xref: /csrg-svn/lib/libc/stdlib/radixsort.c (revision 45935)
1 /*-
2  * Copyright (c) 1990 The Regents of the University of California.
3  * All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  */
7 
8 #if defined(LIBC_SCCS) && !defined(lint)
9 static char sccsid[] = "@(#)radixsort.c	5.6 (Berkeley) 01/13/91";
10 #endif /* LIBC_SCCS and not lint */
11 
12 #include <sys/types.h>
13 #include <limits.h>
14 #include <stdlib.h>
15 #include <stddef.h>
16 
17 /*
18  * __rspartition is the cutoff point for a further partitioning instead
19  * of a shellsort.  If it changes check __rsshell_increments.  Both of
20  * these are exported, as the best values are data dependent.
21  */
22 #define	NPARTITION	40
23 int __rspartition = NPARTITION;
24 int __rsshell_increments[] = { 4, 1, 0, 0, 0, 0, 0, 0 };
25 
26 /*
27  * Stackp points to context structures, where each structure schedules a
28  * partitioning.  Radixsort exits when the stack is empty.
29  *
30  * If the buckets are placed on the stack randomly, the worst case is when
31  * all the buckets but one contain (npartitions + 1) elements and the bucket
32  * pushed on the stack last contains the rest of the elements.  In this case,
33  * stack growth is bounded by:
34  *
35  *	limit = (nelements / (npartitions + 1)) - 1;
36  *
37  * This is a very large number, 52,377,648 for the maximum 32-bit signed int.
38  *
39  * By forcing the largest bucket to be pushed on the stack first, the worst
40  * case is when all but two buckets each contain (npartitions + 1) elements,
41  * with the remaining elements split equally between the first and last
42  * buckets pushed on the stack.  In this case, stack growth is bounded when:
43  *
44  *	for (partition_cnt = 0; nelements > npartitions; ++partition_cnt)
45  *		nelements =
46  *		    (nelements - (npartitions + 1) * (nbuckets - 2)) / 2;
47  * The bound is:
48  *
49  *	limit = partition_cnt * (nbuckets - 1);
50  *
51  * This is a much smaller number, 4590 for the maximum 32-bit signed int.
52  */
53 #define	NBUCKETS	(UCHAR_MAX + 1)
54 
55 typedef struct _stack {
56 	u_char **bot;
57 	int indx, nmemb;
58 } CONTEXT;
59 
60 #define	STACKPUSH { \
61 	stackp->bot = p; \
62 	stackp->nmemb = nmemb; \
63 	stackp->indx = indx; \
64 	++stackp; \
65 }
66 #define	STACKPOP { \
67 	if (stackp == stack) \
68 		break; \
69 	--stackp; \
70 	bot = stackp->bot; \
71 	nmemb = stackp->nmemb; \
72 	indx = stackp->indx; \
73 }
74 
75 /*
76  * A variant of MSD radix sorting; see Knuth Vol. 3, page 177, and 5.2.5,
77  * Ex. 10 and 12.  Also, "Three Partition Refinement Algorithms, Paige
78  * and Tarjan, SIAM J. Comput. Vol. 16, No. 6, December 1987.
79  *
80  * This uses a simple sort as soon as a bucket crosses a cutoff point,
81  * rather than sorting the entire list after partitioning is finished.
82  * This should be an advantage.
83  *
84  * This is pure MSD instead of LSD of some number of MSD, switching to
85  * the simple sort as soon as possible.  Takes linear time relative to
86  * the number of bytes in the strings.
87  */
88 radixsort(l1, nmemb, tab, endbyte)
89 	u_char **l1, *tab, endbyte;
90 	register int nmemb;
91 {
92 	register int i, indx, t1, t2;
93 	register u_char **l2, **p, **bot, *tr;
94 	CONTEXT *stack, *stackp;
95 	int c[NBUCKETS + 1], max;
96 	u_char ltab[NBUCKETS];
97 	static void shellsort();
98 
99 	if (nmemb <= 1)
100 		return(0);
101 
102 	/*
103 	 * T1 is the constant part of the equation, the number of elements
104 	 * represented on the stack between the top and bottom entries.
105 	 * It doesn't get rounded as the divide by 2 rounds down (correct
106 	 * for a value being subtracted).  T2, the nelem value, has to be
107 	 * rounded up before each divide because we want an upper bound;
108 	 * this could overflow if nmemb is the maximum int.
109 	 */
110 	t1 = ((__rspartition + 1) * (NBUCKETS - 2)) >> 1;
111 	for (i = 0, t2 = nmemb; t2 > __rspartition; i += NBUCKETS - 1)
112 		t2 = ((t2 + 1) >> 1) - t1;
113 	if (i) {
114 		if (!(stack = stackp = (CONTEXT *)malloc(i * sizeof(CONTEXT))))
115 			return(-1);
116 	} else
117 		stack = stackp = NULL;
118 
119 	/*
120 	 * There are two arrays, one provided by the user (l1), and the
121 	 * temporary one (l2).  The data is sorted to the temporary stack,
122 	 * and then copied back.  The speedup of using index to determine
123 	 * which stack the data is on and simply swapping stacks back and
124 	 * forth, thus avoiding the copy every iteration, turns out to not
125 	 * be any faster than the current implementation.
