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. All advertising materials mentioning features or use of this software 18 * must display the following acknowledgement: 19 * This product includes software developed by the University of 20 * California, Berkeley and its contributors. 21 * 4. 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 #if defined(LIBC_SCCS) && !defined(lint) 39 static char sccsid[] = "@(#)muldi3.c 8.1 (Berkeley) 6/4/93"; 40 #endif /* LIBC_SCCS and not lint */ 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include "quad.h" 45 46 /* 47 * Multiply two quads. 48 * 49 * Our algorithm is based on the following. Split incoming quad values 50 * u and v (where u,v >= 0) into 51 * 52 * u = 2^n u1 * u0 (n = number of bits in `u_long', usu. 32) 53 * 54 * and 55 * 56 * v = 2^n v1 * v0 57 * 58 * Then 59 * 60 * uv = 2^2n u1 v1 + 2^n u1 v0 + 2^n v1 u0 + u0 v0 61 * = 2^2n u1 v1 + 2^n (u1 v0 + v1 u0) + u0 v0 62 * 63 * Now add 2^n u1 v1 to the first term and subtract it from the middle, 64 * and add 2^n u0 v0 to the last term and subtract it from the middle. 65 * This gives: 66 * 67 * uv = (2^2n + 2^n) (u1 v1) + 68 * (2^n) (u1 v0 - u1 v1 + u0 v1 - u0 v0) + 69 * (2^n + 1) (u0 v0) 70 * 71 * Factoring the middle a bit gives us: 72 * 73 * uv = (2^2n + 2^n) (u1 v1) + [u1v1 = high] 74 * (2^n) (u1 - u0) (v0 - v1) + [(u1-u0)... = mid] 75 * (2^n + 1) (u0 v0) [u0v0 = low] 76 * 77 * The terms (u1 v1), (u1 - u0) (v0 - v1), and (u0 v0) can all be done 78 * in just half the precision of the original. (Note that either or both 79 * of (u1 - u0) or (v0 - v1) may be negative.) 80 * 81 * This algorithm is from Knuth vol. 2 (2nd ed), section 4.3.3, p. 278. 82 * 83 * Since C does not give us a `long * long = quad' operator, we split 84 * our input quads into two longs, then split the two longs into two 85 * shorts. We can then calculate `short * short = long' in native 86 * arithmetic. 87 * 88 * Our product should, strictly speaking, be a `long quad', with 128 89 * bits, but we are going to discard the upper 64. In other words, 90 * we are not interested in uv, but rather in (uv mod 2^2n). This 91 * makes some of the terms above vanish, and we get: 92 * 93 * (2^n)(high) + (2^n)(mid) + (2^n + 1)(low) 94 * 95 * or 96 * 97 * (2^n)(high + mid + low) + low 98 * 99 * Furthermore, `high' and `mid' can be computed mod 2^n, as any factor 100 * of 2^n in either one will also vanish. Only `low' need be computed 101 * mod 2^2n, and only because of the final term above. 102 */ 103 static quad_t __lmulq(u_long, u_long); 104 105 quad_t 106 __muldi3(a, b) 107 quad_t a, b; 108 { 109 union uu u, v, low, prod; 110 u_long high, mid, udiff, vdiff; 111 int negall, negmid; 112 #define u1 u.ul[H] 113 #define u0 u.ul[L] 114 #define v1 v.ul[H] 115 #define v0 v.ul[L] 116 117 /* 118 * Get u and v such that u, v >= 0. When this is finished, 119 * u1, u0, v1, and v0 will be directly accessible through the 120 * longword fields. 121 */ 122 if (a >= 0) 123 u.q = a, negall = 0; 124 else 125 u.q = -a, negall = 1; 126 if (b >= 0) 127 v.q = b; 128 else 129 v.q = -b, negall ^= 1; 130 131 if (u1 == 0 && v1 == 0) { 132 /* 133 * An (I hope) important optimization occurs when u1 and v1 134 * are both 0. This should be common since most numbers 135 * are small. Here the product is just u0*v0. 136 */ 137 prod.q = __lmulq(u0, v0); 138 } else { 139 /* 140 * Compute the three intermediate products, remembering 141 * whether the middle term is negative. We can discard 142 * any upper bits in high and mid, so we can use native 143 * u_long * u_long => u_long arithmetic. 144 */ 145 low.q = __lmulq(u0, v0); 146 147 if (u1 >= u0) 148 negmid = 0, udiff = u1 - u0; 149 else 150 negmid = 1, udiff = u0 - u1; 151 if (v0 >= v1) 152 vdiff = v0 - v1; 153 else 154 vdiff = v1 - v0, negmid ^= 1; 155 mid = udiff * vdiff; 156 157 high = u1 * v1; 158 159 /* 160 * Assemble the final product. 161 */ 162 prod.ul[H] = high + (negmid ? -mid : mid) + low.ul[L] + 163 low.ul[H]; 164 prod.ul[L] = low.ul[L]; 165 } 166 return (negall ? -prod.q : prod.q); 167 #undef u1 168 #undef u0 169 #undef v1 170 #undef v0 171 } 172 173 /* 174 * Multiply two 2N-bit longs to produce a 4N-bit quad, where N is half 175 * the number of bits in a long (whatever that is---the code below 176 * does not care as long as quad.h does its part of the bargain---but 177 * typically N==16). 178 * 179 * We use the same algorithm from Knuth, but this time the modulo refinement 180 * does not apply. On the other hand, since N is half the size of a long, 181 * we can get away with native multiplication---none of our input terms 182 * exceeds (ULONG_MAX >> 1). 183 * 184 * Note that, for u_long l, the quad-precision result 185 * 186 * l << N 187 * 188 * splits into high and low longs as HHALF(l) and LHUP(l) respectively. 189 */ 190 static quad_t 191 __lmulq(u_long u, u_long v) 192 { 193 u_long u1, u0, v1, v0, udiff, vdiff, high, mid, low; 194 u_long prodh, prodl, was; 195 union uu prod; 196 int neg; 197 198 u1 = HHALF(u); 199 u0 = LHALF(u); 200 v1 = HHALF(v); 201 v0 = LHALF(v); 202 203 low = u0 * v0; 204 205 /* This is the same small-number optimization as before. */ 206 if (u1 == 0 && v1 == 0) 207 return (low); 208 209 if (u1 >= u0) 210 udiff = u1 - u0, neg = 0; 211 else 212 udiff = u0 - u1, neg = 1; 213 if (v0 >= v1) 214 vdiff = v0 - v1; 215 else 216 vdiff = v1 - v0, neg ^= 1; 217 mid = udiff * vdiff; 218 219 high = u1 * v1; 220 221 /* prod = (high << 2N) + (high << N); */ 222 prodh = high + HHALF(high); 223 prodl = LHUP(high); 224 225 /* if (neg) prod -= mid << N; else prod += mid << N; */ 226 if (neg) { 227 was = prodl; 228 prodl -= LHUP(mid); 229 prodh -= HHALF(mid) + (prodl > was); 230 } else { 231 was = prodl; 232 prodl += LHUP(mid); 233 prodh += HHALF(mid) + (prodl < was); 234 } 235 236 /* prod += low << N */ 237 was = prodl; 238 prodl += LHUP(low); 239 prodh += HHALF(low) + (prodl < was); 240 /* ... + low; */ 241 if ((prodl += low) < low) 242 prodh++; 243 244 /* return 4N-bit product */ 245 prod.ul[H] = prodh; 246 prod.ul[L] = prodl; 247 return (prod.q); 248 } 249