[unix-history] / usr / src / lib / libc / stdlib / merge.c

/*-
 * Copyright (c) 1992 The Regents of the University of California.
 * All rights reserved.
 *
 * This code is derived from software contributed to Berkeley by
 * Peter McIlroy.
 *
 * %sccs.include.redist.c%
 */

#if defined(LIBC_SCCS) && !defined(lint)
static char sccsid[] = "@(#)merge.c	5.2 (Berkeley) %G%";
#endif /* LIBC_SCCS and not lint */

/*
 * Hybrid exponential search/linear search merge sort with hybrid
 * natural/pairwise first pass.  Requires about .3% more comparisons
 * for random data than LSMS with pairwise first pass alone.
 * It works for objects as small as two bytes.
 */

#define NATURAL
#define THRESHOLD 16	/* Best choice for natural merge cut-off. */

/* #define NATURAL to get hybrid natural merge.
 * (The default is pairwise merging.)
 */

#include <sys/types.h>

#include <errno.h>
#include <stdlib.h>
#include <string.h>

static void setup __P((u_char *, u_char *, size_t, size_t, int (*)()));
static void insertionsort __P((u_char *, size_t, size_t, int (*)()));

#define ISIZE sizeof(int)
#define PSIZE sizeof(u_char *)
#define ICOPY_LIST(src, dst, last)				\
	do							\
	*(int*)dst = *(int*)src, src += ISIZE, dst += ISIZE;	\
	while(src < last)
#define ICOPY_ELT(src, dst, i)					\
	do							\
	*(int*) dst = *(int*) src, src += ISIZE, dst += ISIZE;	\
	while (i -= ISIZE)

#define CCOPY_LIST(src, dst, last)		\
	do					\
		*dst++ = *src++;		\
	while (src < last)
#define CCOPY_ELT(src, dst, i)			\
	do					\
		*dst++ = *src++;		\
	while (i -= 1)
		
/*
 * Find the next possible pointer head.  (Trickery for forcing an array
 * to do double duty as a linked list when objects do not align with word
 * boundaries.
 */
/* Assumption: PSIZE is a power of 2. */
#define EVAL(p) (u_char **)						\
	((u_char *)0 +							\
	    (((u_char *)p + PSIZE - 1 - (u_char *) 0) & ~(PSIZE - 1)))

/*
 * Arguments are as for qsort.
 */
int
mergesort(base, nmemb, size, cmp)
	void *base;
	size_t nmemb;
	register size_t size;
	int (*cmp) __P((const void *, const void *));
{
	register int i, sense;
	int big, iflag;
	register u_char *f1, *f2, *t, *b, *tp2, *q, *l1, *l2;
	u_char *list2, *list1, *p2, *p, *last, **p1;

	if (size < PSIZE / 2) {		/* Pointers must fit into 2 * size. */
		errno = EINVAL;
		return (-1);
	}

	/*
	 * XXX
	 * Stupid subtraction for the Cray.
	 */
	iflag = 0;
	if (!(size % ISIZE) && !(((char *)base - (char *)0) % ISIZE))
		iflag = 1;

	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
		return (-1);

