xref: /netbsd-src/sys/arch/sparc/fpu/fpu_implode.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: fpu_implode.c,v 1.7 2000/08/03 18:32:08 eeh 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 	FPU_ADDS(m3, m3, 1);
137 	FPU_ADDCS(m2, m2, 0);
138 	FPU_ADDCS(m1, m1, 0);
139 	FPU_ADDC(m0, m0, 0);
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 #ifdef SUN4U
237 /*
238  * fpn -> extended int (high bits of int value returned as return value).
239  *
240  * N.B.: this conversion always rounds towards zero (this is a peculiarity
241  * of the SPARC instruction set).
242  */
243 u_int
244 fpu_ftox(fe, fp, res)
245 	struct fpemu *fe;
246 	register struct fpn *fp;
247 	u_int *res;
248 {
249 	register u_int64_t i;
250 	register int sign, exp;
251 
252 	sign = fp->fp_sign;
253 	switch (fp->fp_class) {
254 
255 	case FPC_ZERO:
256 		res[1] = 0;
257 		return (0);
258 
259 	case FPC_NUM:
260 		/*
261 		 * If exp >= 2^64, overflow.  Otherwise shift value right
262 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
263 		 * shifting any guard and round bits out into the sticky
264 		 * bit.  Then ``round'' towards zero, i.e., just set an
265 		 * inexact exception if sticky is set (see round()).
266 		 * If the result is > 0x8000000000000000, or is positive and equals
267 		 * 0x8000000000000000, overflow; otherwise the last fraction word
268 		 * is the result.
269 		 */
270 		if ((exp = fp->fp_exp) >= 64)
271 			break;
272 		/* NB: the following includes exp < 0 cases */
273 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
274 			fe->fe_cx |= FSR_NX;
275 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
276 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
277 			break;
278 		return (sign ? -i : i);
279 
280 	default:		/* Inf, qNaN, sNaN */
281 		break;
282 	}
283 	/* overflow: replace any inexact exception with invalid */
284 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
285 	return (0x7fffffffffffffffLL + sign);
286 }
287 #endif /* SUN4U */
288 
289 /*
290  * fpn -> single (32 bit single returned as return value).
291  * We assume <= 29 bits in a single-precision fraction (1.f part).
292  */
293 u_int
294 fpu_ftos(fe, fp)
295 	struct fpemu *fe;
296 	register struct fpn *fp;
297 {
298 	register u_int sign = fp->fp_sign << 31;
299 	register int exp;
300 
301 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
302 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
303 
304 	/* Take care of non-numbers first. */
305 	if (ISNAN(fp)) {
306 		/*
307 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
308 		 * Note that fp->fp_mant[0] has the quiet bit set,
309 		 * even if it is classified as a signalling NaN.
310 		 */
311 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
312 		exp = SNG_EXP_INFNAN;
313 		goto done;
314 	}
315 	if (ISINF(fp))
316 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
317 	if (ISZERO(fp))
318 		return (sign);
319 
320 	/*
321 	 * Normals (including subnormals).  Drop all the fraction bits
322 	 * (including the explicit ``implied'' 1 bit) down into the
323 	 * single-precision range.  If the number is subnormal, move
324 	 * the ``implied'' 1 into the explicit range as well, and shift
325 	 * right to introduce leading zeroes.  Rounding then acts
326 	 * differently for normals and subnormals: the largest subnormal
327 	 * may round to the smallest normal (1.0 x 2^minexp), or may
328 	 * remain subnormal.  In the latter case, signal an underflow
329 	 * if the result was inexact or if underflow traps are enabled.
330 	 *
331 	 * Rounding a normal, on the other hand, always produces another
332 	 * normal (although either way the result might be too big for
333 	 * single precision, and cause an overflow).  If rounding a
334 	 * normal produces 2.0 in the fraction, we need not adjust that
335 	 * fraction at all, since both 1.0 and 2.0 are zero under the
336 	 * fraction mask.
337 	 *
338 	 * Note that the guard and round bits vanish from the number after
339 	 * rounding.
340 	 */
341 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
342 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
343 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
344 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
345 			return (sign | SNG_EXP(1) | 0);
346 		if ((fe->fe_cx & FSR_NX) ||
347 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
348 			fe->fe_cx |= FSR_UF;
349 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
350 	}
351 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
352 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
353 #ifdef DIAGNOSTIC
354 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
355 		panic("fpu_ftos");
356 #endif
357 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
358 		exp++;
359 	if (exp >= SNG_EXP_INFNAN) {
360 		/* overflow to inf or to max single */
361 		fe->fe_cx |= FSR_OF | FSR_NX;
362 		if (toinf(fe, sign))
363 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
364 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
365 	}
366 done:
367 	/* phew, made it */
368 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
369 }
370 
371 /*
372  * fpn -> double (32 bit high-order result returned; 32-bit low order result
373  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
374  *
375  * This code mimics fpu_ftos; see it for comments.
