xref: /netbsd-src/sys/arch/sparc/fpu/fpu_implode.c (revision 2a399c6883d870daece976daec6ffa7bb7f934ce)
1 /*	$NetBSD: fpu_implode.c,v 1.3 1996/03/14 19:41:59 christos 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_implode.c	8.1 (Berkeley) 6/11/93
45  */
46 
47 /*
48  * FPU subroutines: `implode' internal format numbers into the machine's
49  * `packed binary' format.
50  */
51 
52 #include <sys/types.h>
53 #include <sys/systm.h>
54 
55 #include <machine/ieee.h>
56 #include <machine/instr.h>
57 #include <machine/reg.h>
58 
59 #include <sparc/fpu/fpu_arith.h>
60 #include <sparc/fpu/fpu_emu.h>
61 #include <sparc/fpu/fpu_extern.h>
62 
63 static int round __P((register struct fpemu *, register struct fpn *));
64 static int toinf __P((struct fpemu *, int));
65 
66 /*
67  * Round a number (algorithm from Motorola MC68882 manual, modified for
68  * our internal format).  Set inexact exception if rounding is required.
69  * Return true iff we rounded up.
70  *
71  * After rounding, we discard the guard and round bits by shifting right
72  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
73  * This saves effort later.
74  *
75  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
76  * responsibility to fix this if necessary.
77  */
78 static int
79 round(register struct fpemu *fe, register struct fpn *fp)
80 {
81 	register u_int m0, m1, m2, m3;
82 	register int gr, s;
83 
84 	m0 = fp->fp_mant[0];
85 	m1 = fp->fp_mant[1];
86 	m2 = fp->fp_mant[2];
87 	m3 = fp->fp_mant[3];
88 	gr = m3 & 3;
89 	s = fp->fp_sticky;
90 
91 	/* mant >>= FP_NG */
92 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
93 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
94 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
95 	m0 >>= FP_NG;
96 
97 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
98 		goto rounddown;
99 
100 	fe->fe_cx |= FSR_NX;	/* inexact */
101 
102 	/* Go to rounddown to round down; break to round up. */
103 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
104 
105 	case FSR_RD_RN:
106 	default:
107 		/*
108 		 * Round only if guard is set (gr & 2).  If guard is set,
109 		 * but round & sticky both clear, then we want to round
110 		 * but have a tie, so round to even, i.e., add 1 iff odd.
111 		 */
112 		if ((gr & 2) == 0)
113 			goto rounddown;
114 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
115 			break;
116 		goto rounddown;
117 
118 	case FSR_RD_RZ:
119 		/* Round towards zero, i.e., down. */
120 		goto rounddown;
121 
122 	case FSR_RD_RM:
123 		/* Round towards -Inf: up if negative, down if positive. */
124 		if (fp->fp_sign)
125 			break;
126 		goto rounddown;
127 
128 	case FSR_RD_RP:
129 		/* Round towards +Inf: up if positive, down otherwise. */
130 		if (!fp->fp_sign)
131 			break;
132 		goto rounddown;
133 	}
134 
135 	/* Bump low bit of mantissa, with carry. */
136 #ifdef sparc /* ``cheating'' (left out FPU_DECL_CARRY; know this is faster) */
137 	FPU_ADDS(m3, m3, 1);
138 	FPU_ADDCS(m2, m2, 0);
139 	FPU_ADDCS(m1, m1, 0);
140 	FPU_ADDC(m0, m0, 0);
141 #else
142 	if (++m3 == 0 && ++m2 == 0 && ++m1 == 0)
143 		m0++;
144 #endif
145 	fp->fp_mant[0] = m0;
146 	fp->fp_mant[1] = m1;
147 	fp->fp_mant[2] = m2;
148 	fp->fp_mant[3] = m3;
149 	return (1);
150 
151 rounddown:
152 	fp->fp_mant[0] = m0;
153 	fp->fp_mant[1] = m1;
154 	fp->fp_mant[2] = m2;
155 	fp->fp_mant[3] = m3;
156 	return (0);
157 }
158 
159 /*
160  * For overflow: return true if overflow is to go to +/-Inf, according
161  * to the sign of the overflowing result.  If false, overflow is to go
162  * to the largest magnitude value instead.
163  */
164 static int
165 toinf(struct fpemu *fe, int sign)
166 {
167 	int inf;
168 
169 	/* look at rounding direction */
170 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
171 
172 	default:
173 	case FSR_RD_RN:		/* the nearest value is always Inf */
174 		inf = 1;
175 		break;
176 
177 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
178 		inf = 0;
179 		break;
180 
181 	case FSR_RD_RP:		/* toward +Inf iff positive */
182 		inf = sign == 0;
183 		break;
184 
185 	case FSR_RD_RM:		/* toward -Inf iff negative */
186 		inf = sign;
187 		break;
188 	}
189 	return (inf);
190 }
191 
192 /*
193  * fpn -> int (int value returned as return value).
