xref: /freebsd/sys/libkern/arm/muldi3.c (revision 1669d8afc64812c8d2d1d147ae1fd42ff441e1b1)
1 /*	$NetBSD: muldi3.c,v 1.8 2003/08/07 16:32:09 agc Exp $	*/
2 
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 #if defined(LIBC_SCCS) && !defined(lint)
38 #if 0
39 static char sccsid[] = "@(#)muldi3.c	8.1 (Berkeley) 6/4/93";
40 #else
41 __FBSDID("$FreeBSD$");
42 #endif
43 #endif /* LIBC_SCCS and not lint */
44 
45 #include <libkern/quad.h>
46 
47 /*
48  * Multiply two quads.
49  *
50  * Our algorithm is based on the following.  Split incoming quad values
51  * u and v (where u,v >= 0) into
52  *
53  *	u = 2^n u1  *  u0	(n = number of bits in `u_int', usu. 32)
54  *
55  * and
56  *
57  *	v = 2^n v1  *  v0
58  *
59  * Then
60  *
61  *	uv = 2^2n u1 v1  +  2^n u1 v0  +  2^n v1 u0  +  u0 v0
62  *	   = 2^2n u1 v1  +     2^n (u1 v0 + v1 u0)   +  u0 v0
63  *
64  * Now add 2^n u1 v1 to the first term and subtract it from the middle,
65  * and add 2^n u0 v0 to the last term and subtract it from the middle.
66  * This gives:
67  *
68  *	uv = (2^2n + 2^n) (u1 v1)  +
69  *	         (2^n)    (u1 v0 - u1 v1 + u0 v1 - u0 v0)  +
70  *	       (2^n + 1)  (u0 v0)
71  *
72  * Factoring the middle a bit gives us:
73  *
74  *	uv = (2^2n + 2^n) (u1 v1)  +			[u1v1 = high]
75  *		 (2^n)    (u1 - u0) (v0 - v1)  +	[(u1-u0)... = mid]
76  *	       (2^n + 1)  (u0 v0)			[u0v0 = low]
77  *
78  * The terms (u1 v1), (u1 - u0) (v0 - v1), and (u0 v0) can all be done
79  * in just half the precision of the original.  (Note that either or both
80  * of (u1 - u0) or (v0 - v1) may be negative.)
81  *
82  * This algorithm is from Knuth vol. 2 (2nd ed), section 4.3.3, p. 278.
83  *
84  * Since C does not give us a `int * int = quad' operator, we split
85  * our input quads into two ints, then split the two ints into two
86  * shorts.  We can then calculate `short * short = int' in native
87  * arithmetic.
88  *
89  * Our product should, strictly speaking, be a `long quad', with 128
90  * bits, but we are going to discard the upper 64.  In other words,
91  * we are not interested in uv, but rather in (uv mod 2^2n).  This
92  * makes some of the terms above vanish, and we get:
93  *
94  *	(2^n)(high) + (2^n)(mid) + (2^n + 1)(low)
95  *
96  * or
97  *
98  *	(2^n)(high + mid + low) + low
99  *
100  * Furthermore, `high' and `mid' can be computed mod 2^n, as any factor
101  * of 2^n in either one will also vanish.  Only `low' need be computed
102  * mod 2^2n, and only because of the final term above.
103  */
104 static quad_t __lmulq(u_int, u_int);
105 
106 quad_t __muldi3(quad_t, quad_t);
107 quad_t
108 __muldi3(quad_t a, quad_t b)
109 {
110 	union uu u, v, low, prod;
111 	u_int high, mid, udiff, vdiff;
112 	int negall, negmid;
113 #define	u1	u.ul[H]
114 #define	u0	u.ul[L]
115 #define	v1	v.ul[H]
116 #define	v0	v.ul[L]
117 
118 	/*
119 	 * Get u and v such that u, v >= 0.  When this is finished,
120 	 * u1, u0, v1, and v0 will be directly accessible through the
121 	 * int fields.
