xref: /freebsd/lib/libc/stdlib/merge.c (revision 6b129086dcee14496517fae085b448e3edc69bc7)
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 
33 #if defined(LIBC_SCCS) && !defined(lint)
34 static char sccsid[] = "@(#)merge.c	8.2 (Berkeley) 2/14/94";
35 #endif /* LIBC_SCCS and not lint */
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 /*
40  * Hybrid exponential search/linear search merge sort with hybrid
41  * natural/pairwise first pass.  Requires about .3% more comparisons
42  * for random data than LSMS with pairwise first pass alone.
43  * It works for objects as small as two bytes.
44  */
45 
46 #define NATURAL
47 #define THRESHOLD 16	/* Best choice for natural merge cut-off. */
48 
49 /* #define NATURAL to get hybrid natural merge.
50  * (The default is pairwise merging.)
51  */
52 
53 #include <sys/types.h>
54 
55 #include <errno.h>
56 #include <stdlib.h>
57 #include <string.h>
58 
59 #ifdef I_AM_MERGESORT_B
60 #include "block_abi.h"
61 #define	DECLARE_CMP	DECLARE_BLOCK(int, cmp, const void *, const void *)
62 typedef DECLARE_BLOCK(int, cmp_t, const void *, const void *);
63 #define	CMP(x, y)	CALL_BLOCK(cmp, x, y)
64 #else
65 typedef int (*cmp_t)(const void *, const void *);
66 #define	CMP(x, y)	cmp(x, y)
67 #endif
68 
69 static void setup(u_char *, u_char *, size_t, size_t, cmp_t);
70 static void insertionsort(u_char *, size_t, size_t, cmp_t);
71 
72 #define ISIZE sizeof(int)
73 #define PSIZE sizeof(u_char *)
74 #define ICOPY_LIST(src, dst, last)				\
75 	do							\
76 	*(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE;	\
77 	while(src < last)
78 #define ICOPY_ELT(src, dst, i)					\
79 	do							\
80 	*(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE;	\
81 	while (i -= ISIZE)
82 
83 #define CCOPY_LIST(src, dst, last)		\
84 	do					\
85 		*dst++ = *src++;		\
86 	while (src < last)
87 #define CCOPY_ELT(src, dst, i)			\
88 	do					\
89 		*dst++ = *src++;		\
90 	while (i -= 1)
91 
92 /*
93  * Find the next possible pointer head.  (Trickery for forcing an array
94  * to do double duty as a linked list when objects do not align with word
95  * boundaries.
96  */
97 /* Assumption: PSIZE is a power of 2. */
98 #define EVAL(p) (u_char **)						\
99 	((u_char *)0 +							\
100 	    (((u_char *)p + PSIZE - 1 - (u_char *) 0) & ~(PSIZE - 1)))
101 
102 /*
103  * Arguments are as for qsort.
104  */
105 int
106 #ifdef I_AM_MERGESORT_B
107 mergesort_b(base, nmemb, size, cmp)
108 #else
109 mergesort(base, nmemb, size, cmp)
110 #endif
111 	void *base;
112 	size_t nmemb;
113 	size_t size;
114 	cmp_t cmp;
115 {
116 	size_t i;
117 	int sense;
118 	int big, iflag;
119 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
120 	u_char *list2, *list1, *p2, *p, *last, **p1;
121 
122 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
123 		errno = EINVAL;
124 		return (-1);
125 	}
126 
127 	if (nmemb == 0)
128 		return (0);
129 
130 	/*
131 	 * XXX
132 	 * Stupid subtraction for the Cray.
133 	 */
134 	iflag = 0;
135 	if (!(size % ISIZE) && !(((char *)base - (char *)0) % ISIZE))
136 		iflag = 1;
137 
138 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
139 		return (-1);
140 
141 	list1 = base;
142 	setup(list1, list2, nmemb, size, cmp);
143 	last = list2 + nmemb * size;
144 	i = big = 0;
145 	while (*EVAL(list2) != last) {
146 	    l2 = list1;
147 	    p1 = EVAL(list1);
148 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
149 	    	p2 = *EVAL(p2);
150 	    	f1 = l2;
151 	    	f2 = l1 = list1 + (p2 - list2);
152 	    	if (p2 != last)
153 	    		p2 = *EVAL(p2);
154 	    	l2 = list1 + (p2 - list2);
155 	    	while (f1 < l1 && f2 < l2) {
156 	    		if (CMP(f1, f2) <= 0) {
157 	    			q = f2;
158 	    			b = f1, t = l1;
159 	    			sense = -1;
160 	    		} else {
161 	    			q = f1;
162 	    			b = f2, t = l2;
163 	    			sense = 0;
164 	    		}
165 	    		if (!big) {	/* here i = 0 */
166 				while ((b += size) < t && CMP(q, b) >sense)
167 	    				if (++i == 6) {
168 	    					big = 1;
169 	    					goto EXPONENTIAL;
170 	    				}
171 	    		} else {
172 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
173 	    				if ((p = (b + i)) >= t) {
174 	    					if ((p = t - size) > b &&
175 						    CMP(q, p) <= sense)
176 	    						t = p;
177 	    					else
178 	    						b = p;
179 	    					break;
180 	    				} else if (CMP(q, p) <= sense) {
181 	    					t = p;
182 	    					if (i == size)
183 	    						big = 0;
184 	    					goto FASTCASE;
185 	    				} else
186 	    					b = p;
187 				while (t > b+size) {
188 	    				i = (((t - b) / size) >> 1) * size;
