xref: /freebsd/sys/libkern/arm/muldi3.c (revision d37eb51047221dc3322b34db1038ff3aa533883f)
1 /*	$NetBSD: muldi3.c,v 1.8 2003/08/07 16:32:09 agc Exp $	*/
2 
3 /*-
4  * SPDX-License-Identifier: BSD-3-Clause
5  *
6  * Copyright (c) 1992, 1993
7  *	The Regents of the University of California.  All rights reserved.
8  *
9  * This software was developed by the Computer Systems Engineering group
10  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
11  * contributed to Berkeley.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  */
37 
38 #include <libkern/quad.h>
39 
40 /*
41  * Multiply two quads.
42  *
43  * Our algorithm is based on the following.  Split incoming quad values
44  * u and v (where u,v >= 0) into
45  *
46  *	u = 2^n u1  *  u0	(n = number of bits in `u_int', usu. 32)
47  *
48  * and
49  *
50  *	v = 2^n v1  *  v0
51  *
52  * Then
53  *
54  *	uv = 2^2n u1 v1  +  2^n u1 v0  +  2^n v1 u0  +  u0 v0
55  *	   = 2^2n u1 v1  +     2^n (u1 v0 + v1 u0)   +  u0 v0
56  *
57  * Now add 2^n u1 v1 to the first term and subtract it from the middle,
58  * and add 2^n u0 v0 to the last term and subtract it from the middle.
59  * This gives:
60  *
61  *	uv = (2^2n + 2^n) (u1 v1)  +
62  *	         (2^n)    (u1 v0 - u1 v1 + u0 v1 - u0 v0)  +
63  *	       (2^n + 1)  (u0 v0)
64  *
65  * Factoring the middle a bit gives us:
66  *
67  *	uv = (2^2n + 2^n) (u1 v1)  +			[u1v1 = high]
68  *		 (2^n)    (u1 - u0) (v0 - v1)  +	[(u1-u0)... = mid]
69  *	       (2^n + 1)  (u0 v0)			[u0v0 = low]
70  *
71  * The terms (u1 v1), (u1 - u0) (v0 - v1), and (u0 v0) can all be done
72  * in just half the precision of the original.  (Note that either or both
73  * of (u1 - u0) or (v0 - v1) may be negative.)
74  *
75  * This algorithm is from Knuth vol. 2 (2nd ed), section 4.3.3, p. 278.
76  *
77  * Since C does not give us a `int * int = quad' operator, we split
78  * our input quads into two ints, then split the two ints into two
79  * shorts.  We can then calculate `short * short = int' in native
80  * arithmetic.
81  *
82  * Our product should, strictly speaking, be a `long quad', with 128
83  * bits, but we are going to discard the upper 64.  In other words,
84  * we are not interested in uv, but rather in (uv mod 2^2n).  This
85  * makes some of the terms above vanish, and we get:
86  *
87  *	(2^n)(high) + (2^n)(mid) + (2^n + 1)(low)
88  *
89  * or
90  *
91  *	(2^n)(high + mid + low) + low
92  *
93  * Furthermore, `high' and `mid' can be computed mod 2^n, as any factor
94  * of 2^n in either one will also vanish.  Only `low' need be computed
95  * mod 2^2n, and only because of the final term above.
96  */
97 static quad_t __lmulq(u_int, u_int);
98 
99 quad_t __muldi3(quad_t, quad_t);
100 quad_t
101 __muldi3(quad_t a, quad_t b)
102 {
103 	union uu u, v, low, prod;
104 	u_int high, mid, udiff, vdiff;
105 	int negall, negmid;
106 #define	u1	u.ul[H]
107 #define	u0	u.ul[L]
108 #define	v1	v.ul[H]
109 #define	v0	v.ul[L]
110 
111 	/*
112 	 * Get u and v such that u, v >= 0.  When this is finished,
113 	 * u1, u0, v1, and v0 will be directly accessible through the
114 	 * int fields.
