xref: /openbsd/lib/libc/arch/sparc64/fpu/fpu_div.c (revision 4a39ccd0)
1 /*	$OpenBSD: fpu_div.c,v 1.2 2012/12/05 23:19:59 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  *	$NetBSD: fpu_div.c,v 1.2 1994/11/20 20:52:38 deraadt Exp $
46  */
47 
48 #if 0
49 __FBSDID("$FreeBSD: src/lib/libc/sparc64/fpu/fpu_div.c,v 1.3 2002/03/22 21:52:58 obrien Exp $");
50 #endif
51 
52 /*
53  * Perform an FPU divide (return x / y).
54  */
55 
56 #include <sys/types.h>
57 
58 #include <machine/frame.h>
59 #include <machine/fsr.h>
60 
61 #include "fpu_arith.h"
62 #include "fpu_emu.h"
63 #include "fpu_extern.h"
64 
65 /*
66  * Division of normal numbers is done as follows:
67  *
68  * x and y are floating point numbers, i.e., in the form 1.bbbb * 2^e.
69  * If X and Y are the mantissas (1.bbbb's), the quotient is then:
70  *
71  *	q = (X / Y) * 2^((x exponent) - (y exponent))
72  *
73  * Since X and Y are both in [1.0,2.0), the quotient's mantissa (X / Y)
74  * will be in [0.5,2.0).  Moreover, it will be less than 1.0 if and only
75  * if X < Y.  In that case, it will have to be shifted left one bit to
76  * become a normal number, and the exponent decremented.  Thus, the
77  * desired exponent is:
78  *
79  *	left_shift = x->fp_mant < y->fp_mant;
80  *	result_exp = x->fp_exp - y->fp_exp - left_shift;
81  *
82  * The quotient mantissa X/Y can then be computed one bit at a time
83  * using the following algorithm:
84  *
85  *	Q = 0;			-- Initial quotient.
86  *	R = X;			-- Initial remainder,
87  *	if (left_shift)		--   but fixed up in advance.
88  *		R *= 2;
89  *	for (bit = FP_NMANT; --bit >= 0; R *= 2) {
90  *		if (R >= Y) {
91  *			Q |= 1 << bit;
92  *			R -= Y;
93  *		}
94  *	}
95  *
96  * The subtraction R -= Y always removes the uppermost bit from R (and
97  * can sometimes remove additional lower-order 1 bits); this proof is
98  * left to the reader.
99  *
100  * This loop correctly calculates the guard and round bits since they are
101  * included in the expanded internal representation.  The sticky bit
102  * is to be set if and only if any other bits beyond guard and round
103  * would be set.  From the above it is obvious that this is true if and
104  * only if the remainder R is nonzero when the loop terminates.
105  *
106  * Examining the loop above, we can see that the quotient Q is built
107  * one bit at a time ``from the top down''.  This means that we can
108  * dispense with the multi-word arithmetic and just build it one word
109  * at a time, writing each result word when it is done.
110  *
111  * Furthermore, since X and Y are both in [1.0,2.0), we know that,
112  * initially, R >= Y.  (Recall that, if X < Y, R is set to X * 2 and
113  * is therefore at in [2.0,4.0).)  Thus Q is sure to have bit FP_NMANT-1
114  * set, and R can be set initially to either X - Y (when X >= Y) or
115  * 2X - Y (when X < Y).  In addition, comparing R and Y is difficult,
116  * so we will simply calculate R - Y and see if that underflows.
117  * This leads to the following revised version of the algorithm:
118  *
119  *	R = X;
120  *	bit = FP_1;
121  *	D = R - Y;
122  *	if (D >= 0) {
123  *		result_exp = x->fp_exp - y->fp_exp;
124  *		R = D;
125  *		q = bit;
126  *		bit >>= 1;
127  *	} else {
128  *		result_exp = x->fp_exp - y->fp_exp - 1;
129  *		q = 0;
130  *	}
131  *	R <<= 1;
132  *	do  {
133  *		D = R - Y;
134  *		if (D >= 0) {
135  *			q |= bit;
136  *			R = D;
137  *		}
138  *		R <<= 1;
139  *	} while ((bit >>= 1) != 0);
140  *	Q[0] = q;
141  *	for (i = 1; i < 4; i++) {
142  *		q = 0, bit = 1 << 31;
143  *		do {
144  *			D = R - Y;
145  *			if (D >= 0) {
146  *				q |= bit;
147  *				R = D;
148  *			}
149  *			R <<= 1;
150  *		} while ((bit >>= 1) != 0);
151  *		Q[i] = q;
152  *	}
153  *
154  * This can be refined just a bit further by moving the `R <<= 1'
155  * calculations to the front of the do-loops and eliding the first one.
