1 12c5bde7 2020-10-16 stsp /* $OpenBSD: merge.c,v 1.10 2015/06/21 03:20:56 millert Exp $ */
3 12c5bde7 2020-10-16 stsp * Copyright (c) 1992, 1993
4 12c5bde7 2020-10-16 stsp * The Regents of the University of California. All rights reserved.
6 12c5bde7 2020-10-16 stsp * This code is derived from software contributed to Berkeley by
7 12c5bde7 2020-10-16 stsp * Peter McIlroy.
9 12c5bde7 2020-10-16 stsp * Redistribution and use in source and binary forms, with or without
10 12c5bde7 2020-10-16 stsp * modification, are permitted provided that the following conditions
12 12c5bde7 2020-10-16 stsp * 1. Redistributions of source code must retain the above copyright
13 12c5bde7 2020-10-16 stsp * notice, this list of conditions and the following disclaimer.
14 12c5bde7 2020-10-16 stsp * 2. Redistributions in binary form must reproduce the above copyright
15 12c5bde7 2020-10-16 stsp * notice, this list of conditions and the following disclaimer in the
16 12c5bde7 2020-10-16 stsp * documentation and/or other materials provided with the distribution.
17 12c5bde7 2020-10-16 stsp * 3. Neither the name of the University nor the names of its contributors
18 12c5bde7 2020-10-16 stsp * may be used to endorse or promote products derived from this software
19 12c5bde7 2020-10-16 stsp * without specific prior written permission.
21 12c5bde7 2020-10-16 stsp * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 12c5bde7 2020-10-16 stsp * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 12c5bde7 2020-10-16 stsp * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 12c5bde7 2020-10-16 stsp * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 12c5bde7 2020-10-16 stsp * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 12c5bde7 2020-10-16 stsp * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 12c5bde7 2020-10-16 stsp * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 12c5bde7 2020-10-16 stsp * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 12c5bde7 2020-10-16 stsp * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 12c5bde7 2020-10-16 stsp * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 12c5bde7 2020-10-16 stsp * SUCH DAMAGE.
35 12c5bde7 2020-10-16 stsp * Hybrid exponential search/linear search merge sort with hybrid
36 12c5bde7 2020-10-16 stsp * natural/pairwise first pass. Requires about .3% more comparisons
37 12c5bde7 2020-10-16 stsp * for random data than LSMS with pairwise first pass alone.
38 12c5bde7 2020-10-16 stsp * It works for objects as small as two bytes.
41 12c5bde7 2020-10-16 stsp #define NATURAL
42 12c5bde7 2020-10-16 stsp #define THRESHOLD 16 /* Best choice for natural merge cut-off. */
44 12c5bde7 2020-10-16 stsp /* #define NATURAL to get hybrid natural merge.
45 12c5bde7 2020-10-16 stsp * (The default is pairwise merging.)
48 12c5bde7 2020-10-16 stsp #include <sys/types.h>
50 12c5bde7 2020-10-16 stsp #include <errno.h>
51 12c5bde7 2020-10-16 stsp #include <stdlib.h>
52 12c5bde7 2020-10-16 stsp #include <string.h>
54 12c5bde7 2020-10-16 stsp static void setup(u_char *list1, u_char *list2, size_t n, size_t size,
55 12c5bde7 2020-10-16 stsp int (*cmp)(const void *, const void *));
56 12c5bde7 2020-10-16 stsp static void insertionsort(u_char *a, size_t n, size_t size,
57 12c5bde7 2020-10-16 stsp int (*cmp)(const void *, const void *));
59 12c5bde7 2020-10-16 stsp #define ISIZE sizeof(int)
60 12c5bde7 2020-10-16 stsp #define PSIZE sizeof(u_char *)
61 12c5bde7 2020-10-16 stsp #define ICOPY_LIST(src, dst, last) \
63 12c5bde7 2020-10-16 stsp *(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE; \
64 12c5bde7 2020-10-16 stsp while(src < last)
65 12c5bde7 2020-10-16 stsp #define ICOPY_ELT(src, dst, i) \
67 12c5bde7 2020-10-16 stsp *(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE; \
68 12c5bde7 2020-10-16 stsp while (i -= ISIZE)
70 12c5bde7 2020-10-16 stsp #define CCOPY_LIST(src, dst, last) \
72 12c5bde7 2020-10-16 stsp *dst++ = *src++; \
73 12c5bde7 2020-10-16 stsp while (src < last)
74 12c5bde7 2020-10-16 stsp #define CCOPY_ELT(src, dst, i) \
76 12c5bde7 2020-10-16 stsp *dst++ = *src++; \
77 12c5bde7 2020-10-16 stsp while (i -= 1)
80 12c5bde7 2020-10-16 stsp * Find the next possible pointer head. (Trickery for forcing an array
81 12c5bde7 2020-10-16 stsp * to do double duty as a linked list when objects do not align with word
82 12c5bde7 2020-10-16 stsp * boundaries.
