xref: /freebsd/sys/powerpc/fpu/fpu_div.c (revision 5b31cc94b10d4bb7109c6b27940a0fc76a44a331)
1 /*	$NetBSD: fpu_div.c,v 1.4 2005/12/11 12:18:42 christos Exp $ */
2 
3 /*-
4  * SPDX-License-Identifier: BSD-3-Clause
5  *
6  * Copyright (c) 1992, 1993
7  *	The Regents of the University of California.  All rights reserved.
8  *
9  * This software was developed by the Computer Systems Engineering group
10  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
11  * contributed to Berkeley.
12  *
13  * All advertising materials mentioning features or use of this software
14  * must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Lawrence Berkeley Laboratory.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. 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 
43 /*
44  * Perform an FPU divide (return x / y).
45  */
46 
47 #include <sys/cdefs.h>
48 #include <sys/types.h>
49 #include <sys/systm.h>
50 
51 #include <machine/fpu.h>
52 
53 #include <powerpc/fpu/fpu_arith.h>
54 #include <powerpc/fpu/fpu_emu.h>
55 
56 /*
57  * Division of normal numbers is done as follows:
58  *
59  * x and y are floating point numbers, i.e., in the form 1.bbbb * 2^e.
60  * If X and Y are the mantissas (1.bbbb's), the quotient is then:
61  *
62  *	q = (X / Y) * 2^((x exponent) - (y exponent))
63  *
64  * Since X and Y are both in [1.0,2.0), the quotient's mantissa (X / Y)
65  * will be in [0.5,2.0).  Moreover, it will be less than 1.0 if and only
66  * if X < Y.  In that case, it will have to be shifted left one bit to
67  * become a normal number, and the exponent decremented.  Thus, the
68  * desired exponent is:
69  *
70  *	left_shift = x->fp_mant < y->fp_mant;
71  *	result_exp = x->fp_exp - y->fp_exp - left_shift;
72  *
73  * The quotient mantissa X/Y can then be computed one bit at a time
74  * using the following algorithm:
75  *
76  *	Q = 0;			-- Initial quotient.
77  *	R = X;			-- Initial remainder,
78  *	if (left_shift)		--   but fixed up in advance.
79  *		R *= 2;
80  *	for (bit = FP_NMANT; --bit >= 0; R *= 2) {
81  *		if (R >= Y) {
82  *			Q |= 1 << bit;
83  *			R -= Y;
84  *		}
85  *	}
86  *
87  * The subtraction R -= Y always removes the uppermost bit from R (and
88  * can sometimes remove additional lower-order 1 bits); this proof is
89  * left to the reader.
90  *
91  * This loop correctly calculates the guard and round bits since they are
92  * included in the expanded internal representation.  The sticky bit
93  * is to be set if and only if any other bits beyond guard and round
94  * would be set.  From the above it is obvious that this is true if and
95  * only if the remainder R is nonzero when the loop terminates.
96  *
97  * Examining the loop above, we can see that the quotient Q is built
98  * one bit at a time ``from the top down''.  This means that we can
99  * dispense with the multi-word arithmetic and just build it one word
100  * at a time, writing each result word when it is done.
101  *
102  * Furthermore, since X and Y are both in [1.0,2.0), we know that,
103  * initially, R >= Y.  (Recall that, if X < Y, R is set to X * 2 and
104  * is therefore at in [2.0,4.0).)  Thus Q is sure to have bit FP_NMANT-1
105  * set, and R can be set initially to either X - Y (when X >= Y) or
106  * 2X - Y (when X < Y).  In addition, comparing R and Y is difficult,
107  * so we will simply calculate R - Y and see if that underflows.
108  * This leads to the following revised version of the algorithm:
109  *
110  *	R = X;
111  *	bit = FP_1;
112  *	D = R - Y;
113  *	if (D >= 0) {
114  *		result_exp = x->fp_exp - y->fp_exp;
115  *		R = D;
116  *		q = bit;
117  *		bit >>= 1;
118  *	} else {
119  *		result_exp = x->fp_exp - y->fp_exp - 1;
120  *		q = 0;
121  *	}
122  *	R <<= 1;
123  *	do  {
124  *		D = R - Y;
125  *		if (D >= 0) {
126  *			q |= bit;
127  *			R = D;
128  *		}
129  *		R <<= 1;
130  *	} while ((bit >>= 1) != 0);
131  *	Q[0] = q;
132  *	for (i = 1; i < 4; i++) {
133  *		q = 0, bit = 1 << 31;
134  *		do {
135  *			D = R - Y;
136  *			if (D >= 0) {
137  *				q |= bit;
138  *				R = D;
139  *			}
140  *			R <<= 1;
141  *		} while ((bit >>= 1) != 0);
142  *		Q[i] = q;
143  *	}
144  *
145  * This can be refined just a bit further by moving the `R <<= 1'
146  * calculations to the front of the do-loops and eliding the first one.
147  * The process can be terminated immediately whenever R becomes 0, but
148  * this is relatively rare, and we do not bother.
149  */
150 
151 struct fpn *
152 fpu_div(struct fpemu *fe)
153 {
154 	struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
155 	u_int q, bit;
156 	u_int r0, r1, r2, r3, d0, d1, d2, d3, y0, y1, y2, y3;
157 	FPU_DECL_CARRY
158 
159 	/*
160 	 * Since divide is not commutative, we cannot just use ORDER.