115 	 */
116 	if (a >= 0)
117 		u.q = a, negall = 0;
118 	else
119 		u.q = -a, negall = 1;
120 	if (b >= 0)
121 		v.q = b;
122 	else
123 		v.q = -b, negall ^= 1;
124 
125 	if (u1 == 0 && v1 == 0) {
126 		/*
127 		 * An (I hope) important optimization occurs when u1 and v1
128 		 * are both 0.  This should be common since most numbers
129 		 * are small.  Here the product is just u0*v0.
130 		 */
131 		prod.q = __lmulq(u0, v0);
132 	} else {
133 		/*
134 		 * Compute the three intermediate products, remembering
135 		 * whether the middle term is negative.  We can discard
136 		 * any upper bits in high and mid, so we can use native
137 		 * u_int * u_int => u_int arithmetic.
138 		 */
139 		low.q = __lmulq(u0, v0);
140 
141 		if (u1 >= u0)
142 			negmid = 0, udiff = u1 - u0;
143 		else
144 			negmid = 1, udiff = u0 - u1;
145 		if (v0 >= v1)
146 			vdiff = v0 - v1;
147 		else
148 			vdiff = v1 - v0, negmid ^= 1;
149 		mid = udiff * vdiff;
150 
151 		high = u1 * v1;
152 
153 		/*
154 		 * Assemble the final product.
155 		 */
156 		prod.ul[H] = high + (negmid ? -mid : mid) + low.ul[L] +
157 		    low.ul[H];
158 		prod.ul[L] = low.ul[L];
159 	}
160 	return (negall ? -prod.q : prod.q);
161 #undef u1
162 #undef u0
163 #undef v1
164 #undef v0
165 }
166 
167 /*
168  * Multiply two 2N-bit ints to produce a 4N-bit quad, where N is half
169  * the number of bits in an int (whatever that is---the code below
170  * does not care as long as quad.h does its part of the bargain---but
171  * typically N==16).
172  *
173  * We use the same algorithm from Knuth, but this time the modulo refinement
174  * does not apply.  On the other hand, since N is half the size of an int,
175  * we can get away with native multiplication---none of our input terms
176  * exceeds (UINT_MAX >> 1).
177  *
178  * Note that, for u_int l, the quad-precision result
179  *
180  *	l << N
181  *
182  * splits into high and low ints as HHALF(l) and LHUP(l) respectively.
183  */
184 static quad_t
185 __lmulq(u_int u, u_int v)
186 {
187 	u_int u1, u0, v1, v0, udiff, vdiff, high, mid, low;
188 	u_int prodh, prodl, was;
189 	union uu prod;
190 	int neg;
191 
192 	u1 = HHALF(u);
193 	u0 = LHALF(u);
194 	v1 = HHALF(v);
195 	v0 = LHALF(v);
196 
197 	low = u0 * v0;
198 
199 	/* This is the same small-number optimization as before. */
200 	if (u1 == 0 && v1 == 0)
201 		return (low);
202 
203 	if (u1 >= u0)
204 		udiff = u1 - u0, neg = 0;
205 	else
206 		udiff = u0 - u1, neg = 1;
207 	if (v0 >= v1)
208 		vdiff = v0 - v1;
209 	else
210 		vdiff = v1 - v0, neg ^= 1;
211 	mid = udiff * vdiff;
212 
213 	high = u1 * v1;
214 
215 	/* prod = (high << 2N) + (high << N); */
216 	prodh = high + HHALF(high);
217 	prodl = LHUP(high);
218 
219 	/* if (neg) prod -= mid << N; else prod += mid << N; */
220 	if (neg) {
221 		was = prodl;
222 		prodl -= LHUP(mid);
223 		prodh -= HHALF(mid) + (prodl > was);
224 	} else {
225 		was = prodl;
226 		prodl += LHUP(mid);
227 		prodh += HHALF(mid) + (prodl < was);
228 	}
229 
230 	/* prod += low << N */
231 	was = prodl;
232 	prodl += LHUP(low);
233 	prodh += HHALF(low) + (prodl < was);
234 	/* ... + low; */
235 	if ((prodl += low) < low)
236 		prodh++;
237 
238 	/* return 4N-bit product */
239 	prod.ul[H] = prodh;
240 	prod.ul[L] = prodl;
241 	return (prod.q);
242 }
243