xref: /freebsd/sys/libkern/qdivrem.c (revision 657729a89dd578d8cfc70d6616f5c65a48a8b33a)
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 software was developed by the Computer Systems Engineering group
8  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
9  * contributed to Berkeley.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 /*
40  * Multiprecision divide.  This algorithm is from Knuth vol. 2 (2nd ed),
41  * section 4.3.1, pp. 257--259.
42  */
43 
44 #include <libkern/quad.h>
45 
46 #define	B	(1 << HALF_BITS)	/* digit base */
47 
48 /* Combine two `digits' to make a single two-digit number. */
49 #define	COMBINE(a, b) (((u_long)(a) << HALF_BITS) | (b))
50 
51 /* select a type for digits in base B: use unsigned short if they fit */
52 #if ULONG_MAX == 0xffffffff && USHRT_MAX >= 0xffff
53 typedef unsigned short digit;
54 #else
55 typedef u_long digit;
56 #endif
57 
58 /*
59  * Shift p[0]..p[len] left `sh' bits, ignoring any bits that
60  * `fall out' the left (there never will be any such anyway).
61  * We may assume len >= 0.  NOTE THAT THIS WRITES len+1 DIGITS.
62  */
63 static void
64 __shl(digit *p, int len, int sh)
65 {
66 	int i;
67 
68 	for (i = 0; i < len; i++)
69 		p[i] = LHALF(p[i] << sh) | (p[i + 1] >> (HALF_BITS - sh));
70 	p[i] = LHALF(p[i] << sh);
71 }
72 
73 /*
74  * __qdivrem(u, v, rem) returns u/v and, optionally, sets *rem to u%v.
75  *
76  * We do this in base 2-sup-HALF_BITS, so that all intermediate products
77  * fit within u_long.  As a consequence, the maximum length dividend and
78  * divisor are 4 `digits' in this base (they are shorter if they have
79  * leading zeros).
80  */
81 u_quad_t
82 __qdivrem(u_quad_t uq, u_quad_t vq, u_quad_t *arq)
83 {
84 	union uu tmp;
85 	digit *u, *v, *q;
86 	digit v1, v2;
87 	u_long qhat, rhat, t;
88 	int m, n, d, j, i;
89 	digit uspace[5], vspace[5], qspace[5];
90 
91 	/*
92 	 * Take care of special cases: divide by zero, and u < v.
93 	 */
94 	if (__predict_false(vq == 0)) {
95 		/* divide by zero. */
96 		static volatile const unsigned int zero = 0;
97 
98 		tmp.ul[H] = tmp.ul[L] = 1 / zero;
99 		if (arq)
100 			*arq = uq;
101 		return (tmp.q);
102 	}
103 	if (uq < vq) {
104 		if (arq)
105 			*arq = uq;
106 		return (0);
107 	}
108 	u = &uspace[0];
109 	v = &vspace[0];
110 	q = &qspace[0];
111 
112 	/*
113 	 * Break dividend and divisor into digits in base B, then
114 	 * count leading zeros to determine m and n.  When done, we
115 	 * will have:
116 	 *	u = (u[1]u[2]...u[m+n]) sub B
117 	 *	v = (v[1]v[2]...v[n]) sub B
118 	 *	v[1] != 0
119 	 *	1 < n <= 4 (if n = 1, we use a different division algorithm)
120 	 *	m >= 0 (otherwise u < v, which we already checked)
121 	 *	m + n = 4
122 	 * and thus
123 	 *	m = 4 - n <= 2
124 	 */
125 	tmp.uq = uq;
126 	u[0] = 0;
127 	u[1] = HHALF(tmp.ul[H]);
128 	u[2] = LHALF(tmp.ul[H]);
129 	u[3] = HHALF(tmp.ul[L]);
130 	u[4] = LHALF(tmp.ul[L]);
131 	tmp.uq = vq;
132 	v[1] = HHALF(tmp.ul[H]);
133 	v[2] = LHALF(tmp.ul[H]);
134 	v[3] = HHALF(tmp.ul[L]);
135 	v[4] = LHALF(tmp.ul[L]);
136 	for (n = 4; v[1] == 0; v++) {
137 		if (--n == 1) {
138 			u_long rbj;	/* r*B+u[j] (not root boy jim) */
139 			digit q1, q2, q3, q4;
140 
141 			/*
142 			 * Change of plan, per exercise 16.
143 			 *	r = 0;
144 			 *	for j = 1..4:
145 			 *		q[j] = floor((r*B + u[j]) / v),
146 			 *		r = (r*B + u[j]) % v;
147 			 * We unroll this completely here.