156  * The process can be terminated immediately whenever R becomes 0, but
157  * this is relatively rare, and we do not bother.
158  */
159 
160 struct fpn *
161 __fpu_div(fe)
162 	struct fpemu *fe;
163 {
164 	struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
165 	u_int q, bit;
166 	u_int r0, r1, r2, r3, d0, d1, d2, d3, y0, y1, y2, y3;
167 	FPU_DECL_CARRY
168 
169 	/*
170 	 * Since divide is not commutative, we cannot just use ORDER.
171 	 * Check either operand for NaN first; if there is at least one,
172 	 * order the signalling one (if only one) onto the right, then
173 	 * return it.  Otherwise we have the following cases:
174 	 *
175 	 *	Inf / Inf = NaN, plus NV exception
176 	 *	Inf / num = Inf [i.e., return x]
177 	 *	Inf / 0   = Inf [i.e., return x]
178 	 *	0 / Inf = 0 [i.e., return x]
179 	 *	0 / num = 0 [i.e., return x]
180 	 *	0 / 0   = NaN, plus NV exception
181 	 *	num / Inf = 0
182 	 *	num / num = num (do the divide)
183 	 *	num / 0   = Inf, plus DZ exception
184 	 */
185 	if (ISNAN(x) || ISNAN(y)) {
186 		ORDER(x, y);
187 		return (y);
188 	}
189 	if (ISINF(x) || ISZERO(x)) {
190 		if (x->fp_class == y->fp_class)
191 			return (__fpu_newnan(fe));
192 		return (x);
193 	}
194 
195 	/* all results at this point use XOR of operand signs */
196 	x->fp_sign ^= y->fp_sign;
197 	if (ISINF(y)) {
198 		x->fp_class = FPC_ZERO;
199 		return (x);
200 	}
201 	if (ISZERO(y)) {
202 		fe->fe_cx = FSR_DZ;
203 		x->fp_class = FPC_INF;
204 		return (x);
205 	}
206 
207 	/*
208 	 * Macros for the divide.  See comments at top for algorithm.
209 	 * Note that we expand R, D, and Y here.
210 	 */
211 
212 #define	SUBTRACT		/* D = R - Y */ \
213 	FPU_SUBS(d3, r3, y3); FPU_SUBCS(d2, r2, y2); \
214 	FPU_SUBCS(d1, r1, y1); FPU_SUBC(d0, r0, y0)
215 
216 #define	NONNEGATIVE		/* D >= 0 */ \
217 	((int)d0 >= 0)
218 
219 #ifdef FPU_SHL1_BY_ADD
220 #define	SHL1			/* R <<= 1 */ \
221 	FPU_ADDS(r3, r3, r3); FPU_ADDCS(r2, r2, r2); \
222 	FPU_ADDCS(r1, r1, r1); FPU_ADDC(r0, r0, r0)
223 #else
224 #define	SHL1 \
225 	r0 = (r0 << 1) | (r1 >> 31), r1 = (r1 << 1) | (r2 >> 31), \
226 	r2 = (r2 << 1) | (r3 >> 31), r3 <<= 1
227 #endif
228 
229 #define	LOOP			/* do ... while (bit >>= 1) */ \
230 	do { \
231 		SHL1; \
232 		SUBTRACT; \
233 		if (NONNEGATIVE) { \
234 			q |= bit; \
235 			r0 = d0, r1 = d1, r2 = d2, r3 = d3; \
236 		} \
237 	} while ((bit >>= 1) != 0)
238 
239 #define	WORD(r, i)			/* calculate r->fp_mant[i] */ \
240 	q = 0; \
241 	bit = 1 << 31; \
242 	LOOP; \
243 	(x)->fp_mant[i] = q
244 
245 	/* Setup.  Note that we put our result in x. */
246 	r0 = x->fp_mant[0];
247 	r1 = x->fp_mant[1];
248 	r2 = x->fp_mant[2];
249 	r3 = x->fp_mant[3];
250 	y0 = y->fp_mant[0];
251 	y1 = y->fp_mant[1];
252 	y2 = y->fp_mant[2];
253 	y3 = y->fp_mant[3];
254 
255 	bit = FP_1;
256 	SUBTRACT;
257 	if (NONNEGATIVE) {
258 		x->fp_exp -= y->fp_exp;
259 		r0 = d0, r1 = d1, r2 = d2, r3 = d3;
260 		q = bit;
261 		bit >>= 1;
262 	} else {
263 		x->fp_exp -= y->fp_exp + 1;
264 		q = 0;
265 	}
266 	LOOP;
267 	x->fp_mant[0] = q;
268 	WORD(x, 1);
269 	WORD(x, 2);
270 	WORD(x, 3);
271 	x->fp_sticky = r0 | r1 | r2 | r3;
272 
273 	return (x);
274 }
275