376  */
377 u_int
378 fpu_ftod(fe, fp, res)
379 	struct fpemu *fe;
380 	register struct fpn *fp;
381 	u_int *res;
382 {
383 	register u_int sign = fp->fp_sign << 31;
384 	register int exp;
385 
386 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
387 #define	DBL_MASK	(DBL_EXP(1) - 1)
388 
389 	if (ISNAN(fp)) {
390 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
391 		exp = DBL_EXP_INFNAN;
392 		goto done;
393 	}
394 	if (ISINF(fp)) {
395 		sign |= DBL_EXP(DBL_EXP_INFNAN);
396 		goto zero;
397 	}
398 	if (ISZERO(fp)) {
399 zero:		res[1] = 0;
400 		return (sign);
401 	}
402 
403 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
404 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
405 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
406 			res[1] = 0;
407 			return (sign | DBL_EXP(1) | 0);
408 		}
409 		if ((fe->fe_cx & FSR_NX) ||
410 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
411 			fe->fe_cx |= FSR_UF;
412 		exp = 0;
413 		goto done;
414 	}
415 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
416 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
417 		exp++;
418 	if (exp >= DBL_EXP_INFNAN) {
419 		fe->fe_cx |= FSR_OF | FSR_NX;
420 		if (toinf(fe, sign)) {
421 			res[1] = 0;
422 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
423 		}
424 		res[1] = ~0;
425 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
426 	}
427 done:
428 	res[1] = fp->fp_mant[3];
429 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
430 }
431 
432 /*
433  * fpn -> extended (32 bit high-order result returned; low-order fraction
434  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
435  * our internal format *is* extended precision, plus 2 bits for guard/round,
436  * so we can avoid a small bit of work.
437  */
438 u_int
439 fpu_ftoq(fe, fp, res)
440 	struct fpemu *fe;
441 	register struct fpn *fp;
442 	u_int *res;
443 {
444 	register u_int sign = fp->fp_sign << 31;
445 	register int exp;
446 
447 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
448 #define	EXT_MASK	(EXT_EXP(1) - 1)
449 
450 	if (ISNAN(fp)) {
451 		(void) fpu_shr(fp, 2);	/* since we are not rounding */
452 		exp = EXT_EXP_INFNAN;
453 		goto done;
454 	}
455 	if (ISINF(fp)) {
456 		sign |= EXT_EXP(EXT_EXP_INFNAN);
457 		goto zero;
458 	}
459 	if (ISZERO(fp)) {
460 zero:		res[1] = res[2] = res[3] = 0;
461 		return (sign);
462 	}
463 
464 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
465 		(void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
466 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
467 			res[1] = res[2] = res[3] = 0;
468 			return (sign | EXT_EXP(1) | 0);
469 		}
470 		if ((fe->fe_cx & FSR_NX) ||
471 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
472 			fe->fe_cx |= FSR_UF;
473 		exp = 0;
474 		goto done;
475 	}
476 	/* Since internal == extended, no need to shift here. */
477 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
478 		exp++;
479 	if (exp >= EXT_EXP_INFNAN) {
480 		fe->fe_cx |= FSR_OF | FSR_NX;
481 		if (toinf(fe, sign)) {
482 			res[1] = res[2] = res[3] = 0;
483 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
484 		}
485 		res[1] = res[2] = res[3] = ~0;
486 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
487 	}
488 done:
489 	res[1] = fp->fp_mant[1];
490 	res[2] = fp->fp_mant[2];
491 	res[3] = fp->fp_mant[3];
492 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
493 }
494 
495 /*
496  * Implode an fpn, writing the result into the given space.
497  */
498 void
499 fpu_implode(fe, fp, type, space)
500 	struct fpemu *fe;
501 	register struct fpn *fp;
502 	int type;
503 	register u_int *space;
504 {
505 
506 	switch (type) {
507 
508 #ifdef SUN4U
509 	case FTYPE_LNG:
510 		space[0] = fpu_ftox(fe, fp, space);
511 		break;
512 #endif /* SUN4U */
513 
514 	case FTYPE_INT:
515 		space[0] = fpu_ftoi(fe, fp);
516 		break;
517 
518 	case FTYPE_SNG:
519 		space[0] = fpu_ftos(fe, fp);
520 		break;
521 
522 	case FTYPE_DBL:
523 		space[0] = fpu_ftod(fe, fp, space);
524 		break;
525 
526 	case FTYPE_EXT:
527 		/* funky rounding precision options ?? */
528 		space[0] = fpu_ftoq(fe, fp, space);
529 		break;
530 
531 	default:
532 		panic("fpu_implode");
533 	}
534 }
535