xref: /freebsd/lib/libc/stdlib/merge.c (revision f5dc2263ab1be8a35a7e27e82103f9ccd41ae584)
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 /*
36  * Hybrid exponential search/linear search merge sort with hybrid
37  * natural/pairwise first pass.  Requires about .3% more comparisons
38  * for random data than LSMS with pairwise first pass alone.
39  * It works for objects as small as two bytes.
40  */
41 
42 #define NATURAL
43 #define THRESHOLD 16	/* Best choice for natural merge cut-off. */
44 
45 /* #define NATURAL to get hybrid natural merge.
46  * (The default is pairwise merging.)
47  */
48 
49 #include <sys/param.h>
50 
51 #include <errno.h>
52 #include <stdckdint.h>
53 #include <stdlib.h>
54 #include <string.h>
55 
56 #ifdef I_AM_MERGESORT_B
57 #include "block_abi.h"
58 #define	DECLARE_CMP	DECLARE_BLOCK(int, cmp, const void *, const void *)
59 typedef DECLARE_BLOCK(int, cmp_t, const void *, const void *);
60 #define	CMP(x, y)	CALL_BLOCK(cmp, x, y)
61 #else
62 typedef int (*cmp_t)(const void *, const void *);
63 #define	CMP(x, y)	cmp(x, y)
64 #endif
65 
66 static void setup(u_char *, u_char *, size_t, size_t, cmp_t);
67 static void insertionsort(u_char *, size_t, size_t, cmp_t);
68 
69 #define PSIZE sizeof(u_char *)
70 #define COPY_LIST(src, dst, last)		\
71 	do {					\
72 		memcpy(dst, src, last - src);	\
73 		dst += last - src;		\
74 		src += last - src;		\
75 	} while (0)
76 #define COPY_ELT(src, dst, i)			\
77 	do {					\
78 		memcpy(dst, src, i);		\
79 		src += i;			\
80 		dst += i;			\
81 	} while (0)
82 
83 /*
84  * Find the next possible pointer head.  (Trickery for forcing an array
85  * to do double duty as a linked list when objects do not align with word
86  * boundaries.
87  */
88 /* Assumption: PSIZE is a power of 2. */
89 #define EVAL(p) (u_char **)roundup2((uintptr_t)p, PSIZE)
90 
91 #ifdef I_AM_MERGESORT_B
92 int mergesort_b(void *, size_t, size_t, cmp_t);
93 #else
94 int mergesort(void *, size_t, size_t, cmp_t);
95 #endif
96 
97 /*
98  * Arguments are as for qsort.
99  */
100 int
101 #ifdef I_AM_MERGESORT_B
102 mergesort_b(void *base, size_t nmemb, size_t size, cmp_t cmp)
103 #else
104 mergesort(void *base, size_t nmemb, size_t size, cmp_t cmp)
105 #endif
106 {
107 	size_t i, nbytes, asize;
108 	int sense;
109 	int big;
110 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
111 	u_char *list2, *list1, *p2, *p, *last, **p1;
112 
113 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
114 		errno = EINVAL;
115 		return (-1);
116 	}
117 
118 	if (nmemb == 0)
119 		return (0);
120 
121 	if (ckd_mul(&nbytes, nmemb, size) || ckd_add(&asize, nbytes, PSIZE)) {
122 		errno = EINVAL;
123 		return (-1);
124 	}
125 
126 	if ((list2 = malloc(asize)) == NULL)
127 		return (-1);
128 
129 	list1 = base;
130 	setup(list1, list2, nmemb, size, cmp);
131 	last = list2 + nbytes;
132 	i = big = 0;
133 	while (*EVAL(list2) != last) {
134 	    l2 = list1;
135 	    p1 = EVAL(list1);
136 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
137 	    	p2 = *EVAL(p2);
138 	    	f1 = l2;
139 	    	f2 = l1 = list1 + (p2 - list2);
140 	    	if (p2 != last)
141 	    		p2 = *EVAL(p2);
142 	    	l2 = list1 + (p2 - list2);
143 	    	while (f1 < l1 && f2 < l2) {
144 	    		if (CMP(f1, f2) <= 0) {
145 	    			q = f2;
146 	    			b = f1, t = l1;
147 	    			sense = -1;
148 	    		} else {
149 	    			q = f1;
150 	    			b = f2, t = l2;
151 	    			sense = 0;
152 	    		}
153 	    		if (!big) {	/* here i = 0 */
154 				while ((b += size) < t && CMP(q, b) >sense)
155 	    				if (++i == 6) {
156 	    					big = 1;
157 	    					goto EXPONENTIAL;
158 	    				}
159 	    		} else {
160 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
161 	    				if ((p = (b + i)) >= t) {
162 	    					if ((p = t - size) > b &&
163 						    CMP(q, p) <= sense)
164 	    						t = p;
165 	    					else
166 	    						b = p;
167 	    					break;
168 	    				} else if (CMP(q, p) <= sense) {
169 	    					t = p;
170 	    					if (i == size)
171 	    						big = 0;
172 	    					goto FASTCASE;
173 	    				} else
174 	    					b = p;
175 				while (t > b+size) {