	list1 = base;
	setup(list1, list2, nmemb, size, cmp);
	last = list2 + nmemb * size;
	i = big = 0;
	while (*EVAL(list2) != last) {
	    l2 = list1;
	    p1 = EVAL(list1);
	    for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
	    	p2 = *EVAL(p2);
	    	f1 = l2;
	    	f2 = l1 = list1 + (p2 - list2);
	    	if (p2 != last)
	    		p2 = *EVAL(p2);
	    	l2 = list1 + (p2 - list2);
	    	while (f1 < l1 && f2 < l2) {
	    		if ((*cmp)(f1, f2) <= 0) {
	    			q = f2;
	    			b = f1, t = l1;
	    			sense = -1;
	    		} else {
	    			q = f1;
	    			b = f2, t = l2;
	    			sense = 0;
	    		}
	    		if (!big) {	/* here i = 0 */
LINEAR:	    			while ((b += size) < t && cmp(q, b) >sense)
	    				if (++i == 6) {
	    					big = 1;
	    					goto EXPONENTIAL;
	    				}
	    		} else {
EXPONENTIAL:	    		for (i = size; ; i <<= 1)
	    				if ((p = (b + i)) >= t) {
	    					if ((p = t - size) > b &&
						    (*cmp)(q, p) <= sense)
	    						t = p;
	    					else
	    						b = p;
	    					break;
	    				} else if ((*cmp)(q, p) <= sense) {
	    					t = p;
	    					if (i == size)
	    						big = 0; 
	    					goto FASTCASE;
	    				} else
	    					b = p;
SLOWCASE:	    		while (t > b+size) {
	    				i = (((t - b) / size) >> 1) * size;
	    				if ((*cmp)(q, p = b + i) <= sense)
	    					t = p;
	    				else
	    					b = p;
	    			}
	    			goto COPY;
FASTCASE:	    		while (i > size)
	    				if ((*cmp)(q,
	    					p = b + (i >>= 1)) <= sense)
	    					t = p;
	    				else
	    					b = p;
COPY:	    			b = t;
	    		}
	    		i = size;
	    		if (q == f1) {
	    			if (iflag) {
	    				ICOPY_LIST(f2, tp2, b);
	    				ICOPY_ELT(f1, tp2, i);
	    			} else {
	    				CCOPY_LIST(f2, tp2, b);
	    				CCOPY_ELT(f1, tp2, i);
	    			}
	    		} else {
	    			if (iflag) {
	    				ICOPY_LIST(f1, tp2, b);
	    				ICOPY_ELT(f2, tp2, i);
	    			} else {
	    				CCOPY_LIST(f1, tp2, b);
	    				CCOPY_ELT(f2, tp2, i);
	    			}
	    		}
	    	}
	    	if (f2 < l2) {
	    		if (iflag)
	    			ICOPY_LIST(f2, tp2, l2);
	    		else
	    			CCOPY_LIST(f2, tp2, l2);
	    	} else if (f1 < l1) {
	    		if (iflag)
	    			ICOPY_LIST(f1, tp2, l1);
	    		else
	    			CCOPY_LIST(f1, tp2, l1);
	    	}
	    	*p1 = l2;
	    }
	    tp2 = list1;	/* swap list1, list2 */
	    list1 = list2;
	    list2 = tp2;
	    last = list2 + nmemb*size;
	}
	if (base == list2) {
		memmove(list2, list1, nmemb*size);
		list2 = list1;
	}
	free(list2);
	return (0);
}

#define	swap(a, b) {					\
		s = b;					\
		i = size;				\
		do {					\
			tmp = *a; *a++ = *s; *s++ = tmp; \
		} while (--i);				\
		a -= size;				\
	}
#define reverse(bot, top) {				\
	s = top;					\
	do {						\
		i = size;				\
		do {					\
			tmp = *bot; *bot++ = *s; *s++ = tmp; \
		} while (--i);				\
		s -= size2;				\
	} while(bot < s);				\
}

/*
 * Optional hybrid natural/pairwise first pass.  Eats up list1 in runs of
 * increasing order, list2 in a corresponding linked list.  Checks for runs
 * when THRESHOLD/2 pairs compare with same sense.  (Only used when NATURAL
 * is defined.  Otherwise simple pairwise merging is used.)
 */
void
setup(list1, list2, n, size, cmp)
	size_t n, size;
	int (*cmp) __P((const void *, const void *));
	u_char *list1, *list2;
{
	int i, length, size2, tmp, sense;
	u_char *f1, *f2, *s, *l2, *last, *p2;

	size2 = size*2;
	if (n <= 5) {
		insertionsort(list1, n, size, cmp);
		*EVAL(list2) = (u_char*) list2 + n*size;
		return;
	}
	/*
	 * Avoid running pointers out of bounds; limit n to evens
	 * for simplicity.
	 */
	i = 4 + (n & 1);
	insertionsort(list1 + (n - i) * size, i, size, cmp);
	last = list1 + size * (n - i);
	*EVAL(list2 + (last - list1)) = list2 + n * size;

