xref: /freebsd/sys/powerpc/fpu/fpu_div.c (revision 2b8331622f0b212cf3bb4fc4914a501e5321d506)
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  *	@(#)fpu_div.c	8.1 (Berkeley) 6/11/93
43  */
44 
45 /*
46  * Perform an FPU divide (return x / y).
47  */
48 
49 #include <sys/cdefs.h>
50 __FBSDID("$FreeBSD$");
51 
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 
55 #include <machine/fpu.h>
56 
57 #include <powerpc/fpu/fpu_arith.h>
58 #include <powerpc/fpu/fpu_emu.h>
59 
60 /*
61  * Division of normal numbers is done as follows:
62  *
63  * x and y are floating point numbers, i.e., in the form 1.bbbb * 2^e.
64  * If X and Y are the mantissas (1.bbbb's), the quotient is then:
65  *
66  *	q = (X / Y) * 2^((x exponent) - (y exponent))
67  *
68  * Since X and Y are both in [1.0,2.0), the quotient's mantissa (X / Y)
69  * will be in [0.5,2.0).  Moreover, it will be less than 1.0 if and only
70  * if X < Y.  In that case, it will have to be shifted left one bit to
71  * become a normal number, and the exponent decremented.  Thus, the
72  * desired exponent is:
73  *
74  *	left_shift = x->fp_mant < y->fp_mant;
75  *	result_exp = x->fp_exp - y->fp_exp - left_shift;
76  *
77  * The quotient mantissa X/Y can then be computed one bit at a time
78  * using the following algorithm:
79  *
80  *	Q = 0;			-- Initial quotient.
81  *	R = X;			-- Initial remainder,
82  *	if (left_shift)		--   but fixed up in advance.
83  *		R *= 2;
84  *	for (bit = FP_NMANT; --bit >= 0; R *= 2) {
85  *		if (R >= Y) {
86  *			Q |= 1 << bit;
87  *			R -= Y;
88  *		}
89  *	}
90  *
91  * The subtraction R -= Y always removes the uppermost bit from R (and
92  * can sometimes remove additional lower-order 1 bits); this proof is
93  * left to the reader.
94  *
95  * This loop correctly calculates the guard and round bits since they are
96  * included in the expanded internal representation.  The sticky bit
97  * is to be set if and only if any other bits beyond guard and round
98  * would be set.  From the above it is obvious that this is true if and
99  * only if the remainder R is nonzero when the loop terminates.
100  *
101  * Examining the loop above, we can see that the quotient Q is built
102  * one bit at a time ``from the top down''.  This means that we can
103  * dispense with the multi-word arithmetic and just build it one word
104  * at a time, writing each result word when it is done.
105  *
106  * Furthermore, since X and Y are both in [1.0,2.0), we know that,
107  * initially, R >= Y.  (Recall that, if X < Y, R is set to X * 2 and
108  * is therefore at in [2.0,4.0).)  Thus Q is sure to have bit FP_NMANT-1
109  * set, and R can be set initially to either X - Y (when X >= Y) or
110  * 2X - Y (when X < Y).  In addition, comparing R and Y is difficult,
111  * so we will simply calculate R - Y and see if that underflows.
112  * This leads to the following revised version of the algorithm:
113  *
114  *	R = X;
115  *	bit = FP_1;
116  *	D = R - Y;
117  *	if (D >= 0) {
118  *		result_exp = x->fp_exp - y->fp_exp;
119  *		R = D;
120  *		q = bit;
121  *		bit >>= 1;
122  *	} else {
123  *		result_exp = x->fp_exp - y->fp_exp - 1;
124  *		q = 0;
125  *	}
126  *	R <<= 1;
127  *	do  {
128  *		D = R - Y;
129  *		if (D >= 0) {
130  *			q |= bit;
131  *			R = D;
132  *		}
133  *		R <<= 1;
134  *	} while ((bit >>= 1) != 0);
135  *	Q[0] = q;
136  *	for (i = 1; i < 4; i++) {
137  *		q = 0, bit = 1 << 31;
138  *		do {
139  *			D = R - Y;
140  *			if (D >= 0) {
141  *				q |= bit;
142  *				R = D;
143  *			}
144  *			R <<= 1;
145  *		} while ((bit >>= 1) != 0);
146  *		Q[i] = q;
147  *	}
148  *
149  * This can be refined just a bit further by moving the `R <<= 1'
150  * calculations to the front of the do-loops and eliding the first one.
151  * The process can be terminated immediately whenever R becomes 0, but
152  * this is relatively rare, and we do not bother.
153  */
154 
155 struct fpn *
156 fpu_div(struct fpemu *fe)
157 {
158 	struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
159 	u_int q, bit;
160 	u_int r0, r1, r2, r3, d0, d1, d2, d3, y0, y1, y2, y3;
161 	FPU_DECL_CARRY
162 
163 	/*
164 	 * Since divide is not commutative, we cannot just use ORDER.
