			CCOPY_LIST(f1, tp2, l1);
214 	    	}
215 	    	*p1 = l2;
216 	    }
217 	    tp2 = list1;	/* swap list1, list2 */
218 	    list1 = list2;
219 	    list2 = tp2;
220 	    last = list2 + nmemb*size;
221 	}
222 	if (base == list2) {
223 		memmove(list2, list1, nmemb*size);
224 		list2 = list1;
225 	}
226 	free(list2);
227 	return (0);
228 }
229 
230 #define	swap(a, b) {					\
231 		s = b;					\
232 		i = size;				\
233 		do {					\
234 			tmp = *a; *a++ = *s; *s++ = tmp; \
235 		} while (--i);				\
236 		a -= size;				\
237 	}
238 #define reverse(bot, top) {				\
239 	s = top;					\
240 	do {						\
241 		i = size;				\
242 		do {					\
243 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
244 		} while (--i);				\
245 		s -= size2;				\
246 	} while(bot < s);				\
247 }
248 
249 /*
250  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
251  * increasing order, list2 in a corresponding linked list.  Checks for runs
252  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
253  * is defined.  Otherwise simple pairwise merging is used.)
254  */
255 static void
256 setup(u_char *list1, u_char *list2, size_t n, size_t size,
257       int (*cmp)(const void *, const void *))
258 {
259 	int i, length, size2, tmp, sense;
260 	u_char *f1, *f2, *s, *l2, *last, *p2;
261 
262 	size2 = size*2;
263 	if (n <= 5) {
264 		insertionsort(list1, n, size, cmp);
265 		*EVAL(list2) = (u_char*) list2 + n*size;
266 		return;
267 	}
268 	/*
269 	 * Avoid running pointers out of bounds; limit n to evens
270 	 * for simplicity.
271 	 */
272 	i = 4 + (n & 1);
273 	insertionsort(list1 + (n - i) * size, i, size, cmp);
274 	last = list1 + size * (n - i);
275 	*EVAL(list2 + (last - list1)) = list2 + n * size;
276 
277 #ifdef NATURAL
278 	p2 = list2;
279 	f1 = list1;
280 	sense = (cmp(f1, f1 + size) > 0);
281 	for (; f1 < last; sense = !sense) {
282 		length = 2;
283 					/* Find pairs with same sense. */
284 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
285 			if ((cmp(f2, f2+ size) > 0) != sense)
286 				break;
287 			length += 2;
288 		}
289 		if (length < THRESHOLD) {		/* Pairwise merge */
290 			do {
291 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
292 				if (sense > 0)
293 					swap (f1, f1 + size);
294 			} while ((f1 += size2) < f2);
295 		} else {				/* Natural merge */
296 			l2 = f2;
297 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
298 				if ((cmp(f2-size, f2) > 0) != sense) {
299 					p2 = *EVAL(p2) = f2 - list1 + list2;
300 					if (sense > 0)
301 						reverse(f1, f2-size);
302 					f1 = f2;
303 				}
304 			}
305 			if (sense > 0)
306 				reverse (f1, f2-size);
307 			f1 = f2;
308 			if (f2 < last || cmp(f2 - size, f2) > 0)
309 				p2 = *EVAL(p2) = f2 - list1 + list2;
310 			else
311 				p2 = *EVAL(p2) = list2 + n*size;
312 		}
313 	}
314 #else		/* pairwise merge only. */
315 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
316 		p2 = *EVAL(p2) = p2 + size2;
317 		if (cmp (f1, f1 + size) > 0)
318 			swap(f1, f1 + size);
319 	}
320 #endif /* NATURAL */
321 }
322 
323 /*
324  * This is to avoid out-of-bounds addresses in sorting the
325  * last 4 elements.
326  */
327 static void
328 insertionsort(u_char *a, size_t n, size_t size,
329 	      int (*cmp)(const void *, const void *))
330 {
331 	u_char *ai, *s, *t, *u, tmp;
332 	int i;
333 
334 	for (ai = a+size; --n >= 1; ai += size)
335 		for (t = ai; t > a; t -= size) {
336 			u = t - size;
337 			if (cmp(u, t) <= 0)
338 				break;
339 			swap(u, t);
340 		}
341 }
342