#ifdef NATURAL
	p2 = list2;
	f1 = list1;
	sense = (cmp(f1, f1 + size) > 0);
	for (; f1 < last; sense = !sense) {
		length = 2;
					/* Find pairs with same sense. */
		for (f2 = f1 + size2; f2 < last; f2 += size2) {
			if ((cmp(f2, f2+ size) > 0) != sense)
				break;
			length += 2;
		}
		if (length < THRESHOLD) {		/* Pairwise merge */
			do {
				p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
				if (sense > 0)
					swap (f1, f1 + size);
			} while ((f1 += size2) < f2);
		} else {				/* Natural merge */
			l2 = f2;
			for (f2 = f1 + size2; f2 < l2; f2 += size2) {
				if ((cmp(f2-size, f2) > 0) != sense) {
					p2 = *EVAL(p2) = f2 - list1 + list2;
					if (sense > 0)
						reverse(f1, f2-size);
					f1 = f2;
				}
			}
			if (sense > 0)
				reverse (f1, f2-size);
			f1 = f2;
			if (f2 < last || cmp(f2, f2 + size) > 0)
				p2 = *EVAL(p2) = f2 - list1 + list2;
			else
				p2 = *EVAL(p2) = list2 + n*size;
		}
	}
#else		/* pairwise merge only. */
	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
		p2 = *EVAL(p2) = p2 + size2;
		if (cmp (f1, f1 + size) > 0)
			swap(f1, f1 + size);
	}
#endif /* NATURAL */
}

/*
 * This is to avoid out-of-bounds addresses in sorting the
 * last 4 elements.
 */
static void
insertionsort(a, n, size, cmp)
	u_char *a;
	size_t n, size;
	int (*cmp) __P((const void *, const void *));
{
	u_char *ai, *s, *t, *u, tmp;
	int i;

