xref: /freebsd/lib/libc/stdlib/merge.c (revision 734e82fe33aa764367791a7d603b383996c6b40b)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
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. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #if defined(LIBC_SCCS) && !defined(lint)
36 static char sccsid[] = "@(#)merge.c	8.2 (Berkeley) 2/14/94";
37 #endif /* LIBC_SCCS and not lint */
38 #include <sys/cdefs.h>
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/param.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 **)roundup2((uintptr_t)p, PSIZE)
99 
100 #ifdef I_AM_MERGESORT_B
101 int mergesort_b(void *, size_t, size_t, cmp_t);
102 #else
103 int mergesort(void *, size_t, size_t, cmp_t);
104 #endif
105 
106 /*
107  * Arguments are as for qsort.
108  */
109 int
110 #ifdef I_AM_MERGESORT_B
111 mergesort_b(void *base, size_t nmemb, size_t size, cmp_t cmp)
112 #else
113 mergesort(void *base, size_t nmemb, size_t size, cmp_t cmp)
114 #endif
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 	iflag = 0;
131 	if (__is_aligned(size, ISIZE) && __is_aligned(base, ISIZE))
132 		iflag = 1;
133 
134 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
135 		return (-1);
136 
137 	list1 = base;
138 	setup(list1, list2, nmemb, size, cmp);
139 	last = list2 + nmemb * size;
140 	i = big = 0;
141 	while (*EVAL(list2) != last) {
142 	    l2 = list1;
143 	    p1 = EVAL(list1);
144 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
145 	    	p2 = *EVAL(p2);
146 	    	f1 = l2;
147 	    	f2 = l1 = list1 + (p2 - list2);
148 	    	if (p2 != last)
149 	    		p2 = *EVAL(p2);
150 	    	l2 = list1 + (p2 - list2);
151 	    	while (f1 < l1 && f2 < l2) {
152 	    		if (CMP(f1, f2) <= 0) {
153 	    			q = f2;
154 	    			b = f1, t = l1;
155 	    			sense = -1;
156 	    		} else {
157 	    			q = f1;
158 	    			b = f2, t = l2;
159 	    			sense = 0;
160 	    		}
161 	    		if (!big) {	/* here i = 0 */
162 				while ((b += size) < t && CMP(q, b) >sense)
163 	    				if (++i == 6) {
164 	    					big = 1;
165 	    					goto EXPONENTIAL;
166 	    				}
167 	    		} else {
168 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
169 	    				if ((p = (b + i)) >= t) {
170 	    					if ((p = t - size) > b &&
171 						    CMP(q, p) <= sense)
172 	    						t = p;
173 	    					else
174 	    						b = p;
175 	    					break;
176 	    				} else if (CMP(q, p) <= sense) {
177 	    					t = p;
178 	    					if (i == size)
179 	    						big = 0;
180 	    					goto FASTCASE;
181 	    				} else
182 	    					b = p;
183 				while (t > b+size) {
184 	    				i = (((t - b) / size) >> 1) * size;
185 	    				if (CMP(q, p = b + i) <= sense)
186 	    					t = p;
187 	    				else
188 	    					b = p;
189 	    			}
190 	    			goto COPY;
191 FASTCASE:	    		while (i > size)
192 	    				if (CMP(q,
193 	    					p = b + (i >>= 1)) <= sense)
194 	    					t = p;
195 	    				else
196 	    					b = p;
197 COPY:	    			b = t;
198 	    		}
199 	    		i = size;
200 	    		if (q == f1) {
201 	    			if (iflag) {
202 	    				ICOPY_LIST(f2, tp2, b);
203 	    				ICOPY_ELT(f1, tp2, i);
204 	    			} else {
205 	    				CCOPY_LIST(f2, tp2, b);
206 	    				CCOPY_ELT(f1, tp2, i);
207 	    			}
208 	    		} else {
209 	    			if (iflag) {
210 	    				ICOPY_LIST(f1, tp2, b);
211 	    				ICOPY_ELT(f2, tp2, i);
212 	    			} else {
213 	    				CCOPY_LIST(f1, tp2, b);
214 	    				CCOPY_ELT(f2, tp2, i);
215 	    			}
216 	    		}
217 	    	}
218 	    	if (f2 < l2) {
219 	    		if (iflag)
220 	    			ICOPY_LIST(f2, tp2, l2);
221 	    		else
