xref: /netbsd/sys/arch/sparc/fpu/fpu_div.c (revision bf9ec67e)
1 /*	$NetBSD: fpu_div.c,v 1.2 1994/11/20 20:52:38 deraadt 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  * All advertising materials mentioning features or use of this software
12  * must display the following acknowledgement:
13  *	This product includes software developed by the University of
14  *	California, Lawrence Berkeley Laboratory.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *	This product includes software developed by the University of
27  *	California, Berkeley and its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *	@(#)fpu_div.c	8.1 (Berkeley) 6/11/93
45  */
46 
47 /*
48  * Perform an FPU divide (return x / y).
49  */
50 
51 #include <sys/types.h>
52 
53 #include <machine/reg.h>
54 
55 #include <sparc/fpu/fpu_arith.h>
56 #include <sparc/fpu/fpu_emu.h>
57 
58 /*
59  * Division of normal numbers is done as follows:
60  *
61  * x and y are floating point numbers, i.e., in the form 1.bbbb * 2^e.
62  * If X and Y are the mantissas (1.bbbb's), the quotient is then:
63  *
64  *	q = (X / Y) * 2^((x exponent) - (y exponent))
65  *
66  * Since X and Y are both in [1.0,2.0), the quotient's mantissa (X / Y)
67  * will be in [0.5,2.0).  Moreover, it will be less than 1.0 if and only
68  * if X < Y.  In that case, it will have to be shifted left one bit to
69  * become a normal number, and the exponent decremented.  Thus, the
70  * desired exponent is:
71  *
72  *	left_shift = x->fp_mant < y->fp_mant;
73  *	result_exp = x->fp_exp - y->fp_exp - left_shift;
74  *
75  * The quotient mantissa X/Y can then be computed one bit at a time
76  * using the following algorithm:
77  *
78  *	Q = 0;			-- Initial quotient.
79  *	R = X;			-- Initial remainder,
80  *	if (left_shift)		--   but fixed up in advance.
81  *		R *= 2;
82  *	for (bit = FP_NMANT; --bit >= 0; R *= 2) {
83  *		if (R >= Y) {
84  *			Q |= 1 << bit;
85  *			R -= Y;
86  *		}
87  *	}
88  *
89  * The subtraction R -= Y always removes the uppermost bit from R (and
90  * can sometimes remove additional lower-order 1 bits); this proof is
91  * left to the reader.
92  *
93  * This loop correctly calculates the guard and round bits since they are
94  * included in the expanded internal representation.  The sticky bit
95  * is to be set if and only if any other bits beyond guard and round
96  * would be set.  From the above it is obvious that this is true if and
97  * only if the remainder R is nonzero when the loop terminates.
98  *
99  * Examining the loop above, we can see that the quotient Q is built
100  * one bit at a time ``from the top down''.  This means that we can
101  * dispense with the multi-word arithmetic and just build it one word
102  * at a time, writing each result word when it is done.
103  *
104  * Furthermore, since X and Y are both in [1.0,2.0), we know that,
105  * initially, R >= Y.  (Recall that, if X < Y, R is set to X * 2 and
106  * is therefore at in [2.0,4.0).)  Thus Q is sure to have bit FP_NMANT-1
107  * set, and R can be set initially to either X - Y (when X >= Y) or
108  * 2X - Y (when X < Y).  In addition, comparing R and Y is difficult,
109  * so we will simply calculate R - Y and see if that underflows.
110  * This leads to the following revised version of the algorithm:
111  *
112  *	R = X;
113  *	bit = FP_1;
114  *	D = R - Y;
115  *	if (D >= 0) {
116  *		result_exp = x->fp_exp - y->fp_exp;
117  *		R = D;
118  *		q = bit;
119  *		bit >>= 1;
120  *	} else {
121  *		result_exp = x->fp_exp - y->fp_exp - 1;
122  *		q = 0;
123  *	}
124  *	R <<= 1;
125  *	do  {
126  *		D = R - Y;
127  *		if (D >= 0) {
128  *			q |= bit;
129  *			R = D;
130  *		}
131  *		R <<= 1;
132  *	} while ((bit >>= 1) != 0);
133  *	Q[0] = q;
134  *	for (i = 1; i < 4; i++) {
135  *		q = 0, bit = 1 << 31;
136  *		do {
137  *			D = R - Y;
138  *			if (D >= 0) {
139  *				q |= bit;
140  *				R = D;
141  *			}
142  *			R <<= 1;
143  *		} while ((bit >>= 1) != 0);
144  *		Q[i] = q;
145  *	}
146  *
147  * This can be refined just a bit further by moving the `R <<= 1'
148  * calculations to the front of the do-loops and eliding the first one.
149  * The process can be terminated immediately whenever R becomes 0, but
150  * this is relatively rare, and we do not bother.