148 			 */
149 			t = v[2];	/* nonzero, by definition */
150 			q1 = u[1] / t;
151 			rbj = COMBINE(u[1] % t, u[2]);
152 			q2 = rbj / t;
153 			rbj = COMBINE(rbj % t, u[3]);
154 			q3 = rbj / t;
155 			rbj = COMBINE(rbj % t, u[4]);
156 			q4 = rbj / t;
157 			if (arq)
158 				*arq = rbj % t;
159 			tmp.ul[H] = COMBINE(q1, q2);
160 			tmp.ul[L] = COMBINE(q3, q4);
161 			return (tmp.q);
162 		}
163 	}
164 
165 	/*
166 	 * By adjusting q once we determine m, we can guarantee that
167 	 * there is a complete four-digit quotient at &qspace[1] when
168 	 * we finally stop.
169 	 */
170 	for (m = 4 - n; u[1] == 0; u++)
171 		m--;
172 	for (i = 4 - m; --i >= 0;)
173 		q[i] = 0;
174 	q += 4 - m;
175 
176 	/*
177 	 * Here we run Program D, translated from MIX to C and acquiring
178 	 * a few minor changes.
179 	 *
180 	 * D1: choose multiplier 1 << d to ensure v[1] >= B/2.
181 	 */
182 	d = 0;
183 	for (t = v[1]; t < B / 2; t <<= 1)
184 		d++;
185 	if (d > 0) {
186 		__shl(&u[0], m + n, d);		/* u <<= d */
187 		__shl(&v[1], n - 1, d);		/* v <<= d */
188 	}
189 	/*
190 	 * D2: j = 0.
191 	 */
192 	j = 0;
193 	v1 = v[1];	/* for D3 -- note that v[1..n] are constant */
194 	v2 = v[2];	/* for D3 */
195 	do {
196 		digit uj0, uj1, uj2;
197 
198 		/*
199 		 * D3: Calculate qhat (\^q, in TeX notation).
200 		 * Let qhat = min((u[j]*B + u[j+1])/v[1], B-1), and
201 		 * let rhat = (u[j]*B + u[j+1]) mod v[1].
202 		 * While rhat < B and v[2]*qhat > rhat*B+u[j+2],
203 		 * decrement qhat and increase rhat correspondingly.
204 		 * Note that if rhat >= B, v[2]*qhat < rhat*B.
205 		 */
206 		uj0 = u[j + 0];	/* for D3 only -- note that u[j+...] change */
207 		uj1 = u[j + 1];	/* for D3 only */
208 		uj2 = u[j + 2];	/* for D3 only */
209 		if (uj0 == v1) {
210 			qhat = B;
211 			rhat = uj1;
212 			goto qhat_too_big;
213 		} else {
214 			u_long nn = COMBINE(uj0, uj1);
215 			qhat = nn / v1;
216 			rhat = nn % v1;
217 		}
218 		while (v2 * qhat > COMBINE(rhat, uj2)) {
219 	qhat_too_big:
220 			qhat--;
221 			if ((rhat += v1) >= B)
222 				break;
223 		}
224 		/*
225 		 * D4: Multiply and subtract.
226 		 * The variable `t' holds any borrows across the loop.
227 		 * We split this up so that we do not require v[0] = 0,
228 		 * and to eliminate a final special case.
229 		 */
230 		for (t = 0, i = n; i > 0; i--) {
231 			t = u[i + j] - v[i] * qhat - t;
232 			u[i + j] = LHALF(t);
233 			t = (B - HHALF(t)) & (B - 1);
234 		}
235 		t = u[j] - t;
236 		u[j] = LHALF(t);
237 		/*
238 		 * D5: test remainder.
239 		 * There is a borrow if and only if HHALF(t) is nonzero;
240 		 * in that (rare) case, qhat was too large (by exactly 1).
241 		 * Fix it by adding v[1..n] to u[j..j+n].
242 		 */
243 		if (HHALF(t)) {
244 			qhat--;
245 			for (t = 0, i = n; i > 0; i--) { /* D6: add back. */
246 				t += u[i + j] + v[i];
247 				u[i + j] = LHALF(t);
248 				t = HHALF(t);
249 			}
250 			u[j] = LHALF(u[j] + t);
251 		}
252 		q[j] = qhat;
253 	} while (++j <= m);		/* D7: loop on j. */
254 
255 	/*
256 	 * If caller wants the remainder, we have to calculate it as
257 	 * u[m..m+n] >> d (this is at most n digits and thus fits in
258 	 * u[m+1..m+n], but we may need more source digits).
259 	 */
260 	if (arq) {
261 		if (d) {
262 			for (i = m + n; i > m; --i)
263 				u[i] = (u[i] >> d) |
264 				    LHALF(u[i - 1] << (HALF_BITS - d));
265 			u[i] = 0;
266 		}
267 		tmp.ul[H] = COMBINE(uspace[1], uspace[2]);
268 		tmp.ul[L] = COMBINE(uspace[3], uspace[4]);
269 		*arq = tmp.q;
270 	}
271 
272 	tmp.ul[H] = COMBINE(qspace[1], qspace[2]);
273 	tmp.ul[L] = COMBINE(qspace[3], qspace[4]);
274 	return (tmp.q);
275 }
276