xref: /openbsd-src/lib/libc/quad/qdivrem.c (revision 8500990981f885cbe5e6a4958549cacc238b5ae6)
1 /*-
2  * Copyright (c) 1992, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This software was developed by the Computer Systems Engineering group
6  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
7  * contributed to Berkeley.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 #if defined(LIBC_SCCS) && !defined(lint)
35 static char rcsid[] = "$OpenBSD: qdivrem.c,v 1.4 2003/06/02 20:18:36 millert Exp $";
36 #endif /* LIBC_SCCS and not lint */
37 
38 /*
39  * Multiprecision divide.  This algorithm is from Knuth vol. 2 (2nd ed),
40  * section 4.3.1, pp. 257--259.
41  */
42 
43 #include "quad.h"
44 
45 #define	B	((long)1 << HALF_BITS)	/* digit base */
46 
47 /* Combine two `digits' to make a single two-digit number. */
48 #define	COMBINE(a, b) (((u_long)(a) << HALF_BITS) | (b))
49 
50 /* select a type for digits in base B: use unsigned short if they fit */
51 #if ULONG_MAX == 0xffffffff && USHRT_MAX >= 0xffff
52 typedef unsigned short digit;
53 #else
54 typedef u_long digit;
55 #endif
56 
57 static void shl(digit *p, int len, int sh);
58 
59 /*
60  * __qdivrem(u, v, rem) returns u/v and, optionally, sets *rem to u%v.
61  *
62  * We do this in base 2-sup-HALF_BITS, so that all intermediate products
63  * fit within u_long.  As a consequence, the maximum length dividend and
64  * divisor are 4 `digits' in this base (they are shorter if they have
65  * leading zeros).
66  */
67 u_quad_t
68 __qdivrem(uq, vq, arq)
69 	u_quad_t uq, vq, *arq;
70 {
71 	union uu tmp;
72 	digit *u, *v, *q;
73 	register digit v1, v2;
74 	u_long qhat, rhat, t;
75 	int m, n, d, j, i;
76 	digit uspace[5], vspace[5], qspace[5];
77 
78 	/*
79 	 * Take care of special cases: divide by zero, and u < v.
80 	 */
81 	if (vq == 0) {
82 		/* divide by zero. */
83 		static volatile const unsigned int zero = 0;
84 
85 		tmp.ul[H] = tmp.ul[L] = 1 / zero;
86 		if (arq)
87 			*arq = uq;
88 		return (tmp.q);
89 	}
90 	if (uq < vq) {
91 		if (arq)
92 			*arq = uq;
93 		return (0);
94 	}
95 	u = &uspace[0];
96 	v = &vspace[0];
97 	q = &qspace[0];
98 
99 	/*
100 	 * Break dividend and divisor into digits in base B, then
101 	 * count leading zeros to determine m and n.  When done, we
102 	 * will have:
103 	 *	u = (u[1]u[2]...u[m+n]) sub B
104 	 *	v = (v[1]v[2]...v[n]) sub B
105 	 *	v[1] != 0
106 	 *	1 < n <= 4 (if n = 1, we use a different division algorithm)
107 	 *	m >= 0 (otherwise u < v, which we already checked)
108 	 *	m + n = 4
109 	 * and thus
110 	 *	m = 4 - n <= 2
111 	 */
112 	tmp.uq = uq;
113 	u[0] = 0;
114 	u[1] = HHALF(tmp.ul[H]);
115 	u[2] = LHALF(tmp.ul[H]);
116 	u[3] = HHALF(tmp.ul[L]);
117 	u[4] = LHALF(tmp.ul[L]);
118 	tmp.uq = vq;
119 	v[1] = HHALF(tmp.ul[H]);
120 	v[2] = LHALF(tmp.ul[H]);
121 	v[3] = HHALF(tmp.ul[L]);
122 	v[4] = LHALF(tmp.ul[L]);
123 	for (n = 4; v[1] == 0; v++) {
124 		if (--n == 1) {
125 			u_long rbj;	/* r*B+u[j] (not root boy jim) */
126 			digit q1, q2, q3, q4;
127 
128 			/*
129 			 * Change of plan, per exercise 16.
130 			 *	r = 0;
131 			 *	for j = 1..4:
132 			 *		q[j] = floor((r*B + u[j]) / v),
133 			 *		r = (r*B + u[j]) % v;
134 			 * We unroll this completely here.
135 			 */
136 			t = v[2];	/* nonzero, by definition */
137 			q1 = u[1] / t;
138 			rbj = COMBINE(u[1] % t, u[2]);
139 			q2 = rbj / t;
140 			rbj = COMBINE(rbj % t, u[3]);
141 			q3 = rbj / t;
142 			rbj = COMBINE(rbj % t, u[4]);
143 			q4 = rbj / t;
144 			if (arq)
145 				*arq = rbj % t;
146 			tmp.ul[H] = COMBINE(q1, q2);
147 			tmp.ul[L] = COMBINE(q3, q4);
148 			return (tmp.q);
149 		}
150 	}
151 
152 	/*
153 	 * By adjusting q once we determine m, we can guarantee that
154 	 * there is a complete four-digit quotient at &qspace[1] when
155 	 * we finally stop.
