2 * Copyright (c) 1992, 1993
3 * The Regents of the University of California. All rights reserved.
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.
9 * All advertising materials mentioning features or use of this software
10 * must display the following acknowledgement:
11 * This product includes software developed by the University of
12 * California, Lawrence Berkeley Laboratory.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * @(#)fpu_mul.c 8.1 (Berkeley) 6/11/93
43 * $NetBSD: fpu_mul.c,v 1.2 1994/11/20 20:52:44 deraadt Exp $
46 #include <sys/cdefs.h>
47 __FBSDID("$FreeBSD$");
50 * Perform an FPU multiply (return x * y).
53 #include <sys/types.h>
55 #include <machine/frame.h>
56 #include <machine/fp.h>
58 #include "fpu_arith.h"
60 #include "fpu_extern.h"
63 * The multiplication algorithm for normal numbers is as follows:
65 * The fraction of the product is built in the usual stepwise fashion.
66 * Each step consists of shifting the accumulator right one bit
67 * (maintaining any guard bits) and, if the next bit in y is set,
68 * adding the multiplicand (x) to the accumulator. Then, in any case,
69 * we advance one bit leftward in y. Algorithmically:
72 * for (bit = 0; bit < FP_NMANT; bit++) {
73 * sticky |= A & 1, A >>= 1;
78 * (X and Y here represent the mantissas of x and y respectively.)
79 * The resultant accumulator (A) is the product's mantissa. It may
80 * be as large as 11.11111... in binary and hence may need to be
81 * shifted right, but at most one bit.
83 * Since we do not have efficient multiword arithmetic, we code the
84 * accumulator as four separate words, just like any other mantissa.
85 * We use local `register' variables in the hope that this is faster
86 * than memory. We keep x->fp_mant in locals for the same reason.
88 * In the algorithm above, the bits in y are inspected one at a time.
89 * We will pick them up 32 at a time and then deal with those 32, one
90 * at a time. Note, however, that we know several things about y:
92 * - the guard and round bits at the bottom are sure to be zero;
94 * - often many low bits are zero (y is often from a single or double
97 * - bit FP_NMANT-1 is set, and FP_1*2 fits in a word.
99 * We can also test for 32-zero-bits swiftly. In this case, the center
100 * part of the loop---setting sticky, shifting A, and not adding---will
101 * run 32 times without adding X to A. We can do a 32-bit shift faster
102 * by simply moving words. Since zeros are common, we optimize this case.
103 * Furthermore, since A is initially zero, we can omit the shift as well
104 * until we reach a nonzero word.
110 struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
111 u_int a3, a2, a1, a0, x3, x2, x1, x0, bit, m;
116 * Put the `heavier' operand on the right (see fpu_emu.h).
117 * Then we will have one of the following cases, taken in the
120 * - y = NaN. Implied: if only one is a signalling NaN, y is.
122 * - y = Inf. Implied: x != NaN (is 0, number, or Inf: the NaN
123 * case was taken care of earlier).
124 * If x = 0, the result is NaN. Otherwise the result
125 * is y, with its sign reversed if x is negative.
126 * - x = 0. Implied: y is 0 or number.
127 * The result is 0 (with XORed sign as usual).
128 * - other. Implied: both x and y are numbers.
129 * The result is x * y (XOR sign, multiply bits, add exponents).
133 y->fp_sign ^= x->fp_sign;
138 return (__fpu_newnan(fe));
139 y->fp_sign ^= x->fp_sign;
143 x->fp_sign ^= y->fp_sign;
148 * Setup. In the code below, the mask `m' will hold the current
149 * mantissa byte from y. The variable `bit' denotes the bit
150 * within m. We also define some macros to deal with everything.
156 sticky = a3 = a2 = a1 = a0 = 0;
158 #define ADD /* A += X */ \
159 FPU_ADDS(a3, a3, x3); \
160 FPU_ADDCS(a2, a2, x2); \
161 FPU_ADDCS(a1, a1, x1); \
164 #define SHR1 /* A >>= 1, with sticky */ \
165 sticky |= a3 & 1, a3 = (a3 >> 1) | (a2 << 31), \
166 a2 = (a2 >> 1) | (a1 << 31), a1 = (a1 >> 1) | (a0 << 31), a0 >>= 1
168 #define SHR32 /* A >>= 32, with sticky */ \
169 sticky |= a3, a3 = a2, a2 = a1, a1 = a0, a0 = 0
171 #define STEP /* each 1-bit step of the multiplication */ \
172 SHR1; if (bit & m) { ADD; }; bit <<= 1
175 * We are ready to begin. The multiply loop runs once for each
176 * of the four 32-bit words. Some words, however, are special.
177 * As noted above, the low order bits of Y are often zero. Even
178 * if not, the first loop can certainly skip the guard bits.
179 * The last word of y has its highest 1-bit in position FP_NMANT-1,
180 * so we stop the loop when we move past that bit.
182 if ((m = y->fp_mant[3]) == 0) {
183 /* SHR32; */ /* unneeded since A==0 */
190 if ((m = y->fp_mant[2]) == 0) {
198 if ((m = y->fp_mant[1]) == 0) {
206 m = y->fp_mant[0]; /* definitely != 0 */
213 * Done with mantissa calculation. Get exponent and handle
214 * 11.111...1 case, then put result in place. We reuse x since
215 * it already has the right class (FP_NUM).
217 m = x->fp_exp + y->fp_exp;
222 x->fp_sign ^= y->fp_sign;
224 x->fp_sticky = sticky;