xref: /freebsd/lib/libc/stdlib/merge.c (revision 5f4c09dd85bff675e0ca63c55ea3c517e0fddfcc)
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 /*
39  * Hybrid exponential search/linear search merge sort with hybrid
40  * natural/pairwise first pass.  Requires about .3% more comparisons
41  * for random data than LSMS with pairwise first pass alone.
42  * It works for objects as small as two bytes.
43  */
44 
45 #define NATURAL
46 #define THRESHOLD 16	/* Best choice for natural merge cut-off. */
47 
48 /* #define NATURAL to get hybrid natural merge.
49  * (The default is pairwise merging.)
50  */
51 
52 #include <sys/param.h>
53 
54 #include <errno.h>
55 #include <stdlib.h>
56 #include <string.h>
57 
58 #ifdef I_AM_MERGESORT_B
59 #include "block_abi.h"
60 #define	DECLARE_CMP	DECLARE_BLOCK(int, cmp, const void *, const void *)
61 typedef DECLARE_BLOCK(int, cmp_t, const void *, const void *);
62 #define	CMP(x, y)	CALL_BLOCK(cmp, x, y)
63 #else
64 typedef int (*cmp_t)(const void *, const void *);
65 #define	CMP(x, y)	cmp(x, y)
66 #endif
67 
68 static void setup(u_char *, u_char *, size_t, size_t, cmp_t);
69 static void insertionsort(u_char *, size_t, size_t, cmp_t);
70 
71 #define ISIZE sizeof(int)
72 #define PSIZE sizeof(u_char *)
73 #define ICOPY_LIST(src, dst, last)				\
74 	do							\
75 	*(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE;	\
76 	while(src < last)
77 #define ICOPY_ELT(src, dst, i)					\
78 	do							\
79 	*(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE;	\
80 	while (i -= ISIZE)
81 
82 #define CCOPY_LIST(src, dst, last)		\
83 	do					\
84 		*dst++ = *src++;		\
85 	while (src < last)
86 #define CCOPY_ELT(src, dst, i)			\
87 	do					\
88 		*dst++ = *src++;		\
89 	while (i -= 1)
90 
91 /*
92  * Find the next possible pointer head.  (Trickery for forcing an array
93  * to do double duty as a linked list when objects do not align with word
94  * boundaries.
95  */
96 /* Assumption: PSIZE is a power of 2. */
97 #define EVAL(p) (u_char **)roundup2((uintptr_t)p, PSIZE)
98 
99 #ifdef I_AM_MERGESORT_B
100 int mergesort_b(void *, size_t, size_t, cmp_t);
101 #else
102 int mergesort(void *, size_t, size_t, cmp_t);
103 #endif
104 
105 /*
106  * Arguments are as for qsort.
107  */
108 int
109 #ifdef I_AM_MERGESORT_B
110 mergesort_b(void *base, size_t nmemb, size_t size, cmp_t cmp)
111 #else
112 mergesort(void *base, size_t nmemb, size_t size, cmp_t cmp)
113 #endif
114 {
115 	size_t i;
116 	int sense;
117 	int big, iflag;
118 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
119 	u_char *list2, *list1, *p2, *p, *last, **p1;
120 
121 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
122 		errno = EINVAL;
123 		return (-1);
124 	}
125 
126 	if (nmemb == 0)
127 		return (0);
128 
129 	iflag = 0;
130 	if (__is_aligned(size, ISIZE) && __is_aligned(base, ISIZE))
131 		iflag = 1;
132 
133 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
134 		return (-1);
135 
136 	list1 = base;
137 	setup(list1, list2, nmemb, size, cmp);
138 	last = list2 + nmemb * size;
139 	i = big = 0;
140 	while (*EVAL(list2) != last) {
141 	    l2 = list1;
142 	    p1 = EVAL(list1);
143 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
144 	    	p2 = *EVAL(p2);
145 	    	f1 = l2;
146 	    	f2 = l1 = list1 + (p2 - list2);
147 	    	if (p2 != last)
148 	    		p2 = *EVAL(p2);
149 	    	l2 = list1 + (p2 - list2);
150 	    	while (f1 < l1 && f2 < l2) {
151 	    		if (CMP(f1, f2) <= 0) {
152 	    			q = f2;
153 	    			b = f1, t = l1;
154 	    			sense = -1;
155 	    		} else {
156 	    			q = f1;
157 	    			b = f2, t = l2;
158 	    			sense = 0;
159 	    		}
160 	    		if (!big) {	/* here i = 0 */
161 				while ((b += size) < t && CMP(q, b) >sense)
162 	    				if (++i == 6) {
163 	    					big = 1;
164 	    					goto EXPONENTIAL;
165 	    				}
166 	    		} else {
167 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
168 	    				if ((p = (b + i)) >= t) {
169 	    					if ((p = t - size) > b &&
170 						    CMP(q, p) <= sense)
171 	    						t = p;
172 	    					else
173 	    						b = p;
174 	    					break;
175 	    				} else if (CMP(q, p) <= sense) {
176 	    					t = p;
177 	    					if (i == size)
178 	    						big = 0;
179 	    					goto FASTCASE;
180 	    				} else
181 	    					b = p;
182 				while (t > b+size) {
183 	    				i = (((t - b) / size) >> 1) * size;
184 	    				if (CMP(q, p = b + i) <= sense)
185 	    					t = p;
186 	    				else