84 12c5bde7 2020-10-16 stsp /* Assumption: PSIZE is a power of 2. */
85 12c5bde7 2020-10-16 stsp #define EVAL(p) (u_char **) \
86 12c5bde7 2020-10-16 stsp ((u_char *)0 + \
87 12c5bde7 2020-10-16 stsp (((u_char *)p + PSIZE - 1 - (u_char *) 0) & ~(PSIZE - 1)))
90 12c5bde7 2020-10-16 stsp * Arguments are as for qsort.
93 12c5bde7 2020-10-16 stsp mergesort(void *base, size_t nmemb, size_t size,
94 12c5bde7 2020-10-16 stsp int (*cmp)(const void *, const void *))
96 12c5bde7 2020-10-16 stsp int i, sense;
97 12c5bde7 2020-10-16 stsp int big, iflag;
98 12c5bde7 2020-10-16 stsp u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
99 12c5bde7 2020-10-16 stsp u_char *list2, *list1, *p2, *p, *last, **p1;
101 12c5bde7 2020-10-16 stsp if (size < PSIZE / 2) { /* Pointers must fit into 2 * size. */
102 12c5bde7 2020-10-16 stsp errno = EINVAL;
103 12c5bde7 2020-10-16 stsp return (-1);
106 12c5bde7 2020-10-16 stsp if (nmemb == 0)
107 12c5bde7 2020-10-16 stsp return (0);
111 12c5bde7 2020-10-16 stsp * Stupid subtraction for the Cray.
114 12c5bde7 2020-10-16 stsp if (!(size % ISIZE) && !(((char *)base - (char *)0) % ISIZE))
117 12c5bde7 2020-10-16 stsp if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
118 12c5bde7 2020-10-16 stsp return (-1);
120 12c5bde7 2020-10-16 stsp list1 = base;
121 12c5bde7 2020-10-16 stsp setup(list1, list2, nmemb, size, cmp);
122 12c5bde7 2020-10-16 stsp last = list2 + nmemb * size;
123 12c5bde7 2020-10-16 stsp i = big = 0;
124 12c5bde7 2020-10-16 stsp while (*EVAL(list2) != last) {
125 12c5bde7 2020-10-16 stsp l2 = list1;
126 12c5bde7 2020-10-16 stsp p1 = EVAL(list1);
127 12c5bde7 2020-10-16 stsp for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
128 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2);
130 12c5bde7 2020-10-16 stsp f2 = l1 = list1 + (p2 - list2);
131 12c5bde7 2020-10-16 stsp if (p2 != last)
132 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2);
133 12c5bde7 2020-10-16 stsp l2 = list1 + (p2 - list2);
134 12c5bde7 2020-10-16 stsp while (f1 < l1 && f2 < l2) {
135 12c5bde7 2020-10-16 stsp if ((*cmp)(f1, f2) <= 0) {
137 12c5bde7 2020-10-16 stsp b = f1, t = l1;
138 12c5bde7 2020-10-16 stsp sense = -1;
141 12c5bde7 2020-10-16 stsp b = f2, t = l2;
144 12c5bde7 2020-10-16 stsp if (!big) { /* here i = 0 */
145 12c5bde7 2020-10-16 stsp while ((b += size) < t && cmp(q, b) >sense)
146 12c5bde7 2020-10-16 stsp if (++i == 6) {
148 12c5bde7 2020-10-16 stsp goto EXPONENTIAL;
151 12c5bde7 2020-10-16 stsp EXPONENTIAL: for (i = size; ; i <<= 1)
152 12c5bde7 2020-10-16 stsp if ((p = (b + i)) >= t) {
153 12c5bde7 2020-10-16 stsp if ((p = t - size) > b &&
154 12c5bde7 2020-10-16 stsp (*cmp)(q, p) <= sense)
159 12c5bde7 2020-10-16 stsp } else if ((*cmp)(q, p) <= sense) {
