xref: /netbsd-src/sys/arch/sparc/fpu/fpu_mul.c (revision ae1bfcddc410612bc8c58b807e1830becb69a24c)
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  * 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.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
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.
29  *
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
40  * SUCH DAMAGE.
41  *
42  *	@(#)fpu_mul.c	8.1 (Berkeley) 6/11/93
43  *
44  * from: Header: fpu_mul.c,v 1.3 92/11/26 01:39:50 torek Exp
45  * $Id: fpu_mul.c,v 1.1 1993/10/02 10:22:59 deraadt Exp $
46  */
47 
48 /*
49  * Perform an FPU multiply (return x * y).
50  */
51 
52 #include <sys/types.h>
53 
54 #include <machine/reg.h>
55 
56 #include <sparc/fpu/fpu_arith.h>
57 #include <sparc/fpu/fpu_emu.h>
58 
59 /*
60  * The multiplication algorithm for normal numbers is as follows:
61  *
62  * The fraction of the product is built in the usual stepwise fashion.
63  * Each step consists of shifting the accumulator right one bit
64  * (maintaining any guard bits) and, if the next bit in y is set,
65  * adding the multiplicand (x) to the accumulator.  Then, in any case,
66  * we advance one bit leftward in y.  Algorithmically:
67  *
68  *	A = 0;
69  *	for (bit = 0; bit < FP_NMANT; bit++) {
70  *		sticky |= A & 1, A >>= 1;
71  *		if (Y & (1 << bit))
72  *			A += X;
73  *	}
74  *
75  * (X and Y here represent the mantissas of x and y respectively.)
76  * The resultant accumulator (A) is the product's mantissa.  It may
77  * be as large as 11.11111... in binary and hence may need to be
78  * shifted right, but at most one bit.
79  *
80  * Since we do not have efficient multiword arithmetic, we code the
81  * accumulator as four separate words, just like any other mantissa.
82  * We use local `register' variables in the hope that this is faster
83  * than memory.  We keep x->fp_mant in locals for the same reason.
84  *
85  * In the algorithm above, the bits in y are inspected one at a time.
86  * We will pick them up 32 at a time and then deal with those 32, one
87  * at a time.  Note, however, that we know several things about y:
88  *
89  *    - the guard and round bits at the bottom are sure to be zero;
90  *
91  *    - often many low bits are zero (y is often from a single or double
92  *	precision source);
93  *
94  *    - bit FP_NMANT-1 is set, and FP_1*2 fits in a word.
95  *
96  * We can also test for 32-zero-bits swiftly.  In this case, the center
97  * part of the loop---setting sticky, shifting A, and not adding---will
98  * run 32 times without adding X to A.  We can do a 32-bit shift faster
99  * by simply moving words.  Since zeros are common, we optimize this case.
100  * Furthermore, since A is initially zero, we can omit the shift as well
101  * until we reach a nonzero word.
102  */
103 struct fpn *
104 fpu_mul(fe)
105 	register struct fpemu *fe;
106 {
107 	register struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
108 	register u_int a3, a2, a1, a0, x3, x2, x1, x0, bit, m;
109 	register int sticky;
110 	FPU_DECL_CARRY
111 
112 	/*
113 	 * Put the `heavier' operand on the right (see fpu_emu.h).
114 	 * Then we will have one of the following cases, taken in the
115 	 * following order:
116 	 *
117 	 *  - y = NaN.  Implied: if only one is a signalling NaN, y is.
118 	 *	The result is y.
119 	 *  - y = Inf.  Implied: x != NaN (is 0, number, or Inf: the NaN
120 	 *    case was taken care of earlier).
121 	 *	If x = 0, the result is NaN.  Otherwise the result
122 	 *	is y, with its sign reversed if x is negative.
123 	 *  - x = 0.  Implied: y is 0 or number.
124 	 *	The result is 0 (with XORed sign as usual).
125 	 *  - other.  Implied: both x and y are numbers.