126 	 */
127 	if (!(l2 = (u_char **)malloc(sizeof(u_char *) * nmemb)))
128 		return(-1);
129 
130 	/*
131 	 * Tr references a table of sort weights; multiple entries may
132 	 * map to the same weight; EOS char must have the lowest weight.
133 	 */
134 	if (tab)
135 		tr = tab;
136 	else {
137 		tr = ltab;
138 		for (t1 = 0, t2 = endbyte; t1 < t2; ++t1)
139 			tr[t1] = t1 + 1;
140 		tr[t2] = 0;
141 		for (t1 = endbyte + 1; t1 < NBUCKETS; ++t1)
142 			tr[t1] = t1;
143 	}
144 
145 	/* First sort is entire stack */
146 	bot = l1;
147 	indx = 0;
148 
149 	for (;;) {
150 		/* Clear bucket count array */
151 		bzero((char *)c, sizeof(c));
152 
153 		/*
154 		 * Compute number of items that sort to the same bucket
155 		 * for this index.
156 		 */
157 		for (p = bot, i = nmemb; --i >= 0;)
158 			++c[tr[(*p++)[indx]]];
159 
160 		/*
161 		 * Sum the number of characters into c, dividing the temp
162 		 * stack into the right number of buckets for this bucket,
163 		 * this index.  C contains the cumulative total of keys
164 		 * before and included in this bucket, and will later be
165 		 * used as an index to the bucket.  c[NBUCKETS] contains
166 		 * the total number of elements, for determining how many
167 		 * elements the last bucket contains.  At the same time
168 		 * find the largest bucket so it gets pushed first.
169 		 */
170 		for (i = max = t1 = 0, t2 = __rspartition; i <= NBUCKETS; ++i) {
171 			if (c[i] > t2) {
172 				t2 = c[i];
173 				max = i;
174 			}
175 			t1 = c[i] += t1;
176 		}
177 
178 		/*
179 		 * Partition the elements into buckets; c decrements through
180 		 * the bucket, and ends up pointing to the first element of
181 		 * the bucket.
182 		 */
183 		for (i = nmemb; --i >= 0;) {
184 			--p;
185 			l2[--c[tr[(*p)[indx]]]] = *p;
186 		}
187 
188 		/* Copy the partitioned elements back to user stack */
189 		bcopy(l2, bot, nmemb * sizeof(u_char *));
190 
191 		++indx;
192 		/*
193 		 * Sort buckets as necessary; don't sort c[0], it's the
194 		 * EOS character bucket, and nothing can follow EOS.
195 		 */
196 		for (i = max; i; --i) {
197 			if ((nmemb = c[i + 1] - (t1 = c[i])) < 2)
198 				continue;
199 			p = bot + t1;
200 			if (nmemb > __rspartition)
201 				STACKPUSH
202 			else
203 				shellsort(p, indx, nmemb, tr);
204 		}
205 		for (i = max + 1; i < NBUCKETS; ++i) {
206 			if ((nmemb = c[i + 1] - (t1 = c[i])) < 2)
207 				continue;
208 			p = bot + t1;
209 			if (nmemb > __rspartition)
210 				STACKPUSH
211 			else
212 				shellsort(p, indx, nmemb, tr);
213 		}
214 		/* Break out when stack is empty */
215 		STACKPOP
216 	}
217 
218 	free((char *)l2);
219 	free((char *)stack);
220 	return(0);
221 }
222 
223 /*
224  * Shellsort (diminishing increment sort) from Data Structures and
225  * Algorithms, Aho, Hopcraft and Ullman, 1983 Edition, page 290;
226  * see also Knuth Vol. 3, page 84.  The increments are selected from
227  * formula (8), page 95.  Roughly O(N^3/2).
228  */
229 static void
230 shellsort(p, indx, nmemb, tr)
231 	register u_char **p, *tr;
232 	register int indx, nmemb;
233 {
234 	register u_char ch, *s1, *s2;
235 	register int incr, *incrp, t1, t2;
236 
237 	for (incrp = __rsshell_increments; incr = *incrp++;)
238 		for (t1 = incr; t1 < nmemb; ++t1)
239 			for (t2 = t1 - incr; t2 >= 0;) {
240 				s1 = p[t2] + indx;
241 				s2 = p[t2 + incr] + indx;
242 				while ((ch = tr[*s1++]) == tr[*s2] && ch)
243 					++s2;
244 				if (ch > tr[*s2]) {
245 					s1 = p[t2];
246 					p[t2] = p[t2 + incr];
247 					p[t2 + incr] = s1;
248 					t2 -= incr;
249 				} else
250 					break;
251 			}
252 }
253