176 	    				i = (((t - b) / size) >> 1) * size;
177 	    				if (CMP(q, p = b + i) <= sense)
178 	    					t = p;
179 	    				else
180 	    					b = p;
181 	    			}
182 	    			goto COPY;
183 FASTCASE:	    		while (i > size)
184 	    				if (CMP(q,
185 	    					p = b + (i >>= 1)) <= sense)
186 	    					t = p;
187 	    				else
188 	    					b = p;
189 COPY:	    			b = t;
190 	    		}
191 	    		i = size;
192 	    		if (q == f1) {
193 	    			COPY_LIST(f2, tp2, b);
194 	    			COPY_ELT(f1, tp2, i);
195 	    		} else {
196 	    			COPY_LIST(f1, tp2, b);
197 	    			COPY_ELT(f2, tp2, i);
198 	    		}
199 	    	}
200 	    	if (f2 < l2)
201 	    		COPY_LIST(f2, tp2, l2);
202 	    	else if (f1 < l1)
203 	   		COPY_LIST(f1, tp2, l1);
204 	    	*p1 = l2;
205 	    }
206 	    tp2 = list1;	/* swap list1, list2 */
207 	    list1 = list2;
208 	    list2 = tp2;
209 	    last = list2 + nbytes;
210 	}
211 	if (base == list2) {
212 		memmove(list2, list1, nbytes);
213 		list2 = list1;
214 	}
215 	free(list2);
216 	return (0);
217 }
218 
219 #define	swap(a, b) {					\
220 		s = b;					\
221 		i = size;				\
222 		do {					\
223 			tmp = *a; *a++ = *s; *s++ = tmp; \
224 		} while (--i);				\
225 		a -= size;				\
226 	}
227 #define reverse(bot, top) {				\
228 	s = top;					\
229 	do {						\
230 		i = size;				\
231 		do {					\
232 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
233 		} while (--i);				\
234 		s -= size2;				\
235 	} while(bot < s);				\
236 }
237 
238 /*
239  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
240  * increasing order, list2 in a corresponding linked list.  Checks for runs
241  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
242  * is defined.  Otherwise simple pairwise merging is used.)
243  */
244 void
245 setup(u_char *list1, u_char *list2, size_t n, size_t size, cmp_t cmp)
246 {
247 	int i, length, size2, tmp, sense;
248 	u_char *f1, *f2, *s, *l2, *last, *p2;
249 
250 	size2 = size*2;
251 	if (n <= 5) {
252 		insertionsort(list1, n, size, cmp);
253 		*EVAL(list2) = (u_char*) list2 + n*size;
254 		return;
255 	}
256 	/*
257 	 * Avoid running pointers out of bounds; limit n to evens
258 	 * for simplicity.
259 	 */
260 	i = 4 + (n & 1);
261 	insertionsort(list1 + (n - i) * size, i, size, cmp);
262 	last = list1 + size * (n - i);
263 	*EVAL(list2 + (last - list1)) = list2 + n * size;
264 
265 #ifdef NATURAL
266 	p2 = list2;
267 	f1 = list1;
268 	sense = (CMP(f1, f1 + size) > 0);
269 	for (; f1 < last; sense = !sense) {
270 		length = 2;
271 					/* Find pairs with same sense. */
272 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
273 			if ((CMP(f2, f2+ size) > 0) != sense)
274 				break;
275 			length += 2;
276 		}
277 		if (length < THRESHOLD) {		/* Pairwise merge */
278 			do {
279 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
280 				if (sense > 0)
281 					swap (f1, f1 + size);
282 			} while ((f1 += size2) < f2);
283 		} else {				/* Natural merge */
284 			l2 = f2;
285 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
286 				if ((CMP(f2-size, f2) > 0) != sense) {
287 					p2 = *EVAL(p2) = f2 - list1 + list2;
288 					if (sense > 0)
289 						reverse(f1, f2-size);
290 					f1 = f2;
291 				}
292 			}
293 			if (sense > 0)
294 				reverse (f1, f2-size);
295 			f1 = f2;
296 			if (f2 < last || CMP(f2 - size, f2) > 0)
297 				p2 = *EVAL(p2) = f2 - list1 + list2;
298 			else
299 				p2 = *EVAL(p2) = list2 + n*size;
300 		}
301 	}
302 #else		/* pairwise merge only. */
303 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
304 		p2 = *EVAL(p2) = p2 + size2;
305 		if (CMP (f1, f1 + size) > 0)
306 			swap(f1, f1 + size);
307 	}
308 #endif /* NATURAL */
309 }
310 
311 /*
312  * This is to avoid out-of-bounds addresses in sorting the
313  * last 4 elements.
314  */
315 static void
316 insertionsort(u_char *a, size_t n, size_t size, cmp_t cmp)
317 {
318 	u_char *ai, *s, *t, *u, tmp;
319 	int i;
320 
321 	for (ai = a+size; --n >= 1; ai += size)
322 		for (t = ai; t > a; t -= size) {
323 			u = t - size;
324 			if (CMP(u, t) <= 0)
325 				break;
326 			swap(u, t);
327 		}
328 }
329