161 12c5bde7 2020-10-16 stsp if (i == size)
163 12c5bde7 2020-10-16 stsp goto FASTCASE;
166 12c5bde7 2020-10-16 stsp while (t > b+size) {
167 12c5bde7 2020-10-16 stsp i = (((t - b) / size) >> 1) * size;
168 12c5bde7 2020-10-16 stsp if ((*cmp)(q, p = b + i) <= sense)
174 12c5bde7 2020-10-16 stsp FASTCASE: while (i > size)
175 12c5bde7 2020-10-16 stsp if ((*cmp)(q,
176 12c5bde7 2020-10-16 stsp p = b + (i >>= 1)) <= sense)
180 12c5bde7 2020-10-16 stsp COPY: b = t;
183 12c5bde7 2020-10-16 stsp if (q == f1) {
184 12c5bde7 2020-10-16 stsp if (iflag) {
185 12c5bde7 2020-10-16 stsp ICOPY_LIST(f2, tp2, b);
186 12c5bde7 2020-10-16 stsp ICOPY_ELT(f1, tp2, i);
188 12c5bde7 2020-10-16 stsp CCOPY_LIST(f2, tp2, b);
189 12c5bde7 2020-10-16 stsp CCOPY_ELT(f1, tp2, i);
192 12c5bde7 2020-10-16 stsp if (iflag) {
193 12c5bde7 2020-10-16 stsp ICOPY_LIST(f1, tp2, b);
194 12c5bde7 2020-10-16 stsp ICOPY_ELT(f2, tp2, i);
196 12c5bde7 2020-10-16 stsp CCOPY_LIST(f1, tp2, b);
197 12c5bde7 2020-10-16 stsp CCOPY_ELT(f2, tp2, i);
201 12c5bde7 2020-10-16 stsp if (f2 < l2) {
203 12c5bde7 2020-10-16 stsp ICOPY_LIST(f2, tp2, l2);
205 12c5bde7 2020-10-16 stsp CCOPY_LIST(f2, tp2, l2);
206 12c5bde7 2020-10-16 stsp } else if (f1 < l1) {
208 12c5bde7 2020-10-16 stsp ICOPY_LIST(f1, tp2, l1);
210 12c5bde7 2020-10-16 stsp CCOPY_LIST(f1, tp2, l1);
214 12c5bde7 2020-10-16 stsp tp2 = list1; /* swap list1, list2 */
215 12c5bde7 2020-10-16 stsp list1 = list2;
216 12c5bde7 2020-10-16 stsp list2 = tp2;
217 12c5bde7 2020-10-16 stsp last = list2 + nmemb*size;
219 12c5bde7 2020-10-16 stsp if (base == list2) {
220 12c5bde7 2020-10-16 stsp memmove(list2, list1, nmemb*size);
221 12c5bde7 2020-10-16 stsp list2 = list1;
223 12c5bde7 2020-10-16 stsp free(list2);
224 12c5bde7 2020-10-16 stsp return (0);
227 12c5bde7 2020-10-16 stsp #define swap(a, b) { \
229 12c5bde7 2020-10-16 stsp i = size; \
231 12c5bde7 2020-10-16 stsp tmp = *a; *a++ = *s; *s++ = tmp; \
232 12c5bde7 2020-10-16 stsp } while (--i); \
233 12c5bde7 2020-10-16 stsp a -= size; \
235 12c5bde7 2020-10-16 stsp #define reverse(bot, top) { \
238 12c5bde7 2020-10-16 stsp i = size; \
240 12c5bde7 2020-10-16 stsp tmp = *bot; *bot++ = *s; *s++ = tmp; \
241 12c5bde7 2020-10-16 stsp } while (--i); \
242 12c5bde7 2020-10-16 stsp s -= size2; \
243 12c5bde7 2020-10-16 stsp } while(bot < s); \
247 12c5bde7 2020-10-16 stsp * Optional hybrid natural/pairwise first pass. Eats up list1 in runs of
248 12c5bde7 2020-10-16 stsp * increasing order, list2 in a corresponding linked list. Checks for runs
249 12c5bde7 2020-10-16 stsp * when THRESHOLD/2 pairs compare with same sense. (Only used when NATURAL
250 12c5bde7 2020-10-16 stsp * is defined. Otherwise simple pairwise merging is used.)
