xref: /netbsd-src/sys/arch/sparc/fpu/fpu_implode.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: fpu_implode.c,v 1.12 2003/08/07 16:29:37 agc 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. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	@(#)fpu_implode.c	8.1 (Berkeley) 6/11/93
41  */
42 
43 /*
44  * FPU subroutines: `implode' internal format numbers into the machine's
45  * `packed binary' format.
46  */
47 
48 #include <sys/cdefs.h>
49 __KERNEL_RCSID(0, "$NetBSD: fpu_implode.c,v 1.12 2003/08/07 16:29:37 agc Exp $");
50 
51 #if defined(_KERNEL_OPT)
52 #include "opt_sparc_arch.h"
53 #endif
54 
55 #include <sys/types.h>
56 #include <sys/systm.h>
57 
58 #include <machine/ieee.h>
59 #include <machine/instr.h>
60 #include <machine/reg.h>
61 
62 #include <sparc/fpu/fpu_arith.h>
63 #include <sparc/fpu/fpu_emu.h>
64 #include <sparc/fpu/fpu_extern.h>
65 
66 static int round __P((register struct fpemu *, register struct fpn *));
67 static int toinf __P((struct fpemu *, int));
68 
69 /*
70  * Round a number (algorithm from Motorola MC68882 manual, modified for
71  * our internal format).  Set inexact exception if rounding is required.
72  * Return true iff we rounded up.
73  *
74  * After rounding, we discard the guard and round bits by shifting right
75  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
76  * This saves effort later.
77  *
78  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
79  * responsibility to fix this if necessary.
80  */
81 static int
82 round(register struct fpemu *fe, register struct fpn *fp)
83 {
84 	register u_int m0, m1, m2, m3;
85 	register int gr, s;
86 
87 	m0 = fp->fp_mant[0];
88 	m1 = fp->fp_mant[1];
89 	m2 = fp->fp_mant[2];
90 	m3 = fp->fp_mant[3];
91 	gr = m3 & 3;
92 	s = fp->fp_sticky;
93 
94 	/* mant >>= FP_NG */
95 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
96 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
97 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
98 	m0 >>= FP_NG;
99 
100 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
101 		goto rounddown;
102 
103 	fe->fe_cx |= FSR_NX;	/* inexact */
104 
105 	/* Go to rounddown to round down; break to round up. */
106 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
107 
108 	case FSR_RD_RN:
109 	default:
110 		/*
111 		 * Round only if guard is set (gr & 2).  If guard is set,
112 		 * but round & sticky both clear, then we want to round
113 		 * but have a tie, so round to even, i.e., add 1 iff odd.
114 		 */
115 		if ((gr & 2) == 0)
116 			goto rounddown;
117 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
118 			break;
119 		goto rounddown;
120 
121 	case FSR_RD_RZ:
122 		/* Round towards zero, i.e., down. */
123 		goto rounddown;
124 
125 	case FSR_RD_RM:
126 		/* Round towards -Inf: up if negative, down if positive. */
127 		if (fp->fp_sign)
128 			break;
129 		goto rounddown;
130 
131 	case FSR_RD_RP:
132 		/* Round towards +Inf: up if positive, down otherwise. */
133 		if (!fp->fp_sign)
134 			break;
135 		goto rounddown;
136 	}
137 
138 	/* Bump low bit of mantissa, with carry. */
139 	FPU_ADDS(m3, m3, 1);
140 	FPU_ADDCS(m2, m2, 0);
141 	FPU_ADDCS(m1, m1, 0);
142 	FPU_ADDC(m0, m0, 0);
143 	fp->fp_mant[0] = m0;
144 	fp->fp_mant[1] = m1;
145 	fp->fp_mant[2] = m2;
146 	fp->fp_mant[3] = m3;
147 	return (1);
148 
149 rounddown:
150 	fp->fp_mant[0] = m0;
151 	fp->fp_mant[1] = m1;
152 	fp->fp_mant[2] = m2;
153 	fp->fp_mant[3] = m3;
154 	return (0);
155 }
156 
157 /*
158  * For overflow: return true if overflow is to go to +/-Inf, according
159  * to the sign of the overflowing result.  If false, overflow is to go
160  * to the largest magnitude value instead.
