usr/src/lib/libc/port/fp/muldi3.c

   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 #pragma ident   "%Z%%M% %I%     %E% SMI"
  39
  40 #include "quadint.h"
  41
  42 #pragma weak __muldi3 = ___muldi3
  43
  44 /*
  45  * Multiply two quads.
  46  *
  47  * Our algorithm is based on the following.  Split incoming quad values
  48  * u and v (where u,v >= 0) into
  49  *
  50  *      u = 2^n u1  *  u0       (n = number of bits in `u_long', usu. 32)
  51  *
  52  * and
  53  *
  54  *      v = 2^n v1  *  v0
  55  *
  56  * Then
  57  *
  58  *      uv = 2^2n u1 v1  +  2^n u1 v0  +  2^n v1 u0  +  u0 v0
  59  *         = 2^2n u1 v1  +     2^n (u1 v0 + v1 u0)   +  u0 v0
  60  *
  61  * Now add 2^n u1 v1 to the first term and subtract it from the middle,
  62  * and add 2^n u0 v0 to the last term and subtract it from the middle.
  63  * This gives:
  64  *
  65  *      uv = (2^2n + 2^n) (u1 v1)  +
  66  *               (2^n)    (u1 v0 - u1 v1 + u0 v1 - u0 v0)  +
  67  *             (2^n + 1)  (u0 v0)
  68  *
  69  * Factoring the middle a bit gives us:
  70  *
  71  *      uv = (2^2n + 2^n) (u1 v1)  +                    [u1v1 = high]
  72  *               (2^n)    (u1 - u0) (v0 - v1)  +        [(u1-u0)... = mid]
  73  *             (2^n + 1)  (u0 v0)                       [u0v0 = low]
  74  *
  75  * The terms (u1 v1), (u1 - u0) (v0 - v1), and (u0 v0) can all be done
  76  * in just half the precision of the original.  (Note that either or both
  77  * of (u1 - u0) or (v0 - v1) may be negative.)
  78  *
  79  * This algorithm is from Knuth vol. 2 (2nd ed), section 4.3.3, p. 278.
  80  *
  81  * Since C does not give us a `long * long = quad' operator, we split
  82  * our input quads into two longs, then split the two longs into two
  83  * shorts.  We can then calculate `short * short = long' in native
  84  * arithmetic.
  85  *
  86  * Our product should, strictly speaking, be a `long quad', with 128
  87  * bits, but we are going to discard the upper 64.  In other words,
  88  * we are not interested in uv, but rather in (uv mod 2^2n).  This
  89  * makes some of the terms above vanish, and we get:
  90  *
  91  *      (2^n)(high) + (2^n)(mid) + (2^n + 1)(low)
  92  *
  93  * or
  94  *
  95  *      (2^n)(high + mid + low) + low
  96  *
  97  * Furthermore, `high' and `mid' can be computed mod 2^n, as any factor
  98  * of 2^n in either one will also vanish.  Only `low' need be computed
  99  * mod 2^2n, and only because of the final term above.
 100  */
 101 static longlong_t __lmulq(ulong_t, ulong_t);
 102
 103 longlong_t
 104 ___muldi3(longlong_t a, longlong_t b)
 105 {
 106         union uu u, v, low, prod;
 107         ulong_t high, mid, udiff, vdiff;
 108         int negall, negmid;
 109 #define u1      u.ul[H]
 110 #define u0      u.ul[L]
 111 #define v1      v.ul[H]
 112 #define v0      v.ul[L]
 113
 114         /*
 115          * Get u and v such that u, v >= 0.  When this is finished,
 116          * u1, u0, v1, and v0 will be directly accessible through the
 117          * longword fields.
 118          */
 119         if (a >= 0)
 120                 u.q = a, negall = 0;
 121         else
 122                 u.q = -a, negall = 1;
 123         if (b >= 0)
 124                 v.q = b;
 125         else
 126                 v.q = -b, negall ^= 1;
 127
 128         if (u1 == 0 && v1 == 0) {
 129                 /*
 130                  * An (I hope) important optimization occurs when u1 and v1
 131                  * are both 0.  This should be common since most numbers
 132                  * are small.  Here the product is just u0*v0.
