xref: /freebsd/lib/libc/quad/muldi3.c (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
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 "quad.h"
37 
38 /*
39  * Multiply two quads.
40  *
41  * Our algorithm is based on the following.  Split incoming quad values
42  * u and v (where u,v >= 0) into
43  *
44  *	u = 2^n u1  *  u0	(n = number of bits in `u_long', usu. 32)
45  *
46  * and
47  *
48  *	v = 2^n v1  *  v0
49  *
50  * Then
51  *
52  *	uv = 2^2n u1 v1  +  2^n u1 v0  +  2^n v1 u0  +  u0 v0
53  *	   = 2^2n u1 v1  +     2^n (u1 v0 + v1 u0)   +  u0 v0
54  *
55  * Now add 2^n u1 v1 to the first term and subtract it from the middle,
56  * and add 2^n u0 v0 to the last term and subtract it from the middle.
57  * This gives:
58  *
59  *	uv = (2^2n + 2^n) (u1 v1)  +
60  *	         (2^n)    (u1 v0 - u1 v1 + u0 v1 - u0 v0)  +
61  *	       (2^n + 1)  (u0 v0)
62  *
63  * Factoring the middle a bit gives us:
64  *
65  *	uv = (2^2n + 2^n) (u1 v1)  +			[u1v1 = high]
66  *		 (2^n)    (u1 - u0) (v0 - v1)  +	[(u1-u0)... = mid]
67  *	       (2^n + 1)  (u0 v0)			[u0v0 = low]
68  *
69  * The terms (u1 v1), (u1 - u0) (v0 - v1), and (u0 v0) can all be done
70  * in just half the precision of the original.  (Note that either or both
71  * of (u1 - u0) or (v0 - v1) may be negative.)
72  *
73  * This algorithm is from Knuth vol. 2 (2nd ed), section 4.3.3, p. 278.
74  *
75  * Since C does not give us a `long * long = quad' operator, we split
76  * our input quads into two longs, then split the two longs into two
77  * shorts.  We can then calculate `short * short = long' in native
78  * arithmetic.
79  *
80  * Our product should, strictly speaking, be a `long quad', with 128
81  * bits, but we are going to discard the upper 64.  In other words,
82  * we are not interested in uv, but rather in (uv mod 2^2n).  This
83  * makes some of the terms above vanish, and we get:
84  *
85  *	(2^n)(high) + (2^n)(mid) + (2^n + 1)(low)
86  *
87  * or
88  *
89  *	(2^n)(high + mid + low) + low
90  *
91  * Furthermore, `high' and `mid' can be computed mod 2^n, as any factor
92  * of 2^n in either one will also vanish.  Only `low' need be computed
93  * mod 2^2n, and only because of the final term above.
94  */
95 static quad_t __lmulq(u_long, u_long);
96 
97 quad_t
98 __muldi3(quad_t a, quad_t b)
99 {
100 	union uu u, v, low, prod;
101 	u_long high, mid, udiff, vdiff;
102 	int negall, negmid;
103 #define	u1	u.ul[H]
104 #define	u0	u.ul[L]
105 #define	v1	v.ul[H]
106 #define	v0	v.ul[L]
107 
108 	/*
109 	 * Get u and v such that u, v >= 0.  When this is finished,
110 	 * u1, u0, v1, and v0 will be directly accessible through the
111 	 * longword fields.
112 	 */
113 	if (a >= 0)
114 		u.q = a, negall = 0;
115 	else
116 		u.q = -a, negall = 1;
117 	if (b >= 0)
118 		v.q = b;
119 	else
120 		v.q = -b, negall ^= 1;
121 
122 	if (u1 == 0 && v1 == 0) {
123 		/*
124 		 * An (I hope) important optimization occurs when u1 and v1
125 		 * are both 0.  This should be common since most numbers
126 		 * are small.  Here the product is just u0*v0.
127 		 */
128 		prod.q = __lmulq(u0, v0);
129 	} else {
130 		/*
131 		 * Compute the three intermediate products, remembering
132 		 * whether the middle term is negative.  We can discard
133 		 * any upper bits in high and mid, so we can use native
134 		 * u_long * u_long => u_long arithmetic.
135 		 */
136 		low.q = __lmulq(u0, v0);
137 
138 		if (u1 >= u0)
139 			negmid = 0, udiff = u1 - u0;
140 		else
141 			negmid = 1, udiff = u0 - u1;
142 		if (v0 >= v1)
143 			vdiff = v0 - v1;
144 		else
145 			vdiff = v1 - v0, negmid ^= 1;
146 		mid = udiff * vdiff;
147 
148 		high = u1 * v1;
149 
150 		/*
151 		 * Assemble the final product.
152 		 */
153 		prod.ul[H] = high + (negmid ? -mid : mid) + low.ul[L] +
154 		    low.ul[H];
155 		prod.ul[L] = low.ul[L];
156 	}
157 	return (negall ? -prod.q : prod.q);
158 #undef u1
159 #undef u0
160 #undef v1
161 #undef v0
162 }
163 
164 /*
165  * Multiply two 2N-bit longs to produce a 4N-bit quad, where N is half
166  * the number of bits in a long (whatever that is---the code below
167  * does not care as long as quad.h does its part of the bargain---but
168  * typically N==16).
169  *
170  * We use the same algorithm from Knuth, but this time the modulo refinement
171  * does not apply.  On the other hand, since N is half the size of a long,
172  * we can get away with native multiplication---none of our input terms
173  * exceeds (ULONG_MAX >> 1).
174  *
175  * Note that, for u_long l, the quad-precision result
176  *
177  *	l << N
178  *
179  * splits into high and low longs as HHALF(l) and LHUP(l) respectively.
180  */
181 static quad_t
182 __lmulq(u_long u, u_long v)
183 {
184 	u_long u1, u0, v1, v0, udiff, vdiff, high, mid, low;
185 	u_long prodh, prodl, was;
186 	union uu prod;
187 	int neg;
188 
189 	u1 = HHALF(u);
190 	u0 = LHALF(u);
191 	v1 = HHALF(v);
192 	v0 = LHALF(v);
193 
194 	low = u0 * v0;
195 
196 	/* This is the same small-number optimization as before. */
197 	if (u1 == 0 && v1 == 0)
198 		return (low);
199 
200 	if (u1 >= u0)
201 		udiff = u1 - u0, neg = 0;
202 	else
203 		udiff = u0 - u1, neg = 1;
204 	if (v0 >= v1)
205 		vdiff = v0 - v1;
206 	else
207 		vdiff = v1 - v0, neg ^= 1;
208 	mid = udiff * vdiff;
209 
210 	high = u1 * v1;
211 
212 	/* prod = (high << 2N) + (high << N); */
213 	prodh = high + HHALF(high);
214 	prodl = LHUP(high);
215 
216 	/* if (neg) prod -= mid << N; else prod += mid << N; */
217 	if (neg) {
218 		was = prodl;
219 		prodl -= LHUP(mid);
220 		prodh -= HHALF(mid) + (prodl > was);
221 	} else {
222 		was = prodl;
223 		prodl += LHUP(mid);
224 		prodh += HHALF(mid) + (prodl < was);
225 	}
226 
227 	/* prod += low << N */
228 	was = prodl;
229 	prodl += LHUP(low);
230 	prodh += HHALF(low) + (prodl < was);
231 	/* ... + low; */
232 	if ((prodl += low) < low)
233 		prodh++;
234 
235 	/* return 4N-bit product */
236 	prod.ul[H] = prodh;
237 	prod.ul[L] = prodl;
238 	return (prod.q);
239 }
240