161  */
162 static int
163 toinf(struct fpemu *fe, int sign)
164 {
165 	int inf;
166 
167 	/* look at rounding direction */
168 	switch ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK) {
169 
170 	default:
171 	case FSR_RD_RN:		/* the nearest value is always Inf */
172 		inf = 1;
173 		break;
174 
175 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
176 		inf = 0;
177 		break;
178 
179 	case FSR_RD_RP:		/* toward +Inf iff positive */
180 		inf = sign == 0;
181 		break;
182 
183 	case FSR_RD_RM:		/* toward -Inf iff negative */
184 		inf = sign;
185 		break;
186 	}
187 	return (inf);
188 }
189 
190 /*
191  * fpn -> int (int value returned as return value).
192  *
193  * N.B.: this conversion always rounds towards zero (this is a peculiarity
194  * of the SPARC instruction set).
195  */
196 u_int
197 fpu_ftoi(fe, fp)
198 	struct fpemu *fe;
199 	register struct fpn *fp;
200 {
201 	register u_int i;
202 	register int sign, exp;
203 
204 	sign = fp->fp_sign;
205 	switch (fp->fp_class) {
206 
207 	case FPC_ZERO:
208 		return (0);
209 
210 	case FPC_NUM:
211 		/*
212 		 * If exp >= 2^32, overflow.  Otherwise shift value right
213 		 * into last mantissa word (this will not exceed 0xffffffff),
214 		 * shifting any guard and round bits out into the sticky
215 		 * bit.  Then ``round'' towards zero, i.e., just set an
216 		 * inexact exception if sticky is set (see round()).
217 		 * If the result is > 0x80000000, or is positive and equals
218 		 * 0x80000000, overflow; otherwise the last fraction word
219 		 * is the result.
220 		 */
221 		if ((exp = fp->fp_exp) >= 32)
222 			break;
223 		/* NB: the following includes exp < 0 cases */
224 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
225 			fe->fe_cx |= FSR_NX;
226 		i = fp->fp_mant[3];
227 		if (i >= ((u_int)0x80000000 + sign))
228 			break;
229 		return (sign ? -i : i);
230 
231 	default:		/* Inf, qNaN, sNaN */
232 		break;
233 	}
234 	/* overflow: replace any inexact exception with invalid */
235 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
236 	return (0x7fffffff + sign);
237 }
238 
239 #ifdef SUN4U
240 /*
241  * fpn -> extended int (high bits of int value returned as return value).
242  *
243  * N.B.: this conversion always rounds towards zero (this is a peculiarity
244  * of the SPARC instruction set).
245  */
246 u_int
247 fpu_ftox(fe, fp, res)
248 	struct fpemu *fe;
249 	register struct fpn *fp;
250 	u_int *res;
251 {
252 	register u_int64_t i;
253 	register int sign, exp;
254 
255 	sign = fp->fp_sign;
256 	switch (fp->fp_class) {
257 
258 	case FPC_ZERO:
259 		res[1] = 0;
260 		return (0);
261 
262 	case FPC_NUM:
263 		/*
264 		 * If exp >= 2^64, overflow.  Otherwise shift value right
265 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
266 		 * shifting any guard and round bits out into the sticky
267 		 * bit.  Then ``round'' towards zero, i.e., just set an
268 		 * inexact exception if sticky is set (see round()).
269 		 * If the result is > 0x8000000000000000, or is positive and equals
270 		 * 0x8000000000000000, overflow; otherwise the last fraction word
271 		 * is the result.
272 		 */
273 		if ((exp = fp->fp_exp) >= 64)
274 			break;
275 		/* NB: the following includes exp < 0 cases */
276 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
277 			fe->fe_cx |= FSR_NX;
278 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
279 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
280 			break;
281 		if (sign) i = -i;
282 		res[1] = (int)i;
283 		return (i>>32);
284 
285 	default:		/* Inf, qNaN, sNaN */
286 		break;
287 	}
288 	/* overflow: replace any inexact exception with invalid */
289 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
290 	return (0x7fffffffffffffffLL + sign);
291 }
292 #endif /* SUN4U */
293 
294 /*
295  * fpn -> single (32 bit single returned as return value).