122 	 */
123 	if (a >= 0)
124 		u.q = a, negall = 0;
125 	else
126 		u.q = -a, negall = 1;
127 	if (b >= 0)
128 		v.q = b;
129 	else
130 		v.q = -b, negall ^= 1;
131 
132 	if (u1 == 0 && v1 == 0) {
133 		/*
134 		 * An (I hope) important optimization occurs when u1 and v1
135 		 * are both 0.  This should be common since most numbers
136 		 * are small.  Here the product is just u0*v0.
137 		 */
138 		prod.q = __lmulq(u0, v0);
139 	} else {
140 		/*
141 		 * Compute the three intermediate products, remembering
142 		 * whether the middle term is negative.  We can discard
143 		 * any upper bits in high and mid, so we can use native
144 		 * u_int * u_int => u_int arithmetic.
145 		 */
146 		low.q = __lmulq(u0, v0);
147 
148 		if (u1 >= u0)
149 			negmid = 0, udiff = u1 - u0;
150 		else
151 			negmid = 1, udiff = u0 - u1;
152 		if (v0 >= v1)
153 			vdiff = v0 - v1;
154 		else
155 			vdiff = v1 - v0, negmid ^= 1;
156 		mid = udiff * vdiff;
157 
158 		high = u1 * v1;
159 
160 		/*
161 		 * Assemble the final product.
162 		 */
163 		prod.ul[H] = high + (negmid ? -mid : mid) + low.ul[L] +
164 		    low.ul[H];
165 		prod.ul[L] = low.ul[L];
166 	}
167 	return (negall ? -prod.q : prod.q);
168 #undef u1
169 #undef u0
170 #undef v1
171 #undef v0
172 }
173 
174 /*
175  * Multiply two 2N-bit ints to produce a 4N-bit quad, where N is half
176  * the number of bits in an int (whatever that is---the code below
177  * does not care as long as quad.h does its part of the bargain---but
178  * typically N==16).
179  *
180  * We use the same algorithm from Knuth, but this time the modulo refinement
181  * does not apply.  On the other hand, since N is half the size of an int,
182  * we can get away with native multiplication---none of our input terms
183  * exceeds (UINT_MAX >> 1).
184  *
185  * Note that, for u_int l, the quad-precision result
186  *
187  *	l << N
188  *
189  * splits into high and low ints as HHALF(l) and LHUP(l) respectively.
190  */
191 static quad_t
192 __lmulq(u_int u, u_int v)
193 {
194 	u_int u1, u0, v1, v0, udiff, vdiff, high, mid, low;
195 	u_int prodh, prodl, was;
196 	union uu prod;
197 	int neg;
198 
199 	u1 = HHALF(u);
200 	u0 = LHALF(u);
201 	v1 = HHALF(v);
202 	v0 = LHALF(v);
203 
204 	low = u0 * v0;
205 
206 	/* This is the same small-number optimization as before. */
207 	if (u1 == 0 && v1 == 0)
208 		return (low);
209 
210 	if (u1 >= u0)
211 		udiff = u1 - u0, neg = 0;
212 	else
213 		udiff = u0 - u1, neg = 1;
214 	if (v0 >= v1)
215 		vdiff = v0 - v1;
216 	else
217 		vdiff = v1 - v0, neg ^= 1;
218 	mid = udiff * vdiff;
219 
220 	high = u1 * v1;
221 
222 	/* prod = (high << 2N) + (high << N); */
223 	prodh = high + HHALF(high);
224 	prodl = LHUP(high);
225 
226 	/* if (neg) prod -= mid << N; else prod += mid << N; */
227 	if (neg) {
228 		was = prodl;
229 		prodl -= LHUP(mid);
230 		prodh -= HHALF(mid) + (prodl > was);
231 	} else {
232 		was = prodl;
233 		prodl += LHUP(mid);
234 		prodh += HHALF(mid) + (prodl < was);
235 	}
236 
237 	/* prod += low << N */
238 	was = prodl;
239 	prodl += LHUP(low);
240 	prodh += HHALF(low) + (prodl < was);
241 	/* ... + low; */
242 	if ((prodl += low) < low)
243 		prodh++;
244 
245 	/* return 4N-bit product */
246 	prod.ul[H] = prodh;
247 	prod.ul[L] = prodl;
248 	return (prod.q);
249 }
250