xref: /netbsd-src/sys/arch/sparc/fpu/fpu_implode.c (revision 1f2744e6e4915c9da2a3f980279398c4cf7d5e6d)
1 /*	$NetBSD: fpu_implode.c,v 1.2 1994/11/20 20:52:42 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_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 
54 #include <machine/ieee.h>
55 #include <machine/instr.h>
56 #include <machine/reg.h>
57 
58 #include <sparc/fpu/fpu_arith.h>
59 #include <sparc/fpu/fpu_emu.h>
60 
61 /*
62  * Round a number (algorithm from Motorola MC68882 manual, modified for
63  * our internal format).  Set inexact exception if rounding is required.
64  * Return true iff we rounded up.
65  *
66  * After rounding, we discard the guard and round bits by shifting right
67  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
68  * This saves effort later.
69  *
70  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
71  * responsibility to fix this if necessary.
72  */
73 static int
74 round(register struct fpemu *fe, register struct fpn *fp)
75 {
76 	register u_int m0, m1, m2, m3;
77 	register int gr, s, ret;
78 
79 	m0 = fp->fp_mant[0];
80 	m1 = fp->fp_mant[1];
81 	m2 = fp->fp_mant[2];
82 	m3 = fp->fp_mant[3];
83 	gr = m3 & 3;
84 	s = fp->fp_sticky;
85 
86 	/* mant >>= FP_NG */
87 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
88 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
89 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
90 	m0 >>= FP_NG;
91 
92 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
93 		goto rounddown;
94 
95 	fe->fe_cx |= FSR_NX;	/* inexact */
96 
97 	/* Go to rounddown to round down; break to round up. */
98 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
99 
100 	case FSR_RD_RN:
101 	default:
102 		/*
103 		 * Round only if guard is set (gr & 2).  If guard is set,
104 		 * but round & sticky both clear, then we want to round
105 		 * but have a tie, so round to even, i.e., add 1 iff odd.
106 		 */
107 		if ((gr & 2) == 0)
108 			goto rounddown;
109 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
110 			break;
111 		goto rounddown;
112 
113 	case FSR_RD_RZ:
114 		/* Round towards zero, i.e., down. */
115 		goto rounddown;
116 
117 	case FSR_RD_RM:
118 		/* Round towards -Inf: up if negative, down if positive. */
119 		if (fp->fp_sign)
120 			break;
121 		goto rounddown;
122 
123 	case FSR_RD_RP:
124 		/* Round towards +Inf: up if positive, down otherwise. */
125 		if (!fp->fp_sign)
126 			break;
127 		goto rounddown;
128 	}
129 
130 	/* Bump low bit of mantissa, with carry. */
131 #ifdef sparc /* ``cheating'' (left out FPU_DECL_CARRY; know this is faster) */
132 	FPU_ADDS(m3, m3, 1);
133 	FPU_ADDCS(m2, m2, 0);
134 	FPU_ADDCS(m1, m1, 0);
135 	FPU_ADDC(m0, m0, 0);
136 #else
137 	if (++m3 == 0 && ++m2 == 0 && ++m1 == 0)
138 		m0++;
139 #endif
140 	fp->fp_mant[0] = m0;
141 	fp->fp_mant[1] = m1;
142 	fp->fp_mant[2] = m2;
143 	fp->fp_mant[3] = m3;
144 	return (1);
145 
146 rounddown:
147 	fp->fp_mant[0] = m0;
148 	fp->fp_mant[1] = m1;
149 	fp->fp_mant[2] = m2;
150 	fp->fp_mant[3] = m3;
151 	return (0);
152 }
153 
154 /*
155  * For overflow: return true if overflow is to go to +/-Inf, according
156  * to the sign of the overflowing result.  If false, overflow is to go
157  * to the largest magnitude value instead.
158  */
159 static int
160 toinf(struct fpemu *fe, int sign)
161 {
162 	int inf;
163 
164 	/* look at rounding direction */
165 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
166 
167 	default:
168 	case FSR_RD_RN:		/* the nearest value is always Inf */
169 		inf = 1;
170 		break;
171 
172 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
173 		inf = 0;
174 		break;
175 
176 	case FSR_RD_RP:		/* toward +Inf iff positive */
177 		inf = sign == 0;
178 		break;
179 
180 	case FSR_RD_RM:		/* toward -Inf iff negative */
181 		inf = sign;
182 		break;
183 	}
184 	return (inf);
185 }
186 
187 /*
188  * fpn -> int (int value returned as return value).