	for (ai = a+size; --n >= 1; ai += size)
		for (t = ai; t > a; t -= size) {
			u = t - size;
			if (cmp(u, t) <= 0)
				break;
			swap(u, t);
		}
}
Commit	Line	Data
ad59b6a4 KB	1	/*-
	2	* Copyright (c) 1992 The Regents of the University of California.
	3	* All rights reserved.
	4	*
	5	* This code is derived from software contributed to Berkeley by
	6	* Peter McIlroy.
	7	*
	8	* %sccs.include.redist.c%
	9	*/
	10
	11	#if defined(LIBC_SCCS) && !defined(lint)
6f40f148	12	static char sccsid[] = "@(#)merge.c 5.2 (Berkeley) %G%";
ad59b6a4 KB	13	#endif /* LIBC_SCCS and not lint */
	14
	15	/*
	16	* Hybrid exponential search/linear search merge sort with hybrid
	17	* natural/pairwise first pass. Requires about .3% more comparisons
	18	* for random data than LSMS with pairwise first pass alone.
	19	* It works for objects as small as two bytes.
	20	*/
	21
	22	#define NATURAL
	23	#define THRESHOLD 16 /* Best choice for natural merge cut-off. */
	24
	25	/* #define NATURAL to get hybrid natural merge.
	26	* (The default is pairwise merging.)
	27	*/
	28
	29	#include <sys/types.h>
	30
	31	#include <errno.h>
	32	#include <stdlib.h>
	33	#include <string.h>
	34
	35	static void setup __P((u_char , u_char , size_t, size_t, int (*)()));
	36	static void insertionsort __P((u_char , size_t, size_t, int ()()));
	37
	38	#define ISIZE sizeof(int)
	39	#define PSIZE sizeof(u_char *)
	40	#define ICOPY_LIST(src, dst, last) \
	41	do \
	42	(int)dst = (int)src, src += ISIZE, dst += ISIZE; \
	43	while(src < last)
	44	#define ICOPY_ELT(src, dst, i) \
	45	do \
	46	(int) dst = (int) src, src += ISIZE, dst += ISIZE; \
	47	while (i -= ISIZE)
	48
	49	#define CCOPY_LIST(src, dst, last) \
	50	do \
	51	dst++ = src++; \
	52	while (src < last)
	53	#define CCOPY_ELT(src, dst, i) \
	54	do \
	55	dst++ = src++; \
	56	while (i -= 1)
	57
	58	/*
	59	* Find the next possible pointer head. (Trickery for forcing an array
	60	* to do double duty as a linked list when objects do not align with word
	61	* boundaries.
	62	*/
	63	/* Assumption: PSIZE is a power of 2. */
	64	#define EVAL(p) (u_char **) \
	65	((u_char *)0 + \
	66	(((u_char )p + PSIZE - 1 - (u_char ) 0) & ~(PSIZE - 1)))
	67
	68	/*
	69	* Arguments are as for qsort.
	70	*/
	71	int
	72	mergesort(base, nmemb, size, cmp)
	73	void *base;
	74	size_t nmemb;
	75	register size_t size;
	76	int (cmp) __P((const void , const void *));
77	{
78	register int i, sense;
79	int big, iflag;
80	register u_char f1, f2, t, b, tp2, q, l1, l2;
81	u_char list2, list1, p2, p, last, *p1;
82
6f40f148	83	if (size < PSIZE / 2) { /* Pointers must fit into 2 * size. */
ad59b6a4 KB	84	errno = EINVAL;
	85	return (-1);
	86	}
	87
	88	/*
	89	* XXX
	90	* Stupid subtraction for the Cray.
	91	*/
	92	iflag = 0;
	93	if (!(size % ISIZE) && !(((char )base - (char )0) % ISIZE))
	94	iflag = 1;
	95
	96	if ((list2 = malloc(nmemb * size + PSIZE)) == NULL)
	97	return (-1);
	98
	99	list1 = base;
	100	setup(list1, list2, nmemb, size, cmp);
	101	last = list2 + nmemb * size;
	102	i = big = 0;
	103	while (*EVAL(list2) != last) {
	104	l2 = list1;
	105	p1 = EVAL(list1);
	106	for (tp2 = p2 = list2; p2 != last; p1 = EVAL(l2)) {
	107	p2 = *EVAL(p2);
	108	f1 = l2;
	109	f2 = l1 = list1 + (p2 - list2);
	110	if (p2 != last)
	111	p2 = *EVAL(p2);
	112	l2 = list1 + (p2 - list2);
	113	while (f1 < l1 && f2 < l2) {
	114	if ((*cmp)(f1, f2) <= 0) {
	115	q = f2;
	116	b = f1, t = l1;
	117	sense = -1;
	118	} else {
	119	q = f1;
	120	b = f2, t = l2;
	121	sense = 0;
	122	}
	123	if (!big) { /* here i = 0 */
	124	LINEAR: while ((b += size) < t && cmp(q, b) >sense)
	125	if (++i == 6) {
	126	big = 1;
	127	goto EXPONENTIAL;
	128	}
	129	} else {
	130	EXPONENTIAL: for (i = size; ; i <<= 1)
	131	if ((p = (b + i)) >= t) {
	132	if ((p = t - size) > b &&
	133	(*cmp)(q, p) <= sense)
	134	t = p;
	135	else
	136	b = p;
	137	break;
	138	} else if ((*cmp)(q, p) <= sense) {
	139	t = p;
	140	if (i == size)
	141	big = 0;
	142	goto FASTCASE;
	143	} else
	144	b = p;
	145	SLOWCASE: while (t > b+size) {