161 	 * Check either operand for NaN first; if there is at least one,
162 	 * order the signalling one (if only one) onto the right, then
163 	 * return it.  Otherwise we have the following cases:
164 	 *
165 	 *	Inf / Inf = NaN, plus NV exception
166 	 *	Inf / num = Inf [i.e., return x]
167 	 *	Inf / 0   = Inf [i.e., return x]
168 	 *	0 / Inf = 0 [i.e., return x]
169 	 *	0 / num = 0 [i.e., return x]
170 	 *	0 / 0   = NaN, plus NV exception
171 	 *	num / Inf = 0
172 	 *	num / num = num (do the divide)
173 	 *	num / 0   = Inf, plus DZ exception
174 	 */
175 	DPRINTF(FPE_REG, ("fpu_div:\n"));
176 	DUMPFPN(FPE_REG, x);
177 	DUMPFPN(FPE_REG, y);
178 	DPRINTF(FPE_REG, ("=>\n"));
179 	if (ISNAN(x) || ISNAN(y)) {
180 		ORDER(x, y);
181 		fe->fe_cx |= FPSCR_VXSNAN;
182 		DUMPFPN(FPE_REG, y);
183 		return (y);
184 	}
185 	/*
186 	 * Need to split the following out cause they generate different
187 	 * exceptions.
188 	 */
189 	if (ISINF(x)) {
190 		if (x->fp_class == y->fp_class) {
191 			fe->fe_cx |= FPSCR_VXIDI;
192 			return (fpu_newnan(fe));
193 		}
194 		DUMPFPN(FPE_REG, x);
195 		return (x);
196 	}
197 	if (ISZERO(x)) {
198 		fe->fe_cx |= FPSCR_ZX;
199 		if (x->fp_class == y->fp_class) {
200 			fe->fe_cx |= FPSCR_VXZDZ;
201 			return (fpu_newnan(fe));
202 		}
203 		DUMPFPN(FPE_REG, x);
204 		return (x);
205 	}
206 
207 	/* all results at this point use XOR of operand signs */
208 	x->fp_sign ^= y->fp_sign;
209 	if (ISINF(y)) {
210 		x->fp_class = FPC_ZERO;
211 		DUMPFPN(FPE_REG, x);
212 		return (x);
213 	}
214 	if (ISZERO(y)) {
215 		fe->fe_cx = FPSCR_ZX;
216 		x->fp_class = FPC_INF;
217 		DUMPFPN(FPE_REG, x);
218 		return (x);
219 	}
220 
221 	/*
222 	 * Macros for the divide.  See comments at top for algorithm.
223 	 * Note that we expand R, D, and Y here.
224 	 */
225 
226 #define	SUBTRACT		/* D = R - Y */ \
227 	FPU_SUBS(d3, r3, y3); FPU_SUBCS(d2, r2, y2); \
228 	FPU_SUBCS(d1, r1, y1); FPU_SUBC(d0, r0, y0)
229 
230 #define	NONNEGATIVE		/* D >= 0 */ \
231 	((int)d0 >= 0)
232 
233 #ifdef FPU_SHL1_BY_ADD
234 #define	SHL1			/* R <<= 1 */ \
235 	FPU_ADDS(r3, r3, r3); FPU_ADDCS(r2, r2, r2); \
236 	FPU_ADDCS(r1, r1, r1); FPU_ADDC(r0, r0, r0)
237 #else
238 #define	SHL1 \
239 	r0 = (r0 << 1) | (r1 >> 31), r1 = (r1 << 1) | (r2 >> 31), \
240 	r2 = (r2 << 1) | (r3 >> 31), r3 <<= 1
241 #endif
242 
243 #define	LOOP			/* do ... while (bit >>= 1) */ \
244 	do { \
245 		SHL1; \
246 		SUBTRACT; \
247 		if (NONNEGATIVE) { \
248 			q |= bit; \
249 			r0 = d0, r1 = d1, r2 = d2, r3 = d3; \
250 		} \
251 	} while ((bit >>= 1) != 0)
252 
253 #define	WORD(r, i)			/* calculate r->fp_mant[i] */ \
254 	q = 0; \
255 	bit = 1 << 31; \
256 	LOOP; \
257 	(x)->fp_mant[i] = q
258 
259 	/* Setup.  Note that we put our result in x. */
260 	r0 = x->fp_mant[0];
261 	r1 = x->fp_mant[1];
262 	r2 = x->fp_mant[2];
263 	r3 = x->fp_mant[3];
264 	y0 = y->fp_mant[0];
265 	y1 = y->fp_mant[1];
266 	y2 = y->fp_mant[2];
267 	y3 = y->fp_mant[3];
268 
269 	bit = FP_1;
270 	SUBTRACT;
271 	if (NONNEGATIVE) {
272 		x->fp_exp -= y->fp_exp;
273 		r0 = d0, r1 = d1, r2 = d2, r3 = d3;
274 		q = bit;
275 		bit >>= 1;
276 	} else {
277 		x->fp_exp -= y->fp_exp + 1;
278 		q = 0;
279 	}
280 	LOOP;
281 	x->fp_mant[0] = q;
282 	WORD(x, 1);
283 	WORD(x, 2);
284 	WORD(x, 3);
285 	x->fp_sticky = r0 | r1 | r2 | r3;
286 
287 	DUMPFPN(FPE_REG, x);
288 	return (x);
289 }
290