187 	    					b = p;
188 	    			}
189 	    			goto COPY;
190 FASTCASE:	    		while (i > size)
191 	    				if (CMP(q,
192 	    					p = b + (i >>= 1)) <= sense)
193 	    					t = p;
194 	    				else
195 	    					b = p;
196 COPY:	    			b = t;
197 	    		}
198 	    		i = size;
199 	    		if (q == f1) {
200 	    			if (iflag) {
201 	    				ICOPY_LIST(f2, tp2, b);
202 	    				ICOPY_ELT(f1, tp2, i);
203 	    			} else {
204 	    				CCOPY_LIST(f2, tp2, b);
205 	    				CCOPY_ELT(f1, tp2, i);
206 	    			}
207 	    		} else {
208 	    			if (iflag) {
209 	    				ICOPY_LIST(f1, tp2, b);
210 	    				ICOPY_ELT(f2, tp2, i);
211 	    			} else {
212 	    				CCOPY_LIST(f1, tp2, b);
213 	    				CCOPY_ELT(f2, tp2, i);
214 	    			}
215 	    		}
216 	    	}
217 	    	if (f2 < l2) {
218 	    		if (iflag)
219 	    			ICOPY_LIST(f2, tp2, l2);
220 	    		else
221 	    			CCOPY_LIST(f2, tp2, l2);
222 	    	} else if (f1 < l1) {
223 	    		if (iflag)
224 	    			ICOPY_LIST(f1, tp2, l1);
225 	    		else
226 	    			CCOPY_LIST(f1, tp2, l1);
227 	    	}
228 	    	*p1 = l2;
229 	    }
230 	    tp2 = list1;	/* swap list1, list2 */
231 	    list1 = list2;
232 	    list2 = tp2;
233 	    last = list2 + nmemb*size;
234 	}
235 	if (base == list2) {
236 		memmove(list2, list1, nmemb*size);
237 		list2 = list1;
238 	}
239 	free(list2);
240 	return (0);
241 }
242 
243 #define	swap(a, b) {					\
244 		s = b;					\
245 		i = size;				\
246 		do {					\
247 			tmp = *a; *a++ = *s; *s++ = tmp; \
248 		} while (--i);				\
249 		a -= size;				\
250 	}
251 #define reverse(bot, top) {				\
252 	s = top;					\
253 	do {						\
254 		i = size;				\
255 		do {					\
256 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
257 		} while (--i);				\
258 		s -= size2;				\
259 	} while(bot < s);				\
260 }
261 
262 /*
263  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
264  * increasing order, list2 in a corresponding linked list.  Checks for runs
265  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
266  * is defined.  Otherwise simple pairwise merging is used.)
267  */
268 void
269 setup(u_char *list1, u_char *list2, size_t n, size_t size, cmp_t cmp)
270 {
271 	int i, length, size2, tmp, sense;
272 	u_char *f1, *f2, *s, *l2, *last, *p2;
273 
274 	size2 = size*2;
275 	if (n <= 5) {
276 		insertionsort(list1, n, size, cmp);
277 		*EVAL(list2) = (u_char*) list2 + n*size;
278 		return;
279 	}
280 	/*
281 	 * Avoid running pointers out of bounds; limit n to evens
282 	 * for simplicity.
283 	 */
284 	i = 4 + (n & 1);
285 	insertionsort(list1 + (n - i) * size, i, size, cmp);
286 	last = list1 + size * (n - i);
287 	*EVAL(list2 + (last - list1)) = list2 + n * size;
288 
289 #ifdef NATURAL
290 	p2 = list2;
291 	f1 = list1;
292 	sense = (CMP(f1, f1 + size) > 0);
293 	for (; f1 < last; sense = !sense) {
294 		length = 2;
295 					/* Find pairs with same sense. */
296 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
297 			if ((CMP(f2, f2+ size) > 0) != sense)
298 				break;
299 			length += 2;
300 		}
301 		if (length < THRESHOLD) {		/* Pairwise merge */
302 			do {
303 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
304 				if (sense > 0)
305 					swap (f1, f1 + size);
306 			} while ((f1 += size2) < f2);
307 		} else {				/* Natural merge */
308 			l2 = f2;
309 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
310 				if ((CMP(f2-size, f2) > 0) != sense) {
311 					p2 = *EVAL(p2) = f2 - list1 + list2;
312 					if (sense > 0)
313 						reverse(f1, f2-size);
314 					f1 = f2;
315 				}
316 			}
317 			if (sense > 0)
318 				reverse (f1, f2-size);
319 			f1 = f2;
320 			if (f2 < last || CMP(f2 - size, f2) > 0)
321 				p2 = *EVAL(p2) = f2 - list1 + list2;
322 			else
323 				p2 = *EVAL(p2) = list2 + n*size;
324 		}
325 	}
326 #else		/* pairwise merge only. */
327 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
328 		p2 = *EVAL(p2) = p2 + size2;
329 		if (CMP (f1, f1 + size) > 0)
330 			swap(f1, f1 + size);
331 	}
332 #endif /* NATURAL */
333 }
334 
335 /*
336  * This is to avoid out-of-bounds addresses in sorting the
337  * last 4 elements.
338  */
339 static void
340 insertionsort(u_char *a, size_t n, size_t size, cmp_t cmp)
341 {
342 	u_char *ai, *s, *t, *u, tmp;
343 	int i;
344 
345 	for (ai = a+size; --n >= 1; ai += size)
346 		for (t = ai; t > a; t -= size) {
347 			u = t - size;
348 			if (CMP(u, t) <= 0)
349 				break;
350 			swap(u, t);
351 		}
352 }
353