222 	    			CCOPY_LIST(f2, tp2, l2);
223 	    	} else if (f1 < l1) {
224 	    		if (iflag)
225 	    			ICOPY_LIST(f1, tp2, l1);
226 	    		else
227 	    			CCOPY_LIST(f1, tp2, l1);
228 	    	}
229 	    	*p1 = l2;
230 	    }
231 	    tp2 = list1;	/* swap list1, list2 */
232 	    list1 = list2;
233 	    list2 = tp2;
234 	    last = list2 + nmemb*size;
235 	}
236 	if (base == list2) {
237 		memmove(list2, list1, nmemb*size);
238 		list2 = list1;
239 	}
240 	free(list2);
241 	return (0);
242 }
243 
244 #define	swap(a, b) {					\
245 		s = b;					\
246 		i = size;				\
247 		do {					\
248 			tmp = *a; *a++ = *s; *s++ = tmp; \
249 		} while (--i);				\
250 		a -= size;				\
251 	}
252 #define reverse(bot, top) {				\
253 	s = top;					\
254 	do {						\
255 		i = size;				\
256 		do {					\
257 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
258 		} while (--i);				\
259 		s -= size2;				\
260 	} while(bot < s);				\
261 }
262 
263 /*
264  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
265  * increasing order, list2 in a corresponding linked list.  Checks for runs
266  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
267  * is defined.  Otherwise simple pairwise merging is used.)
268  */
269 void
270 setup(u_char *list1, u_char *list2, size_t n, size_t size, cmp_t cmp)
271 {
272 	int i, length, size2, tmp, sense;
273 	u_char *f1, *f2, *s, *l2, *last, *p2;
274 
275 	size2 = size*2;
276 	if (n <= 5) {
277 		insertionsort(list1, n, size, cmp);
278 		*EVAL(list2) = (u_char*) list2 + n*size;
279 		return;
280 	}
281 	/*
282 	 * Avoid running pointers out of bounds; limit n to evens
283 	 * for simplicity.
284 	 */
285 	i = 4 + (n & 1);
286 	insertionsort(list1 + (n - i) * size, i, size, cmp);
287 	last = list1 + size * (n - i);
288 	*EVAL(list2 + (last - list1)) = list2 + n * size;
289 
290 #ifdef NATURAL
291 	p2 = list2;
292 	f1 = list1;
293 	sense = (CMP(f1, f1 + size) > 0);
294 	for (; f1 < last; sense = !sense) {
295 		length = 2;
296 					/* Find pairs with same sense. */
297 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
298 			if ((CMP(f2, f2+ size) > 0) != sense)
299 				break;
300 			length += 2;
301 		}
302 		if (length < THRESHOLD) {		/* Pairwise merge */
303 			do {
304 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
305 				if (sense > 0)
306 					swap (f1, f1 + size);
307 			} while ((f1 += size2) < f2);
308 		} else {				/* Natural merge */
309 			l2 = f2;
310 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
311 				if ((CMP(f2-size, f2) > 0) != sense) {
312 					p2 = *EVAL(p2) = f2 - list1 + list2;
313 					if (sense > 0)
314 						reverse(f1, f2-size);
315 					f1 = f2;
316 				}
317 			}
318 			if (sense > 0)
319 				reverse (f1, f2-size);
320 			f1 = f2;
321 			if (f2 < last || CMP(f2 - size, f2) > 0)
322 				p2 = *EVAL(p2) = f2 - list1 + list2;
323 			else
324 				p2 = *EVAL(p2) = list2 + n*size;
325 		}
326 	}
327 #else		/* pairwise merge only. */
328 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
329 		p2 = *EVAL(p2) = p2 + size2;
330 		if (CMP (f1, f1 + size) > 0)
331 			swap(f1, f1 + size);
332 	}
333 #endif /* NATURAL */
334 }
335 
336 /*
337  * This is to avoid out-of-bounds addresses in sorting the
338  * last 4 elements.
339  */
340 static void
341 insertionsort(u_char *a, size_t n, size_t size, cmp_t cmp)
342 {
343 	u_char *ai, *s, *t, *u, tmp;
344 	int i;
345 
346 	for (ai = a+size; --n >= 1; ai += size)
347 		for (t = ai; t > a; t -= size) {
348 			u = t - size;
349 			if (CMP(u, t) <= 0)
350 				break;
351 			swap(u, t);
352 		}
353 }
354