	146	i = (((t - b) / size) >> 1) * size;
	147	if ((*cmp)(q, p = b + i) <= sense)
148	t = p;
149	else
150	b = p;
151	}
152	goto COPY;
153	FASTCASE: while (i > size)
154	if ((*cmp)(q,
155	p = b + (i >>= 1)) <= sense)
156	t = p;
157	else
158	b = p;
159	COPY: b = t;
160	}
161	i = size;
162	if (q == f1) {
163	if (iflag) {
164	ICOPY_LIST(f2, tp2, b);
165	ICOPY_ELT(f1, tp2, i);
166	} else {
167	CCOPY_LIST(f2, tp2, b);
168	CCOPY_ELT(f1, tp2, i);
169	}
170	} else {
171	if (iflag) {
172	ICOPY_LIST(f1, tp2, b);
173	ICOPY_ELT(f2, tp2, i);
174	} else {
175	CCOPY_LIST(f1, tp2, b);
176	CCOPY_ELT(f2, tp2, i);
177	}
178	}
179	}
180	if (f2 < l2) {
181	if (iflag)
182	ICOPY_LIST(f2, tp2, l2);
183	else
184	CCOPY_LIST(f2, tp2, l2);
185	} else if (f1 < l1) {
186	if (iflag)
187	ICOPY_LIST(f1, tp2, l1);
188	else
189	CCOPY_LIST(f1, tp2, l1);
190	}
191	*p1 = l2;
192	}
193	tp2 = list1; /* swap list1, list2 */
194	list1 = list2;
195	list2 = tp2;
196	last = list2 + nmemb*size;
197	}
198	if (base == list2) {
199	memmove(list2, list1, nmemb*size);
200	list2 = list1;
201	}
202	free(list2);
203	return (0);
204	}
205
206	#define swap(a, b) { \
207	s = b; \
208	i = size; \
209	do { \
210	tmp = a; a++ = s; s++ = tmp; \
211	} while (--i); \
212	a -= size; \
213	}
214	#define reverse(bot, top) { \
215	s = top; \
216	do { \
217	i = size; \
218	do { \
219	tmp = bot; bot++ = s; s++ = tmp; \
220	} while (--i); \
221	s -= size2; \
222	} while(bot < s); \
223	}
224
225	/*
226	* Optional hybrid natural/pairwise first pass. Eats up list1 in runs of
227	* increasing order, list2 in a corresponding linked list. Checks for runs
228	* when THRESHOLD/2 pairs compare with same sense. (Only used when NATURAL
229	* is defined. Otherwise simple pairwise merging is used.)
230	*/
231	void
232	setup(list1, list2, n, size, cmp)
233	size_t n, size;
234	int (cmp) __P((const void , const void *));
235	u_char list1, list2;
236	{
237	int i, length, size2, tmp, sense;
238	u_char f1, f2, s, l2, last, p2;
239
240	size2 = size*2;
241	if (n <= 5) {
242	insertionsort(list1, n, size, cmp);
243	EVAL(list2) = (u_char) list2 + n*size;
244	return;
245	}
246	/*
247	* Avoid running pointers out of bounds; limit n to evens
248	* for simplicity.
249	*/
250	i = 4 + (n & 1);
251	insertionsort(list1 + (n - i) * size, i, size, cmp);
252	last = list1 + size * (n - i);
253	EVAL(list2 + (last - list1)) = list2 + n size;
254
255	#ifdef NATURAL
256	p2 = list2;
257	f1 = list1;
258	sense = (cmp(f1, f1 + size) > 0);
259	for (; f1 < last; sense = !sense) {
260	length = 2;
261	/* Find pairs with same sense. */
262	for (f2 = f1 + size2; f2 < last; f2 += size2) {
263	if ((cmp(f2, f2+ size) > 0) != sense)
264	break;
265	length += 2;
266	}
267	if (length < THRESHOLD) { /* Pairwise merge */
268	do {
269	p2 = *EVAL(p2) = f1 + size2 - list1 + list2;
270	if (sense > 0)
271	swap (f1, f1 + size);
272	} while ((f1 += size2) < f2);
273	} else { /* Natural merge */
274	l2 = f2;
275	for (f2 = f1 + size2; f2 < l2; f2 += size2) {
276	if ((cmp(f2-size, f2) > 0) != sense) {
277	p2 = *EVAL(p2) = f2 - list1 + list2;
278	if (sense > 0)
279	reverse(f1, f2-size);
280	f1 = f2;
281	}
282	}
283	if (sense > 0)
284	reverse (f1, f2-size);
285	f1 = f2;
286	if (f2 < last \|\| cmp(f2, f2 + size) > 0)
287	p2 = *EVAL(p2) = f2 - list1 + list2;
288	else
289	p2 = EVAL(p2) = list2 + nsize;
290	}
291	}
292	#else /* pairwise merge only. */
293	for (f1 = list1, p2 = list2; f1 < last; f1 += size2) {
294	p2 = *EVAL(p2) = p2 + size2;
295	if (cmp (f1, f1 + size) > 0)
296	swap(f1, f1 + size);
297	}
298	#endif /* NATURAL */
299	}
300
301	/*
302	* This is to avoid out-of-bounds addresses in sorting the
303	* last 4 elements.
304	*/
305	static void
306	insertionsort(a, n, size, cmp)
307	u_char *a;
308	size_t n, size;
309	int (cmp) __P((const void , const void *));
310	{
311	u_char ai, s, t, u, tmp;
312	int i;
313
314	for (ai = a+size; --n >= 1; ai += size)
315	for (t = ai; t > a; t -= size) {
316	u = t - size;
317	if (cmp(u, t) <= 0)
318	break;
319	swap(u, t);
320	}
321	}