151  */
152 
153 struct fpn *
154 fpu_div(fe)
155 	register struct fpemu *fe;
156 {
157 	register struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
158 	register u_int q, bit;
159 	register u_int r0, r1, r2, r3, d0, d1, d2, d3, y0, y1, y2, y3;
160 	FPU_DECL_CARRY
161 
162 	/*
163 	 * Since divide is not commutative, we cannot just use ORDER.
164 	 * Check either operand for NaN first; if there is at least one,
165 	 * order the signalling one (if only one) onto the right, then
166 	 * return it.  Otherwise we have the following cases:
167 	 *
168 	 *	Inf / Inf = NaN, plus NV exception
169 	 *	Inf / num = Inf [i.e., return x]
170 	 *	Inf / 0   = Inf [i.e., return x]
171 	 *	0 / Inf = 0 [i.e., return x]
172 	 *	0 / num = 0 [i.e., return x]
173 	 *	0 / 0   = NaN, plus NV exception
174 	 *	num / Inf = 0
175 	 *	num / num = num (do the divide)
176 	 *	num / 0   = Inf, plus DZ exception
177 	 */
178 	if (ISNAN(x) || ISNAN(y)) {
179 		ORDER(x, y);
180 		return (y);
181 	}
182 	if (ISINF(x) || ISZERO(x)) {
183 		if (x->fp_class == y->fp_class)
184 			return (fpu_newnan(fe));
185 		return (x);
186 	}
187 
188 	/* all results at this point use XOR of operand signs */
189 	x->fp_sign ^= y->fp_sign;
190 	if (ISINF(y)) {
191 		x->fp_class = FPC_ZERO;
192 		return (x);
193 	}
194 	if (ISZERO(y)) {
195 		fe->fe_cx = FSR_DZ;
196 		x->fp_class = FPC_INF;
197 		return (x);
198 	}
199 
200 	/*
201 	 * Macros for the divide.  See comments at top for algorithm.
202 	 * Note that we expand R, D, and Y here.
203 	 */
204 
205 #define	SUBTRACT		/* D = R - Y */ \
206 	FPU_SUBS(d3, r3, y3); FPU_SUBCS(d2, r2, y2); \
207 	FPU_SUBCS(d1, r1, y1); FPU_SUBC(d0, r0, y0)
208 
209 #define	NONNEGATIVE		/* D >= 0 */ \
210 	((int)d0 >= 0)
211 
212 #ifdef FPU_SHL1_BY_ADD
213 #define	SHL1			/* R <<= 1 */ \
214 	FPU_ADDS(r3, r3, r3); FPU_ADDCS(r2, r2, r2); \
215 	FPU_ADDCS(r1, r1, r1); FPU_ADDC(r0, r0, r0)
216 #else
217 #define	SHL1 \
218 	r0 = (r0 << 1) | (r1 >> 31), r1 = (r1 << 1) | (r2 >> 31), \
219 	r2 = (r2 << 1) | (r3 >> 31), r3 <<= 1
220 #endif
221 
222 #define	LOOP			/* do ... while (bit >>= 1) */ \
223 	do { \
224 		SHL1; \
225 		SUBTRACT; \
226 		if (NONNEGATIVE) { \
227 			q |= bit; \
228 			r0 = d0, r1 = d1, r2 = d2, r3 = d3; \
229 		} \
230 	} while ((bit >>= 1) != 0)
231 
232 #define	WORD(r, i)			/* calculate r->fp_mant[i] */ \
233 	q = 0; \
234 	bit = 1 << 31; \
235 	LOOP; \
236 	(x)->fp_mant[i] = q
237 
238 	/* Setup.  Note that we put our result in x. */
239 	r0 = x->fp_mant[0];
240 	r1 = x->fp_mant[1];
241 	r2 = x->fp_mant[2];
242 	r3 = x->fp_mant[3];
243 	y0 = y->fp_mant[0];
244 	y1 = y->fp_mant[1];
245 	y2 = y->fp_mant[2];
246 	y3 = y->fp_mant[3];
247 
248 	bit = FP_1;
249 	SUBTRACT;
250 	if (NONNEGATIVE) {
251 		x->fp_exp -= y->fp_exp;
252 		r0 = d0, r1 = d1, r2 = d2, r3 = d3;
253 		q = bit;
254 		bit >>= 1;
255 	} else {
256 		x->fp_exp -= y->fp_exp + 1;
257 		q = 0;
258 	}
259 	LOOP;
260 	x->fp_mant[0] = q;
261 	WORD(x, 1);
262 	WORD(x, 2);
263 	WORD(x, 3);
264 	x->fp_sticky = r0 | r1 | r2 | r3;
265 
266 	return (x);
267 }
268