296  * We assume <= 29 bits in a single-precision fraction (1.f part).
297  */
298 u_int
299 fpu_ftos(fe, fp)
300 	struct fpemu *fe;
301 	register struct fpn *fp;
302 {
303 	register u_int sign = fp->fp_sign << 31;
304 	register int exp;
305 
306 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
307 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
308 
309 	/* Take care of non-numbers first. */
310 	if (ISNAN(fp)) {
311 		/*
312 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
313 		 * Note that fp->fp_mant[0] has the quiet bit set,
314 		 * even if it is classified as a signalling NaN.
315 		 */
316 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
317 		exp = SNG_EXP_INFNAN;
318 		goto done;
319 	}
320 	if (ISINF(fp))
321 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
322 	if (ISZERO(fp))
323 		return (sign);
324 
325 	/*
326 	 * Normals (including subnormals).  Drop all the fraction bits
327 	 * (including the explicit ``implied'' 1 bit) down into the
328 	 * single-precision range.  If the number is subnormal, move
329 	 * the ``implied'' 1 into the explicit range as well, and shift
330 	 * right to introduce leading zeroes.  Rounding then acts
331 	 * differently for normals and subnormals: the largest subnormal
332 	 * may round to the smallest normal (1.0 x 2^minexp), or may
333 	 * remain subnormal.  In the latter case, signal an underflow
334 	 * if the result was inexact or if underflow traps are enabled.
335 	 *
336 	 * Rounding a normal, on the other hand, always produces another
337 	 * normal (although either way the result might be too big for
338 	 * single precision, and cause an overflow).  If rounding a
339 	 * normal produces 2.0 in the fraction, we need not adjust that
340 	 * fraction at all, since both 1.0 and 2.0 are zero under the
341 	 * fraction mask.
342 	 *
343 	 * Note that the guard and round bits vanish from the number after
344 	 * rounding.
345 	 */
346 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
347 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
348 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
349 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
350 			return (sign | SNG_EXP(1) | 0);
351 		if ((fe->fe_cx & FSR_NX) ||
352 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
353 			fe->fe_cx |= FSR_UF;
354 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
355 	}
356 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
357 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
358 #ifdef DIAGNOSTIC
359 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
360 		panic("fpu_ftos");
361 #endif
362 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
363 		exp++;
364 	if (exp >= SNG_EXP_INFNAN) {
365 		/* overflow to inf or to max single */
366 		fe->fe_cx |= FSR_OF | FSR_NX;
367 		if (toinf(fe, sign))
368 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
369 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
370 	}
371 done:
372 	/* phew, made it */
373 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
374 }
375 
376 /*
377  * fpn -> double (32 bit high-order result returned; 32-bit low order result
378  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
379  *
380  * This code mimics fpu_ftos; see it for comments.
381  */
382 u_int
383 fpu_ftod(fe, fp, res)
384 	struct fpemu *fe;
385 	register struct fpn *fp;
386 	u_int *res;
387 {
388 	register u_int sign = fp->fp_sign << 31;
389 	register int exp;
390 
391 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
392 #define	DBL_MASK	(DBL_EXP(1) - 1)
393 
394 	if (ISNAN(fp)) {
395 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
396 		exp = DBL_EXP_INFNAN;
397 		goto done;
398 	}
399 	if (ISINF(fp)) {
400 		sign |= DBL_EXP(DBL_EXP_INFNAN);
401 		goto zero;
402 	}
403 	if (ISZERO(fp)) {
404 zero:		res[1] = 0;
405 		return (sign);
406 	}
407 
408 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
409 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
410 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
411 			res[1] = 0;
412 			return (sign | DBL_EXP(1) | 0);
413 		}
414 		if ((fe->fe_cx & FSR_NX) ||
415 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
416 			fe->fe_cx |= FSR_UF;
417 		exp = 0;
418 		goto done;
419 	}
420 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
421 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
422 		exp++;
423 	if (exp >= DBL_EXP_INFNAN) {
424 		fe->fe_cx |= FSR_OF | FSR_NX;
425 		if (toinf(fe, sign)) {
426 			res[1] = 0;
427 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
428 		}
429 		res[1] = ~0;
430 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
431 	}
432 done:
433 	res[1] = fp->fp_mant[3];
434 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
435 }
436 
437 /*
438  * fpn -> extended (32 bit high-order result returned; low-order fraction
439  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
440  * our internal format *is* extended precision, plus 2 bits for guard/round,
441  * so we can avoid a small bit of work.
