xref: /freebsd/lib/libc/stdlib/merge.c (revision 4b9d605768acabc460aa6dcfe8a1f8db35b16794)
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 <stdlib.h>
53 #include <string.h>
54 
55 #ifdef I_AM_MERGESORT_B
56 #include "block_abi.h"
57 #define	DECLARE_CMP	DECLARE_BLOCK(int, cmp, const void *, const void *)
58 typedef DECLARE_BLOCK(int, cmp_t, const void *, const void *);
59 #define	CMP(x, y)	CALL_BLOCK(cmp, x, y)
60 #else
61 typedef int (*cmp_t)(const void *, const void *);
62 #define	CMP(x, y)	cmp(x, y)
63 #endif
64 
65 static void setup(u_char *, u_char *, size_t, size_t, cmp_t);
66 static void insertionsort(u_char *, size_t, size_t, cmp_t);
67 
68 #define ISIZE sizeof(int)
69 #define PSIZE sizeof(u_char *)
70 #define ICOPY_LIST(src, dst, last)				\
71 	do							\
72 	*(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE;	\
73 	while(src < last)
74 #define ICOPY_ELT(src, dst, i)					\
75 	do							\
76 	*(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE;	\
77 	while (i -= ISIZE)
78 
79 #define CCOPY_LIST(src, dst, last)		\
80 	do					\
81 		*dst++ = *src++;		\
82 	while (src < last)
83 #define CCOPY_ELT(src, dst, i)			\
84 	do					\
85 		*dst++ = *src++;		\
86 	while (i -= 1)
87 
88 /*
89  * Find the next possible pointer head.  (Trickery for forcing an array
90  * to do double duty as a linked list when objects do not align with word
91  * boundaries.
92  */
93 /* Assumption: PSIZE is a power of 2. */
94 #define EVAL(p) (u_char **)roundup2((uintptr_t)p, PSIZE)
95 
96 #ifdef I_AM_MERGESORT_B
97 int mergesort_b(void *, size_t, size_t, cmp_t);
98 #else
99 int mergesort(void *, size_t, size_t, cmp_t);
100 #endif
101 
102 /*
103  * Arguments are as for qsort.
104  */
105 int
106 #ifdef I_AM_MERGESORT_B
107 mergesort_b(void *base, size_t nmemb, size_t size, cmp_t cmp)
108 #else
109 mergesort(void *base, size_t nmemb, size_t size, cmp_t cmp)
110 #endif
111 {
112 	size_t i;
113 	int sense;
114 	int big, iflag;
115 	u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
116 	u_char *list2, *list1, *p2, *p, *last, **p1;
117 
118 	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
119 		errno = EINVAL;
120 		return (-1);
121 	}
122 
123 	if (nmemb == 0)
124 		return (0);
125 
126 	iflag = 0;
127 	if (__is_aligned(size, ISIZE) && __is_aligned(base, ISIZE))
128 		iflag = 1;
129 
130 	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
131 		return (-1);
132 
133 	list1 = base;
134 	setup(list1, list2, nmemb, size, cmp);
135 	last = list2 + nmemb * size;
136 	i = big = 0;
137 	while (*EVAL(list2) != last) {
138 	    l2 = list1;
139 	    p1 = EVAL(list1);
140 	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
141 	    	p2 = *EVAL(p2);
142 	    	f1 = l2;
143 	    	f2 = l1 = list1 + (p2 - list2);
144 	    	if (p2 != last)
145 	    		p2 = *EVAL(p2);
146 	    	l2 = list1 + (p2 - list2);
147 	    	while (f1 < l1 && f2 < l2) {
148 	    		if (CMP(f1, f2) <= 0) {
149 	    			q = f2;
150 	    			b = f1, t = l1;
151 	    			sense = -1;
152 	    		} else {
153 	    			q = f1;
154 	    			b = f2, t = l2;
155 	    			sense = 0;
156 	    		}
157 	    		if (!big) {	/* here i = 0 */
158 				while ((b += size) < t && CMP(q, b) >sense)
159 	    				if (++i == 6) {
160 	    					big = 1;
161 	    					goto EXPONENTIAL;
162 	    				}
163 	    		} else {
164 EXPONENTIAL:	    		for (i = size; ; i <<= 1)
165 	    				if ((p = (b + i)) >= t) {
166 	    					if ((p = t - size) > b &&
167 						    CMP(q, p) <= sense)
168 	    						t = p;
169 	    					else
170 	    						b = p;
171 	    					break;
172 	    				} else if (CMP(q, p) <= sense) {
173 	    					t = p;
174 	    					if (i == size)
175 	    						big = 0;
176 	    					goto FASTCASE;
177 	    				} else
178 	    					b = p;
179 				while (t > b+size) {
180 	    				i = (((t - b) / size) >> 1) * size;
181 	    				if (CMP(q, p = b + i) <= sense)
182 	    					t = p;
183 	    				else
184 	    					b = p;
185 	    			}
186 	    			goto COPY;
187 FASTCASE:	    		while (i > size)
188 	    				if (CMP(q,
189 	    					p = b + (i >>= 1)) <= sense)
190 	    					t = p;
191 	    				else
192 	    					b = p;
193 COPY:	    			b = t;
194 	    		}