165 	 * Check either operand for NaN first; if there is at least one,
166 	 * order the signalling one (if only one) onto the right, then
167 	 * return it.  Otherwise we have the following cases:
168 	 *
169 	 *	Inf / Inf = NaN, plus NV exception
170 	 *	Inf / num = Inf [i.e., return x]
171 	 *	Inf / 0   = Inf [i.e., return x]
172 	 *	0 / Inf = 0 [i.e., return x]
173 	 *	0 / num = 0 [i.e., return x]
174 	 *	0 / 0   = NaN, plus NV exception
175 	 *	num / Inf = 0
176 	 *	num / num = num (do the divide)
177 	 *	num / 0   = Inf, plus DZ exception
178 	 */
179 	DPRINTF(FPE_REG, ("fpu_div:\n"));
180 	DUMPFPN(FPE_REG, x);
181 	DUMPFPN(FPE_REG, y);
182 	DPRINTF(FPE_REG, ("=>\n"));
183 	if (ISNAN(x) || ISNAN(y)) {
184 		ORDER(x, y);
185 		fe->fe_cx |= FPSCR_VXSNAN;
186 		DUMPFPN(FPE_REG, y);
187 		return (y);
188 	}
189 	/*
190 	 * Need to split the following out cause they generate different
191 	 * exceptions.
192 	 */
193 	if (ISINF(x)) {
194 		if (x->fp_class == y->fp_class) {
195 			fe->fe_cx |= FPSCR_VXIDI;
196 			return (fpu_newnan(fe));
197 		}
198 		DUMPFPN(FPE_REG, x);
199 		return (x);
200 	}
201 	if (ISZERO(x)) {
202 		fe->fe_cx |= FPSCR_ZX;
203 		if (x->fp_class == y->fp_class) {
204 			fe->fe_cx |= FPSCR_VXZDZ;
205 			return (fpu_newnan(fe));
206 		}
207 		DUMPFPN(FPE_REG, x);
208 		return (x);
209 	}
210 
211 	/* all results at this point use XOR of operand signs */
212 	x->fp_sign ^= y->fp_sign;
213 	if (ISINF(y)) {
214 		x->fp_class = FPC_ZERO;
215 		DUMPFPN(FPE_REG, x);
216 		return (x);
217 	}
218 	if (ISZERO(y)) {
219 		fe->fe_cx = FPSCR_ZX;
220 		x->fp_class = FPC_INF;
221 		DUMPFPN(FPE_REG, x);
222 		return (x);
223 	}
224 
225 	/*
226 	 * Macros for the divide.  See comments at top for algorithm.
227 	 * Note that we expand R, D, and Y here.
228 	 */
229 
230 #define	SUBTRACT		/* D = R - Y */ \
231 	FPU_SUBS(d3, r3, y3); FPU_SUBCS(d2, r2, y2); \
232 	FPU_SUBCS(d1, r1, y1); FPU_SUBC(d0, r0, y0)
233 
234 #define	NONNEGATIVE		/* D >= 0 */ \
235 	((int)d0 >= 0)
236 
237 #ifdef FPU_SHL1_BY_ADD
238 #define	SHL1			/* R <<= 1 */ \
239 	FPU_ADDS(r3, r3, r3); FPU_ADDCS(r2, r2, r2); \
240 	FPU_ADDCS(r1, r1, r1); FPU_ADDC(r0, r0, r0)
241 #else
242 #define	SHL1 \
243 	r0 = (r0 << 1) | (r1 >> 31), r1 = (r1 << 1) | (r2 >> 31), \
244 	r2 = (r2 << 1) | (r3 >> 31), r3 <<= 1
245 #endif
246 
247 #define	LOOP			/* do ... while (bit >>= 1) */ \
248 	do { \
249 		SHL1; \
250 		SUBTRACT; \
251 		if (NONNEGATIVE) { \
252 			q |= bit; \
253 			r0 = d0, r1 = d1, r2 = d2, r3 = d3; \
254 		} \
255 	} while ((bit >>= 1) != 0)
256 
257 #define	WORD(r, i)			/* calculate r->fp_mant[i] */ \
258 	q = 0; \
259 	bit = 1 << 31; \
260 	LOOP; \
261 	(x)->fp_mant[i] = q
262 
263 	/* Setup.  Note that we put our result in x. */
264 	r0 = x->fp_mant[0];
265 	r1 = x->fp_mant[1];
266 	r2 = x->fp_mant[2];
267 	r3 = x->fp_mant[3];
268 	y0 = y->fp_mant[0];
269 	y1 = y->fp_mant[1];
270 	y2 = y->fp_mant[2];
271 	y3 = y->fp_mant[3];
272 
273 	bit = FP_1;
274 	SUBTRACT;
275 	if (NONNEGATIVE) {
276 		x->fp_exp -= y->fp_exp;
277 		r0 = d0, r1 = d1, r2 = d2, r3 = d3;
278 		q = bit;
279 		bit >>= 1;
280 	} else {
281 		x->fp_exp -= y->fp_exp + 1;
282 		q = 0;
283 	}
284 	LOOP;
285 	x->fp_mant[0] = q;
286 	WORD(x, 1);
287 	WORD(x, 2);
288 	WORD(x, 3);
289 	x->fp_sticky = r0 | r1 | r2 | r3;
290 
291 	DUMPFPN(FPE_REG, x);
292 	return (x);
293 }
294