 133                  */
 134                 prod.q = __lmulq(u0, v0);
 135         } else {
 136                 /*
 137                  * Compute the three intermediate products, remembering
 138                  * whether the middle term is negative.  We can discard
 139                  * any upper bits in high and mid, so we can use native
 140                  * ulong_t * ulong_t => ulong_t arithmetic.
 141                  */
 142                 low.q = __lmulq(u0, v0);
 143
 144                 if (u1 >= u0)
 145                         negmid = 0, udiff = u1 - u0;
 146                 else
 147                         negmid = 1, udiff = u0 - u1;
 148                 if (v0 >= v1)
 149                         vdiff = v0 - v1;
 150                 else
 151                         vdiff = v1 - v0, negmid ^= 1;
 152                 mid = udiff * vdiff;
 153
 154                 high = u1 * v1;
 155
 156                 /*
 157                  * Assemble the final product.
 158                  */
 159                 prod.ul[H] = high + (negmid ? -mid : mid) + low.ul[L] +
 160                     low.ul[H];
 161                 prod.ul[L] = low.ul[L];
 162         }
 163         return (negall ? -prod.q : prod.q);
 164 #undef u1
 165 #undef u0
 166 #undef v1
 167 #undef v0
 168 }
 169
 170 /*
 171  * Multiply two 2N-bit longs to produce a 4N-bit quad, where N is half
 172  * the number of bits in a long (whatever that is---the code below
 173  * does not care as long as quad.h does its part of the bargain---but
 174  * typically N==16).
 175  *
 176  * We use the same algorithm from Knuth, but this time the modulo refinement
 177  * does not apply.  On the other hand, since N is half the size of a long,
 178  * we can get away with native multiplication---none of our input terms
 179  * exceeds (ULONG_MAX >> 1).
 180  *
 181  * Note that, for ulong_t l, the quad-precision result
 182  *
 183  *      l << N
 184  *
 185  * splits into high and low longs as HHALF(l) and LHUP(l) respectively.
 186  */
 187 static longlong_t
 188 __lmulq(ulong_t u, ulong_t v)
 189 {
 190         ulong_t u1, u0, v1, v0, udiff, vdiff, high, mid, low;
 191         ulong_t prodh, prodl, was;
 192         union uu prod;
 193         int neg;
 194
 195         u1 = HHALF(u);
 196         u0 = LHALF(u);
 197         v1 = HHALF(v);
 198         v0 = LHALF(v);
 199
 200         low = u0 * v0;
 201
 202         /* This is the same small-number optimization as before. */
 203         if (u1 == 0 && v1 == 0)
 204                 return (low);
 205
 206         if (u1 >= u0)
 207                 udiff = u1 - u0, neg = 0;
 208         else
 209                 udiff = u0 - u1, neg = 1;
 210         if (v0 >= v1)
 211                 vdiff = v0 - v1;
 212         else
 213                 vdiff = v1 - v0, neg ^= 1;
 214         mid = udiff * vdiff;
 215
 216         high = u1 * v1;
 217
 218         /* prod = (high << 2N) + (high << N); */
 219         prodh = high + HHALF(high);
 220         prodl = LHUP(high);
 221
 222         /* if (neg) prod -= mid << N; else prod += mid << N; */
 223         if (neg) {
 224                 was = prodl;
 225                 prodl -= LHUP(mid);
 226                 prodh -= HHALF(mid) + (prodl > was);
 227         } else {
 228                 was = prodl;
 229                 prodl += LHUP(mid);
 230                 prodh += HHALF(mid) + (prodl < was);
 231         }
 232
 233         /* prod += low << N */
 234         was = prodl;
 235         prodl += LHUP(low);
 236         prodh += HHALF(low) + (prodl < was);
 237         /* ... + low; */
 238         if ((prodl += low) < low)
 239                 prodh++;
 240
 241         /* return 4N-bit product */
 242         prod.ul[H] = prodh;
 243         prod.ul[L] = prodl;
 244         return (prod.q);
 245 }