126 	 *	The result is x * y (XOR sign, multiply bits, add exponents).
127 	 */
128 	ORDER(x, y);
129 	if (ISNAN(y)) {
130 		y->fp_sign ^= x->fp_sign;
131 		return (y);
132 	}
133 	if (ISINF(y)) {
134 		if (ISZERO(x))
135 			return (fpu_newnan(fe));
136 		y->fp_sign ^= x->fp_sign;
137 		return (y);
138 	}
139 	if (ISZERO(x)) {
140 		x->fp_sign ^= y->fp_sign;
141 		return (x);
142 	}
143 
144 	/*
145 	 * Setup.  In the code below, the mask `m' will hold the current
146 	 * mantissa byte from y.  The variable `bit' denotes the bit
147 	 * within m.  We also define some macros to deal with everything.
148 	 */
149 	x3 = x->fp_mant[3];
150 	x2 = x->fp_mant[2];
151 	x1 = x->fp_mant[1];
152 	x0 = x->fp_mant[0];
153 	sticky = a3 = a2 = a1 = a0 = 0;
154 
155 #define	ADD	/* A += X */ \
156 	FPU_ADDS(a3, a3, x3); \
157 	FPU_ADDCS(a2, a2, x2); \
158 	FPU_ADDCS(a1, a1, x1); \
159 	FPU_ADDC(a0, a0, x0)
160 
161 #define	SHR1	/* A >>= 1, with sticky */ \
162 	sticky |= a3 & 1, a3 = (a3 >> 1) | (a2 << 31), \
163 	a2 = (a2 >> 1) | (a1 << 31), a1 = (a1 >> 1) | (a0 << 31), a0 >>= 1
164 
165 #define	SHR32	/* A >>= 32, with sticky */ \
166 	sticky |= a3, a3 = a2, a2 = a1, a1 = a0, a0 = 0
167 
168 #define	STEP	/* each 1-bit step of the multiplication */ \
169 	SHR1; if (bit & m) { ADD; }; bit <<= 1
170 
171 	/*
172 	 * We are ready to begin.  The multiply loop runs once for each
173 	 * of the four 32-bit words.  Some words, however, are special.
174 	 * As noted above, the low order bits of Y are often zero.  Even
175 	 * if not, the first loop can certainly skip the guard bits.
176 	 * The last word of y has its highest 1-bit in position FP_NMANT-1,
177 	 * so we stop the loop when we move past that bit.
178 	 */
179 	if ((m = y->fp_mant[3]) == 0) {
180 		/* SHR32; */			/* unneeded since A==0 */
181 	} else {
182 		bit = 1 << FP_NG;
183 		do {
184 			STEP;
185 		} while (bit != 0);
186 	}
187 	if ((m = y->fp_mant[2]) == 0) {
188 		SHR32;
189 	} else {
190 		bit = 1;
191 		do {
192 			STEP;
193 		} while (bit != 0);
194 	}
195 	if ((m = y->fp_mant[1]) == 0) {
196 		SHR32;
197 	} else {
198 		bit = 1;
199 		do {
200 			STEP;
201 		} while (bit != 0);
202 	}
203 	m = y->fp_mant[0];		/* definitely != 0 */
204 	bit = 1;
205 	do {
206 		STEP;
207 	} while (bit <= m);
208 
209 	/*
210 	 * Done with mantissa calculation.  Get exponent and handle
211 	 * 11.111...1 case, then put result in place.  We reuse x since
212 	 * it already has the right class (FP_NUM).
213 	 */
214 	m = x->fp_exp + y->fp_exp;
215 	if (a0 >= FP_2) {
216 		SHR1;
217 		m++;
218 	}
219 	x->fp_sign ^= y->fp_sign;
220 	x->fp_exp = m;
221 	x->fp_sticky = sticky;
222 	x->fp_mant[3] = a3;
223 	x->fp_mant[2] = a2;
224 	x->fp_mant[1] = a1;
225 	x->fp_mant[0] = a0;
226 	return (x);
227 }
228