195 	    		i = size;
196 	    		if (q == f1) {
197 	    			if (iflag) {
198 	    				ICOPY_LIST(f2, tp2, b);
199 	    				ICOPY_ELT(f1, tp2, i);
200 	    			} else {
201 	    				CCOPY_LIST(f2, tp2, b);
202 	    				CCOPY_ELT(f1, tp2, i);
203 	    			}
204 	    		} else {
205 	    			if (iflag) {
206 	    				ICOPY_LIST(f1, tp2, b);
207 	    				ICOPY_ELT(f2, tp2, i);
208 	    			} else {
209 	    				CCOPY_LIST(f1, tp2, b);
210 	    				CCOPY_ELT(f2, tp2, i);
211 	    			}
212 	    		}
213 	    	}
214 	    	if (f2 < l2) {
215 	    		if (iflag)
216 	    			ICOPY_LIST(f2, tp2, l2);
217 	    		else
218 	    			CCOPY_LIST(f2, tp2, l2);
219 	    	} else if (f1 < l1) {
220 	    		if (iflag)
221 	    			ICOPY_LIST(f1, tp2, l1);
222 	    		else
223 	    			CCOPY_LIST(f1, tp2, l1);
224 	    	}
225 	    	*p1 = l2;
226 	    }
227 	    tp2 = list1;	/* swap list1, list2 */
228 	    list1 = list2;
229 	    list2 = tp2;
230 	    last = list2 + nmemb*size;
231 	}
232 	if (base == list2) {
233 		memmove(list2, list1, nmemb*size);
234 		list2 = list1;
235 	}
236 	free(list2);
237 	return (0);
238 }
239 
240 #define	swap(a, b) {					\
241 		s = b;					\
242 		i = size;				\
243 		do {					\
244 			tmp = *a; *a++ = *s; *s++ = tmp; \
245 		} while (--i);				\
246 		a -= size;				\
247 	}
248 #define reverse(bot, top) {				\
249 	s = top;					\
250 	do {						\
251 		i = size;				\
252 		do {					\
253 			tmp = *bot; *bot++ = *s; *s++ = tmp; \
254 		} while (--i);				\
255 		s -= size2;				\
256 	} while(bot < s);				\
257 }
258 
259 /*
260  * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
261  * increasing order, list2 in a corresponding linked list.  Checks for runs
262  * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
263  * is defined.  Otherwise simple pairwise merging is used.)
264  */
265 void
266 setup(u_char *list1, u_char *list2, size_t n, size_t size, cmp_t cmp)
267 {
268 	int i, length, size2, tmp, sense;
269 	u_char *f1, *f2, *s, *l2, *last, *p2;
270 
271 	size2 = size*2;
272 	if (n <= 5) {
273 		insertionsort(list1, n, size, cmp);
274 		*EVAL(list2) = (u_char*) list2 + n*size;
275 		return;
276 	}
277 	/*
278 	 * Avoid running pointers out of bounds; limit n to evens
279 	 * for simplicity.
280 	 */
281 	i = 4 + (n & 1);
282 	insertionsort(list1 + (n - i) * size, i, size, cmp);
283 	last = list1 + size * (n - i);
284 	*EVAL(list2 + (last - list1)) = list2 + n * size;
285 
286 #ifdef NATURAL
287 	p2 = list2;
288 	f1 = list1;
289 	sense = (CMP(f1, f1 + size) > 0);
290 	for (; f1 < last; sense = !sense) {
291 		length = 2;
292 					/* Find pairs with same sense. */
293 		for (f2 = f1 + size2; f2 < last; f2 += size2) {
294 			if ((CMP(f2, f2+ size) > 0) != sense)
295 				break;
296 			length += 2;
297 		}
298 		if (length < THRESHOLD) {		/* Pairwise merge */
299 			do {
300 				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
301 				if (sense > 0)
302 					swap (f1, f1 + size);
303 			} while ((f1 += size2) < f2);
304 		} else {				/* Natural merge */
305 			l2 = f2;
306 			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
307 				if ((CMP(f2-size, f2) > 0) != sense) {
308 					p2 = *EVAL(p2) = f2 - list1 + list2;
309 					if (sense > 0)
310 						reverse(f1, f2-size);
311 					f1 = f2;
312 				}
313 			}
314 			if (sense > 0)
315 				reverse (f1, f2-size);
316 			f1 = f2;
317 			if (f2 < last || CMP(f2 - size, f2) > 0)
318 				p2 = *EVAL(p2) = f2 - list1 + list2;
319 			else
320 				p2 = *EVAL(p2) = list2 + n*size;
321 		}
322 	}
323 #else		/* pairwise merge only. */
324 	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
325 		p2 = *EVAL(p2) = p2 + size2;
326 		if (CMP (f1, f1 + size) > 0)
327 			swap(f1, f1 + size);
328 	}
329 #endif /* NATURAL */
330 }
331 
332 /*
333  * This is to avoid out-of-bounds addresses in sorting the
334  * last 4 elements.
335  */
336 static void
337 insertionsort(u_char *a, size_t n, size_t size, cmp_t cmp)
338 {
339 	u_char *ai, *s, *t, *u, tmp;
340 	int i;
341 
342 	for (ai = a+size; --n >= 1; ai += size)
343 		for (t = ai; t > a; t -= size) {
344 			u = t - size;
345 			if (CMP(u, t) <= 0)
346 				break;
347 			swap(u, t);
348 		}
349 }
350