194  *
195  * N.B.: this conversion always rounds towards zero (this is a peculiarity
196  * of the SPARC instruction set).
197  */
198 u_int
199 fpu_ftoi(fe, fp)
200 	struct fpemu *fe;
201 	register struct fpn *fp;
202 {
203 	register u_int i;
204 	register int sign, exp;
205 
206 	sign = fp->fp_sign;
207 	switch (fp->fp_class) {
208 
209 	case FPC_ZERO:
210 		return (0);
211 
212 	case FPC_NUM:
213 		/*
214 		 * If exp >= 2^32, overflow.  Otherwise shift value right
215 		 * into last mantissa word (this will not exceed 0xffffffff),
216 		 * shifting any guard and round bits out into the sticky
217 		 * bit.  Then ``round'' towards zero, i.e., just set an
218 		 * inexact exception if sticky is set (see round()).
219 		 * If the result is > 0x80000000, or is positive and equals
220 		 * 0x80000000, overflow; otherwise the last fraction word
221 		 * is the result.
222 		 */
223 		if ((exp = fp->fp_exp) >= 32)
224 			break;
225 		/* NB: the following includes exp < 0 cases */
226 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
227 			fe->fe_cx |= FSR_NX;
228 		i = fp->fp_mant[3];
229 		if (i >= ((u_int)0x80000000 + sign))
230 			break;
231 		return (sign ? -i : i);
232 
233 	default:		/* Inf, qNaN, sNaN */
234 		break;
235 	}
236 	/* overflow: replace any inexact exception with invalid */
237 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
238 	return (0x7fffffff + sign);
239 }
240 
241 /*
242  * fpn -> single (32 bit single returned as return value).
243  * We assume <= 29 bits in a single-precision fraction (1.f part).
244  */
245 u_int
246 fpu_ftos(fe, fp)
247 	struct fpemu *fe;
248 	register struct fpn *fp;
249 {
250 	register u_int sign = fp->fp_sign << 31;
251 	register int exp;
252 
253 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
254 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
255 
256 	/* Take care of non-numbers first. */
257 	if (ISNAN(fp)) {
258 		/*
259 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
260 		 * Note that fp->fp_mant[0] has the quiet bit set,
261 		 * even if it is classified as a signalling NaN.
262 		 */
263 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
264 		exp = SNG_EXP_INFNAN;
265 		goto done;
266 	}
267 	if (ISINF(fp))
268 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
269 	if (ISZERO(fp))
270 		return (sign);
271 
272 	/*
273 	 * Normals (including subnormals).  Drop all the fraction bits
274 	 * (including the explicit ``implied'' 1 bit) down into the
275 	 * single-precision range.  If the number is subnormal, move
276 	 * the ``implied'' 1 into the explicit range as well, and shift
277 	 * right to introduce leading zeroes.  Rounding then acts
278 	 * differently for normals and subnormals: the largest subnormal
279 	 * may round to the smallest normal (1.0 x 2^minexp), or may
280 	 * remain subnormal.  In the latter case, signal an underflow
281 	 * if the result was inexact or if underflow traps are enabled.
282 	 *
283 	 * Rounding a normal, on the other hand, always produces another
284 	 * normal (although either way the result might be too big for
285 	 * single precision, and cause an overflow).  If rounding a
286 	 * normal produces 2.0 in the fraction, we need not adjust that
287 	 * fraction at all, since both 1.0 and 2.0 are zero under the
288 	 * fraction mask.
289 	 *
290 	 * Note that the guard and round bits vanish from the number after
291 	 * rounding.
292 	 */
293 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
294 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
295 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
296 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
297 			return (sign | SNG_EXP(1) | 0);
298 		if ((fe->fe_cx & FSR_NX) ||
299 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
300 			fe->fe_cx |= FSR_UF;
301 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
302 	}
303 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
304 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
305 #ifdef DIAGNOSTIC
306 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
307 		panic("fpu_ftos");
308 #endif
309 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
310 		exp++;
311 	if (exp >= SNG_EXP_INFNAN) {
312 		/* overflow to inf or to max single */
313 		fe->fe_cx |= FSR_OF | FSR_NX;
314 		if (toinf(fe, sign))
315 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
316 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
317 	}
318 done:
319 	/* phew, made it */
320 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
321 }
322 
323 /*
324  * fpn -> double (32 bit high-order result returned; 32-bit low order result
325  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
326  *
327  * This code mimics fpu_ftos; see it for comments.