189  *
190  * N.B.: this conversion always rounds towards zero (this is a peculiarity
191  * of the SPARC instruction set).
192  */
193 u_int
194 fpu_ftoi(fe, fp)
195 	struct fpemu *fe;
196 	register struct fpn *fp;
197 {
198 	register u_int i;
199 	register int sign, exp;
200 
201 	sign = fp->fp_sign;
202 	switch (fp->fp_class) {
203 
204 	case FPC_ZERO:
205 		return (0);
206 
207 	case FPC_NUM:
208 		/*
209 		 * If exp >= 2^32, overflow.  Otherwise shift value right
210 		 * into last mantissa word (this will not exceed 0xffffffff),
211 		 * shifting any guard and round bits out into the sticky
212 		 * bit.  Then ``round'' towards zero, i.e., just set an
213 		 * inexact exception if sticky is set (see round()).
214 		 * If the result is > 0x80000000, or is positive and equals
215 		 * 0x80000000, overflow; otherwise the last fraction word
216 		 * is the result.
217 		 */
218 		if ((exp = fp->fp_exp) >= 32)
219 			break;
220 		/* NB: the following includes exp < 0 cases */
221 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
222 			fe->fe_cx |= FSR_NX;
223 		i = fp->fp_mant[3];
224 		if (i >= ((u_int)0x80000000 + sign))
225 			break;
226 		return (sign ? -i : i);
227 
228 	default:		/* Inf, qNaN, sNaN */
229 		break;
230 	}
231 	/* overflow: replace any inexact exception with invalid */
232 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
233 	return (0x7fffffff + sign);
234 }
235 
236 /*
237  * fpn -> single (32 bit single returned as return value).
238  * We assume <= 29 bits in a single-precision fraction (1.f part).
239  */
240 u_int
241 fpu_ftos(fe, fp)
242 	struct fpemu *fe;
243 	register struct fpn *fp;
244 {
245 	register u_int sign = fp->fp_sign << 31;
246 	register int exp;
247 
248 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
249 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
250 
251 	/* Take care of non-numbers first. */
252 	if (ISNAN(fp)) {
253 		/*
254 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
255 		 * Note that fp->fp_mant[0] has the quiet bit set,
256 		 * even if it is classified as a signalling NaN.
257 		 */
258 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
259 		exp = SNG_EXP_INFNAN;
260 		goto done;
261 	}
262 	if (ISINF(fp))
263 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
264 	if (ISZERO(fp))
265 		return (sign);
266 
267 	/*
268 	 * Normals (including subnormals).  Drop all the fraction bits
269 	 * (including the explicit ``implied'' 1 bit) down into the
270 	 * single-precision range.  If the number is subnormal, move
271 	 * the ``implied'' 1 into the explicit range as well, and shift
272 	 * right to introduce leading zeroes.  Rounding then acts
273 	 * differently for normals and subnormals: the largest subnormal
274 	 * may round to the smallest normal (1.0 x 2^minexp), or may
275 	 * remain subnormal.  In the latter case, signal an underflow
276 	 * if the result was inexact or if underflow traps are enabled.
277 	 *
278 	 * Rounding a normal, on the other hand, always produces another
279 	 * normal (although either way the result might be too big for
280 	 * single precision, and cause an overflow).  If rounding a
281 	 * normal produces 2.0 in the fraction, we need not adjust that
282 	 * fraction at all, since both 1.0 and 2.0 are zero under the
283 	 * fraction mask.
284 	 *
285 	 * Note that the guard and round bits vanish from the number after
286 	 * rounding.
287 	 */
288 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
289 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
290 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
291 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
292 			return (sign | SNG_EXP(1) | 0);
293 		if ((fe->fe_cx & FSR_NX) ||
294 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
295 			fe->fe_cx |= FSR_UF;
296 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
297 	}
298 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
299 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
300 #ifdef DIAGNOSTIC
301 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
302 		panic("fpu_ftos");
303 #endif
304 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
305 		exp++;
306 	if (exp >= SNG_EXP_INFNAN) {
307 		/* overflow to inf or to max single */
308 		fe->fe_cx |= FSR_OF | FSR_NX;
309 		if (toinf(fe, sign))
310 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
311 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
312 	}
313 done:
314 	/* phew, made it */
315 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
316 }
317 
318 /*
319  * fpn -> double (32 bit high-order result returned; 32-bit low order result
320  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
321  *
322  * This code mimics fpu_ftos; see it for comments.