189 	    				if (CMP(q, p = b + i) <= sense)
190 	    					t = p;
191 	    				else
192 	    					b = p;
193 	    			}
194 	    			goto COPY;
195 FASTCASE:	    		while (i > size)
196 	    				if (CMP(q,
197 	    					p = b + (i >>= 1)) <= sense)
198 	    					t = p;
199 	    				else
200 	    					b = p;
201 COPY:	    			b = t;
202 	    		}
203 	    		i = size;
204 	    		if (q == f1) {
205 	    			if (iflag) {
206 	    				ICOPY_LIST(f2, tp2, b);
207 	    				ICOPY_ELT(f1, tp2, i);
208 	    			} else {
209 	    				CCOPY_LIST(f2, tp2, b);
210 	    				CCOPY_ELT(f1, tp2, i);
211 	    			}
212 	    		} else {
213 	    			if (iflag) {
214 	    				ICOPY_LIST(f1, tp2, b);
215 	    				ICOPY_ELT(f2, tp2, i);
216 	    			} else {
217 	    				CCOPY_LIST(f1, tp2, b);
218 	    				CCOPY_ELT(f2, tp2, i);
219 	    			}
220 	    		}
221 	    	}
222 	    	if (f2 < l2) {
223 	    		if (iflag)
224 	    			ICOPY_LIST(f2, tp2, l2);
225 	    		else
226 	    			CCOPY_LIST(f2, tp2, l2);
227 	    	} else if (f1 < l1) {
228 	    		if (iflag)
229 	    			ICOPY_LIST(f1, tp2, l1);
230 	    		else
231 	    			CCOPY_LIST(f1, tp2, l1);
232 	    	}
233 	    	*p1 = l2;
234 	    }
235 	    tp2 = list1;	/* swap list1, list2 */
236 	    list1 = list2;
237 	    list2 = tp2;
238 	    last = list2 + nmemb*size;
239 	}
240 	if (base == list2) {
241 		memmove(list2, list1, nmemb*size);
242 		list2 = list1;
243 	}
244 	free(list2);
245 	return (0);
246 }
247 
248 #define	swap(a, b) {					\
249 		s = b;					\
250 		i = size;				\
251 		do {					\
252 			tmp = *a; *a++ = *s; *s++ = tmp; \
253 		} while (--i);				\
254 		a -= size;				\
255 	}
256 #define reverse(bot, top) {				\
257 	s = top;					\
258 	do {						\
259 		i = size;				\
260 		do {					\
261 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
262 		} while (--i);				\
263 		s -= size2;				\
264 	} while(bot < s);				\
265 }
266 
267 /*
268  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
269  * increasing order, list2 in a corresponding linked list.  Checks for runs
270  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
271  * is defined.  Otherwise simple pairwise merging is used.)
272  */
273 void
274 setup(list1, list2, n, size, cmp)
275 	size_t n, size;
276 	u_char *list1, *list2;
277 	cmp_t cmp;
278 {
279 	int i, length, size2, tmp, sense;
280 	u_char *f1, *f2, *s, *l2, *last, *p2;
281 
282 	size2 = size*2;
283 	if (n <= 5) {
284 		insertionsort(list1, n, size, cmp);
285 		*EVAL(list2) = (u_char*) list2 + n*size;
286 		return;
287 	}
288 	/*
289 	 * Avoid running pointers out of bounds; limit n to evens
290 	 * for simplicity.
291 	 */
292 	i = 4 + (n & 1);
293 	insertionsort(list1 + (n - i) * size, i, size, cmp);
294 	last = list1 + size * (n - i);
295 	*EVAL(list2 + (last - list1)) = list2 + n * size;
296 
297 #ifdef NATURAL
298 	p2 = list2;
299 	f1 = list1;
300 	sense = (CMP(f1, f1 + size) > 0);
301 	for (; f1 < last; sense = !sense) {
302 		length = 2;
303 					/* Find pairs with same sense. */
304 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
305 			if ((CMP(f2, f2+ size) > 0) != sense)
306 				break;
307 			length += 2;
308 		}
309 		if (length < THRESHOLD) {		/* Pairwise merge */
310 			do {
311 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
312 				if (sense > 0)
313 					swap (f1, f1 + size);
314 			} while ((f1 += size2) < f2);
315 		} else {				/* Natural merge */
316 			l2 = f2;
317 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
318 				if ((CMP(f2-size, f2) > 0) != sense) {
319 					p2 = *EVAL(p2) = f2 - list1 + list2;
320 					if (sense > 0)
321 						reverse(f1, f2-size);
322 					f1 = f2;
323 				}
324 			}
325 			if (sense > 0)
326 				reverse (f1, f2-size);
327 			f1 = f2;
328 			if (f2 < last || CMP(f2 - size, f2) > 0)
329 				p2 = *EVAL(p2) = f2 - list1 + list2;
330 			else
331 				p2 = *EVAL(p2) = list2 + n*size;
332 		}
333 	}
334 #else		/* pairwise merge only. */
335 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
336 		p2 = *EVAL(p2) = p2 + size2;
337 		if (CMP (f1, f1 + size) > 0)
338 			swap(f1, f1 + size);
339 	}
340 #endif /* NATURAL */
341 }
342 
343 /*
344  * This is to avoid out-of-bounds addresses in sorting the
345  * last 4 elements.
346  */
347 static void
348 insertionsort(a, n, size, cmp)
349 	u_char *a;
350 	size_t n, size;
351 	cmp_t cmp;
352 {
353 	u_char *ai, *s, *t, *u, tmp;
354 	int i;
355 
356 	for (ai = a+size; --n >= 1; ai += size)
357 		for (t = ai; t > a; t -= size) {
358 			u = t - size;
359 			if (CMP(u, t) <= 0)
360 				break;
361 			swap(u, t);
362 		}
363 }
364