253 12c5bde7 2020-10-16 stsp setup(u_char *list1, u_char *list2, size_t n, size_t size,
254 12c5bde7 2020-10-16 stsp int (*cmp)(const void *, const void *))
256 12c5bde7 2020-10-16 stsp int i, length, size2, sense;
257 12c5bde7 2020-10-16 stsp u_char tmp, *f1, *f2, *s, *l2, *last, *p2;
259 12c5bde7 2020-10-16 stsp size2 = size*2;
260 12c5bde7 2020-10-16 stsp if (n <= 5) {
261 12c5bde7 2020-10-16 stsp insertionsort(list1, n, size, cmp);
262 12c5bde7 2020-10-16 stsp *EVAL(list2) = (u_char*) list2 + n*size;
266 12c5bde7 2020-10-16 stsp * Avoid running pointers out of bounds; limit n to evens
267 12c5bde7 2020-10-16 stsp * for simplicity.
269 12c5bde7 2020-10-16 stsp i = 4 + (n & 1);
270 12c5bde7 2020-10-16 stsp insertionsort(list1 + (n - i) * size, i, size, cmp);
271 12c5bde7 2020-10-16 stsp last = list1 + size * (n - i);
272 12c5bde7 2020-10-16 stsp *EVAL(list2 + (last - list1)) = list2 + n * size;
274 12c5bde7 2020-10-16 stsp #ifdef NATURAL
275 12c5bde7 2020-10-16 stsp p2 = list2;
276 12c5bde7 2020-10-16 stsp f1 = list1;
277 12c5bde7 2020-10-16 stsp sense = (cmp(f1, f1 + size) > 0);
278 12c5bde7 2020-10-16 stsp for (; f1 < last; sense = !sense) {
279 12c5bde7 2020-10-16 stsp length = 2;
280 12c5bde7 2020-10-16 stsp /* Find pairs with same sense. */
281 12c5bde7 2020-10-16 stsp for (f2 = f1 + size2; f2 < last; f2 += size2) {
282 12c5bde7 2020-10-16 stsp if ((cmp(f2, f2+ size) > 0) != sense)
284 12c5bde7 2020-10-16 stsp length += 2;
286 12c5bde7 2020-10-16 stsp if (length < THRESHOLD) { /* Pairwise merge */
288 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
289 12c5bde7 2020-10-16 stsp if (sense > 0)
290 12c5bde7 2020-10-16 stsp swap (f1, f1 + size);
291 12c5bde7 2020-10-16 stsp } while ((f1 += size2) < f2);
292 12c5bde7 2020-10-16 stsp } else { /* Natural merge */
294 12c5bde7 2020-10-16 stsp for (f2 = f1 + size2; f2 < l2; f2 += size2) {
295 12c5bde7 2020-10-16 stsp if ((cmp(f2-size, f2) > 0) != sense) {
296 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2) = f2 - list1 + list2;
297 12c5bde7 2020-10-16 stsp if (sense > 0)
298 12c5bde7 2020-10-16 stsp reverse(f1, f2-size);
302 12c5bde7 2020-10-16 stsp if (sense > 0)
303 12c5bde7 2020-10-16 stsp reverse (f1, f2-size);
305 12c5bde7 2020-10-16 stsp if (f2 < last || cmp(f2 - size, f2) > 0)
306 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2) = f2 - list1 + list2;
308 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2) = list2 + n*size;
311 12c5bde7 2020-10-16 stsp #else /* pairwise merge only. */
312 12c5bde7 2020-10-16 stsp for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
313 12c5bde7 2020-10-16 stsp p2 = *EVAL(p2) = p2 + size2;
314 12c5bde7 2020-10-16 stsp if (cmp (f1, f1 + size) > 0)
315 12c5bde7 2020-10-16 stsp swap(f1, f1 + size);
317 12c5bde7 2020-10-16 stsp #endif /* NATURAL */
321 12c5bde7 2020-10-16 stsp * This is to avoid out-of-bounds addresses in sorting the
322 12c5bde7 2020-10-16 stsp * last 4 elements.
324 12c5bde7 2020-10-16 stsp static void
325 12c5bde7 2020-10-16 stsp insertionsort(u_char *a, size_t n, size_t size,
326 12c5bde7 2020-10-16 stsp int (*cmp)(const void *, const void *))
328 12c5bde7 2020-10-16 stsp u_char *ai, *s, *t, *u, tmp;
331 12c5bde7 2020-10-16 stsp for (ai = a+size; --n >= 1; ai += size)
332 12c5bde7 2020-10-16 stsp for (t = ai; t > a; t -= size) {
333 12c5bde7 2020-10-16 stsp u = t - size;
334 12c5bde7 2020-10-16 stsp if (cmp(u, t) <= 0)
336 12c5bde7 2020-10-16 stsp swap(u, t);