442  */
443 u_int
444 fpu_ftoq(fe, fp, res)
445 	struct fpemu *fe;
446 	register struct fpn *fp;
447 	u_int *res;
448 {
449 	register u_int sign = fp->fp_sign << 31;
450 	register int exp;
451 
452 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
453 #define	EXT_MASK	(EXT_EXP(1) - 1)
454 
455 	if (ISNAN(fp)) {
456 		(void) fpu_shr(fp, 2);	/* since we are not rounding */
457 		exp = EXT_EXP_INFNAN;
458 		goto done;
459 	}
460 	if (ISINF(fp)) {
461 		sign |= EXT_EXP(EXT_EXP_INFNAN);
462 		goto zero;
463 	}
464 	if (ISZERO(fp)) {
465 zero:		res[1] = res[2] = res[3] = 0;
466 		return (sign);
467 	}
468 
469 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
470 		(void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
471 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
472 			res[1] = res[2] = res[3] = 0;
473 			return (sign | EXT_EXP(1) | 0);
474 		}
475 		if ((fe->fe_cx & FSR_NX) ||
476 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
477 			fe->fe_cx |= FSR_UF;
478 		exp = 0;
479 		goto done;
480 	}
481 	/* Since internal == extended, no need to shift here. */
482 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
483 		exp++;
484 	if (exp >= EXT_EXP_INFNAN) {
485 		fe->fe_cx |= FSR_OF | FSR_NX;
486 		if (toinf(fe, sign)) {
487 			res[1] = res[2] = res[3] = 0;
488 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
489 		}
490 		res[1] = res[2] = res[3] = ~0;
491 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
492 	}
493 done:
494 	res[1] = fp->fp_mant[1];
495 	res[2] = fp->fp_mant[2];
496 	res[3] = fp->fp_mant[3];
497 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
498 }
499 
500 /*
501  * Implode an fpn, writing the result into the given space.
502  */
503 void
504 fpu_implode(fe, fp, type, space)
505 	struct fpemu *fe;
506 	register struct fpn *fp;
507 	int type;
508 	register u_int *space;
509 {
510 
511 	DPRINTF(FPE_REG, ("\n imploding: "));
512 	DUMPFPN(FPE_REG, fp);
513 	DPRINTF(FPE_REG, ("\n"));
514 
515 	switch (type) {
516 
517 #ifdef SUN4U
518 	case FTYPE_LNG:
519 		space[0] = fpu_ftox(fe, fp, space);
520 		break;
521 #endif /* SUN4U */
522 
523 	case FTYPE_INT:
524 		space[0] = fpu_ftoi(fe, fp);
525 		break;
526 
527 	case FTYPE_SNG:
528 		space[0] = fpu_ftos(fe, fp);
529 		break;
530 
531 	case FTYPE_DBL:
532 		space[0] = fpu_ftod(fe, fp, space);
533 		break;
534 
535 	case FTYPE_EXT:
536 		/* funky rounding precision options ?? */
537 		space[0] = fpu_ftoq(fe, fp, space);
538 		break;
539 
540 	default:
541 		panic("fpu_implode");
542 	}
543 #ifdef SUN4U
544 	DPRINTF(FPE_REG, ("fpu_implode: %x %x %x %x\n",
545 		space[0], space[1], space[2], space[3]));
546 #else
547 	DPRINTF(FPE_REG, ("fpu_implode: %x %x\n",
548 		space[0], space[1]));
549 #endif
550 }
551