323  */
324 u_int
325 fpu_ftod(fe, fp, res)
326 	struct fpemu *fe;
327 	register struct fpn *fp;
328 	u_int *res;
329 {
330 	register u_int sign = fp->fp_sign << 31;
331 	register int exp;
332 
333 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
334 #define	DBL_MASK	(DBL_EXP(1) - 1)
335 
336 	if (ISNAN(fp)) {
337 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
338 		exp = DBL_EXP_INFNAN;
339 		goto done;
340 	}
341 	if (ISINF(fp)) {
342 		sign |= DBL_EXP(DBL_EXP_INFNAN);
343 		goto zero;
344 	}
345 	if (ISZERO(fp)) {
346 zero:		res[1] = 0;
347 		return (sign);
348 	}
349 
350 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
351 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
352 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
353 			res[1] = 0;
354 			return (sign | DBL_EXP(1) | 0);
355 		}
356 		if ((fe->fe_cx & FSR_NX) ||
357 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
358 			fe->fe_cx |= FSR_UF;
359 		exp = 0;
360 		goto done;
361 	}
362 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
363 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
364 		exp++;
365 	if (exp >= DBL_EXP_INFNAN) {
366 		fe->fe_cx |= FSR_OF | FSR_NX;
367 		if (toinf(fe, sign)) {
368 			res[1] = 0;
369 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
370 		}
371 		res[1] = ~0;
372 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
373 	}
374 done:
375 	res[1] = fp->fp_mant[3];
376 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
377 }
378 
379 /*
380  * fpn -> extended (32 bit high-order result returned; low-order fraction
381  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
382  * our internal format *is* extended precision, plus 2 bits for guard/round,
383  * so we can avoid a small bit of work.
384  */
385 u_int
386 fpu_ftox(fe, fp, res)
387 	struct fpemu *fe;
388 	register struct fpn *fp;
389 	u_int *res;
390 {
391 	register u_int sign = fp->fp_sign << 31;
392 	register int exp;
393 
394 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
395 #define	EXT_MASK	(EXT_EXP(1) - 1)
396 
397 	if (ISNAN(fp)) {
398 		(void) fpu_shr(fp, 2);	/* since we are not rounding */
399 		exp = EXT_EXP_INFNAN;
400 		goto done;
401 	}
402 	if (ISINF(fp)) {
403 		sign |= EXT_EXP(EXT_EXP_INFNAN);
404 		goto zero;
405 	}
406 	if (ISZERO(fp)) {
407 zero:		res[1] = res[2] = res[3] = 0;
408 		return (sign);
409 	}
410 
411 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
412 		(void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
413 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
414 			res[1] = res[2] = res[3] = 0;
415 			return (sign | EXT_EXP(1) | 0);
416 		}
417 		if ((fe->fe_cx & FSR_NX) ||
418 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
419 			fe->fe_cx |= FSR_UF;
420 		exp = 0;
421 		goto done;
422 	}
423 	/* Since internal == extended, no need to shift here. */
424 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
425 		exp++;
426 	if (exp >= EXT_EXP_INFNAN) {
427 		fe->fe_cx |= FSR_OF | FSR_NX;
428 		if (toinf(fe, sign)) {
429 			res[1] = res[2] = res[3] = 0;
430 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
431 		}
432 		res[1] = res[2] = res[3] = ~0;
433 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
434 	}
435 done:
436 	res[1] = fp->fp_mant[1];
437 	res[2] = fp->fp_mant[2];
438 	res[3] = fp->fp_mant[3];
439 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
440 }
441 
442 /*
443  * Implode an fpn, writing the result into the given space.
444  */
445 void
446 fpu_implode(fe, fp, type, space)
447 	struct fpemu *fe;
448 	register struct fpn *fp;
449 	int type;
450 	register u_int *space;
451 {
452 
453 	switch (type) {
454 
455 	case FTYPE_INT:
456 		space[0] = fpu_ftoi(fe, fp);
457 		break;
458 
459 	case FTYPE_SNG:
460 		space[0] = fpu_ftos(fe, fp);
461 		break;
462 
463 	case FTYPE_DBL:
464 		space[0] = fpu_ftod(fe, fp, space);
465 		break;
466 
467 	case FTYPE_EXT:
468 		/* funky rounding precision options ?? */
469 		space[0] = fpu_ftox(fe, fp, space);
470 		break;
471 
472 	default:
473 		panic("fpu_implode");
474 	}
475 }
476