328  */
329 u_int
330 fpu_ftod(fe, fp, res)
331 	struct fpemu *fe;
332 	register struct fpn *fp;
333 	u_int *res;
334 {
335 	register u_int sign = fp->fp_sign << 31;
336 	register int exp;
337 
338 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
339 #define	DBL_MASK	(DBL_EXP(1) - 1)
340 
341 	if (ISNAN(fp)) {
342 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
343 		exp = DBL_EXP_INFNAN;
344 		goto done;
345 	}
346 	if (ISINF(fp)) {
347 		sign |= DBL_EXP(DBL_EXP_INFNAN);
348 		goto zero;
349 	}
350 	if (ISZERO(fp)) {
351 zero:		res[1] = 0;
352 		return (sign);
353 	}
354 
355 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
356 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
357 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
358 			res[1] = 0;
359 			return (sign | DBL_EXP(1) | 0);
360 		}
361 		if ((fe->fe_cx & FSR_NX) ||
362 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
363 			fe->fe_cx |= FSR_UF;
364 		exp = 0;
365 		goto done;
366 	}
367 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
368 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
369 		exp++;
370 	if (exp >= DBL_EXP_INFNAN) {
371 		fe->fe_cx |= FSR_OF | FSR_NX;
372 		if (toinf(fe, sign)) {
373 			res[1] = 0;
374 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
375 		}
376 		res[1] = ~0;
377 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
378 	}
379 done:
380 	res[1] = fp->fp_mant[3];
381 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
382 }
383 
384 /*
385  * fpn -> extended (32 bit high-order result returned; low-order fraction
386  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
387  * our internal format *is* extended precision, plus 2 bits for guard/round,
388  * so we can avoid a small bit of work.
389  */
390 u_int
391 fpu_ftox(fe, fp, res)
392 	struct fpemu *fe;
393 	register struct fpn *fp;
394 	u_int *res;
395 {
396 	register u_int sign = fp->fp_sign << 31;
397 	register int exp;
398 
399 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
400 #define	EXT_MASK	(EXT_EXP(1) - 1)
401 
402 	if (ISNAN(fp)) {
403 		(void) fpu_shr(fp, 2);	/* since we are not rounding */
404 		exp = EXT_EXP_INFNAN;
405 		goto done;
406 	}
407 	if (ISINF(fp)) {
408 		sign |= EXT_EXP(EXT_EXP_INFNAN);
409 		goto zero;
410 	}
411 	if (ISZERO(fp)) {
412 zero:		res[1] = res[2] = res[3] = 0;
413 		return (sign);
414 	}
415 
416 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
417 		(void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
418 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
419 			res[1] = res[2] = res[3] = 0;
420 			return (sign | EXT_EXP(1) | 0);
421 		}
422 		if ((fe->fe_cx & FSR_NX) ||
423 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
424 			fe->fe_cx |= FSR_UF;
425 		exp = 0;
426 		goto done;
427 	}
428 	/* Since internal == extended, no need to shift here. */
429 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
430 		exp++;
431 	if (exp >= EXT_EXP_INFNAN) {
432 		fe->fe_cx |= FSR_OF | FSR_NX;
433 		if (toinf(fe, sign)) {
434 			res[1] = res[2] = res[3] = 0;
435 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
436 		}
437 		res[1] = res[2] = res[3] = ~0;
438 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
439 	}
440 done:
441 	res[1] = fp->fp_mant[1];
442 	res[2] = fp->fp_mant[2];
443 	res[3] = fp->fp_mant[3];
444 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
445 }
446 
447 /*
448  * Implode an fpn, writing the result into the given space.
449  */
450 void
451 fpu_implode(fe, fp, type, space)
452 	struct fpemu *fe;
453 	register struct fpn *fp;
454 	int type;
455 	register u_int *space;
456 {
457 
458 	switch (type) {
459 
460 	case FTYPE_INT:
461 		space[0] = fpu_ftoi(fe, fp);
462 		break;
463 
464 	case FTYPE_SNG:
465 		space[0] = fpu_ftos(fe, fp);
466 		break;
467 
468 	case FTYPE_DBL:
469 		space[0] = fpu_ftod(fe, fp, space);
470 		break;
471 
472 	case FTYPE_EXT:
473 		/* funky rounding precision options ?? */
474 		space[0] = fpu_ftox(fe, fp, space);
475 		break;
476 
477 	default:
478 		panic("fpu_implode");
479 	}
480 }
481