156 	 */
157 	for (m = 4 - n; u[1] == 0; u++)
158 		m--;
159 	for (i = 4 - m; --i >= 0;)
160 		q[i] = 0;
161 	q += 4 - m;
162 
163 	/*
164 	 * Here we run Program D, translated from MIX to C and acquiring
165 	 * a few minor changes.
166 	 *
167 	 * D1: choose multiplier 1 << d to ensure v[1] >= B/2.
168 	 */
169 	d = 0;
170 	for (t = v[1]; t < B / 2; t <<= 1)
171 		d++;
172 	if (d > 0) {
173 		shl(&u[0], m + n, d);		/* u <<= d */
174 		shl(&v[1], n - 1, d);		/* v <<= d */
175 	}
176 	/*
177 	 * D2: j = 0.
178 	 */
179 	j = 0;
180 	v1 = v[1];	/* for D3 -- note that v[1..n] are constant */
181 	v2 = v[2];	/* for D3 */
182 	do {
183 		register digit uj0, uj1, uj2;
184 
185 		/*
186 		 * D3: Calculate qhat (\^q, in TeX notation).
187 		 * Let qhat = min((u[j]*B + u[j+1])/v[1], B-1), and
188 		 * let rhat = (u[j]*B + u[j+1]) mod v[1].
189 		 * While rhat < B and v[2]*qhat > rhat*B+u[j+2],
190 		 * decrement qhat and increase rhat correspondingly.
191 		 * Note that if rhat >= B, v[2]*qhat < rhat*B.
192 		 */
193 		uj0 = u[j + 0];	/* for D3 only -- note that u[j+...] change */
194 		uj1 = u[j + 1];	/* for D3 only */
195 		uj2 = u[j + 2];	/* for D3 only */
196 		if (uj0 == v1) {
197 			qhat = B;
198 			rhat = uj1;
199 			goto qhat_too_big;
200 		} else {
201 			u_long n = COMBINE(uj0, uj1);
202 			qhat = n / v1;
203 			rhat = n % v1;
204 		}
205 		while (v2 * qhat > COMBINE(rhat, uj2)) {
206 	qhat_too_big:
207 			qhat--;
208 			if ((rhat += v1) >= B)
209 				break;
210 		}
211 		/*
212 		 * D4: Multiply and subtract.
213 		 * The variable `t' holds any borrows across the loop.
214 		 * We split this up so that we do not require v[0] = 0,
215 		 * and to eliminate a final special case.
216 		 */
217 		for (t = 0, i = n; i > 0; i--) {
218 			t = u[i + j] - v[i] * qhat - t;
219 			u[i + j] = LHALF(t);
220 			t = (B - HHALF(t)) & (B - 1);
221 		}
222 		t = u[j] - t;
223 		u[j] = LHALF(t);
224 		/*
225 		 * D5: test remainder.
226 		 * There is a borrow if and only if HHALF(t) is nonzero;
227 		 * in that (rare) case, qhat was too large (by exactly 1).
228 		 * Fix it by adding v[1..n] to u[j..j+n].
229 		 */
230 		if (HHALF(t)) {
231 			qhat--;
232 			for (t = 0, i = n; i > 0; i--) { /* D6: add back. */
233 				t += u[i + j] + v[i];
234 				u[i + j] = LHALF(t);
235 				t = HHALF(t);
236 			}
237 			u[j] = LHALF(u[j] + t);
238 		}
239 		q[j] = qhat;
240 	} while (++j <= m);		/* D7: loop on j. */
241 
242 	/*
243 	 * If caller wants the remainder, we have to calculate it as
244 	 * u[m..m+n] >> d (this is at most n digits and thus fits in
245 	 * u[m+1..m+n], but we may need more source digits).
246 	 */
247 	if (arq) {
248 		if (d) {
249 			for (i = m + n; i > m; --i)
250 				u[i] = (u[i] >> d) |
251 				    LHALF(u[i - 1] << (HALF_BITS - d));
252 			u[i] = 0;
253 		}
254 		tmp.ul[H] = COMBINE(uspace[1], uspace[2]);
255 		tmp.ul[L] = COMBINE(uspace[3], uspace[4]);
256 		*arq = tmp.q;
257 	}
258 
259 	tmp.ul[H] = COMBINE(qspace[1], qspace[2]);
260 	tmp.ul[L] = COMBINE(qspace[3], qspace[4]);
261 	return (tmp.q);
262 }
263 
264 /*
265  * Shift p[0]..p[len] left `sh' bits, ignoring any bits that
266  * `fall out' the left (there never will be any such anyway).
267  * We may assume len >= 0.  NOTE THAT THIS WRITES len+1 DIGITS.
268  */
269 static void
270 shl(register digit *p, register int len, register int sh)
271 {
272 	register int i;
273 
274 	for (i = 0; i < len; i++)
275 		p[i] = LHALF(p[i] << sh) | (p[i + 1] >> (HALF_BITS - sh));
276 	p[i] = LHALF(p[i] << sh);
277 }
278