xref: /openbsd-src/lib/libc/arch/sparc64/fpu/fpu_implode.c (revision 91f110e064cd7c194e59e019b83bb7496c1c84d4)
1 /*	$OpenBSD: fpu_implode.c,v 1.3 2012/12/05 23:19:59 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  *	$NetBSD: fpu_implode.c,v 1.8 2001/08/26 05:44:46 eeh Exp $
46  */
47 
48 #if 0
49 __FBSDID("$FreeBSD: src/lib/libc/sparc64/fpu/fpu_implode.c,v 1.5 2002/04/27 21:56:28 jake Exp $");
50 #endif
51 
52 /*
53  * FPU subroutines: `implode' internal format numbers into the machine's
54  * `packed binary' format.
55  */
56 
57 #include <sys/param.h>
58 
59 #include <machine/frame.h>
60 #include <machine/fsr.h>
61 #include <machine/ieee.h>
62 #include <machine/instr.h>
63 
64 #include "fpu_arith.h"
65 #include "fpu_emu.h"
66 #include "fpu_extern.h"
67 
68 static int round(struct fpemu *, struct fpn *);
69 static int toinf(struct fpemu *, int);
70 
71 #define	FSR_GET_RD(fsr)		(((fsr) >> FSR_RD_SHIFT) & FSR_RD_MASK)
72 
73 /*
74  * Round a number (algorithm from Motorola MC68882 manual, modified for
75  * our internal format).  Set inexact exception if rounding is required.
76  * Return true iff we rounded up.
77  *
78  * After rounding, we discard the guard and round bits by shifting right
79  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
80  * This saves effort later.
81  *
82  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
83  * responsibility to fix this if necessary.
84  */
85 static int
86 round(struct fpemu *fe, struct fpn *fp)
87 {
88 	u_int m0, m1, m2, m3;
89 	int gr, s;
90 
91 	m0 = fp->fp_mant[0];
92 	m1 = fp->fp_mant[1];
93 	m2 = fp->fp_mant[2];
94 	m3 = fp->fp_mant[3];
95 	gr = m3 & 3;
96 	s = fp->fp_sticky;
97 
98 	/* mant >>= FP_NG */
99 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
100 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
101 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
102 	m0 >>= FP_NG;
103 
104 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
105 		goto rounddown;
106 
107 	fe->fe_cx |= FSR_NX;	/* inexact */
108 
109 	/* Go to rounddown to round down; break to round up. */
110 	switch (FSR_GET_RD(fe->fe_fsr)) {
111 	case FSR_RD_RN:
112 	default:
113 		/*
114 		 * Round only if guard is set (gr & 2).  If guard is set,
115 		 * but round & sticky both clear, then we want to round
116 		 * but have a tie, so round to even, i.e., add 1 iff odd.
117 		 */
118 		if ((gr & 2) == 0)
119 			goto rounddown;
120 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
121 			break;
122 		goto rounddown;
123 
124 	case FSR_RD_RZ:
125 		/* Round towards zero, i.e., down. */
126 		goto rounddown;
127 
128 	case FSR_RD_RM:
129 		/* Round towards -Inf: up if negative, down if positive. */
130 		if (fp->fp_sign)
131 			break;
132 		goto rounddown;
133 
134 	case FSR_RD_RP:
135 		/* Round towards +Inf: up if positive, down otherwise. */
136 		if (!fp->fp_sign)
137 			break;
138 		goto rounddown;
139 	}
140 
141 	/* Bump low bit of mantissa, with carry. */
142 	FPU_ADDS(m3, m3, 1);
143 	FPU_ADDCS(m2, m2, 0);
144 	FPU_ADDCS(m1, m1, 0);
145 	FPU_ADDC(m0, m0, 0);
146 	fp->fp_mant[0] = m0;
147 	fp->fp_mant[1] = m1;
148 	fp->fp_mant[2] = m2;
149 	fp->fp_mant[3] = m3;
150 	return (1);
151 
152 rounddown:
153 	fp->fp_mant[0] = m0;
154 	fp->fp_mant[1] = m1;
155 	fp->fp_mant[2] = m2;
156 	fp->fp_mant[3] = m3;
157 	return (0);
158 }
159 
160 /*
161  * For overflow: return true if overflow is to go to +/-Inf, according
162  * to the sign of the overflowing result.  If false, overflow is to go
163  * to the largest magnitude value instead.
164  */
165 static int
166 toinf(struct fpemu *fe, int sign)
167 {
168 	int inf;
169 
170 	/* look at rounding direction */
171 	switch (FSR_GET_RD(fe->fe_fsr)) {
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 	struct fpn *fp;
202 {
203 	u_int i;
204 	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 -> extended int (high bits of int value returned as return value).
243  *
244  * N.B.: this conversion always rounds towards zero (this is a peculiarity
245  * of the SPARC instruction set).
246  */
247 u_int
248 __fpu_ftox(fe, fp, res)
249 	struct fpemu *fe;
250 	struct fpn *fp;
251 	u_int *res;
252 {
253 	u_int64_t i;
254 	int sign, exp;
255 
256 	sign = fp->fp_sign;
257 	switch (fp->fp_class) {
258 
259 	case FPC_ZERO:
260 		res[1] = 0;
261 		return (0);
262 
263 	case FPC_NUM:
264 		/*
265 		 * If exp >= 2^64, overflow.  Otherwise shift value right
266 		 * into last mantissa word (this will not exceed
267 		 * 0xffffffffffffffff), shifting any guard and round bits out
268 		 * into the sticky bit.  Then ``round'' towards zero, i.e.,
269 		 * just set an inexact exception if sticky is set (see round()).
270 		 * If the result is > 0x8000000000000000, or is positive and
271 		 * equals 0x8000000000000000, overflow; otherwise the
272 		 * last fraction word is the result.
273 		 */
274 		if ((exp = fp->fp_exp) >= 64)
275 			break;
276 		/* NB: the following includes exp < 0 cases */
277 		if (__fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
278 			fe->fe_cx |= FSR_NX;
279 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
280 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
281 			break;
282 		if (sign)
283 			i = -i;
284 		res[1] = (int)i;
285 		return (i >> 32);
286 
287 	default:		/* Inf, qNaN, sNaN */
288 		break;
289 	}
290 	/* overflow: replace any inexact exception with invalid */
291 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
292 	return (0x7fffffffffffffffLL + sign);
293 }
294 
295 /*
296  * fpn -> single (32 bit single returned as return value).
297  * We assume <= 29 bits in a single-precision fraction (1.f part).
298  */
299 u_int
300 __fpu_ftos(fe, fp)
301 	struct fpemu *fe;
302 	struct fpn *fp;
303 {
304 	u_int sign = fp->fp_sign << 31;
305 	int exp;
306 
307 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
308 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
309 
310 	/* Take care of non-numbers first. */
311 	if (ISNAN(fp)) {
312 		/*
313 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
314 		 * Note that fp->fp_mant[0] has the quiet bit set,
315 		 * even if it is classified as a signalling NaN.
316 		 */
317 		(void) __fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
318 		exp = SNG_EXP_INFNAN;
319 		goto done;
320 	}
321 	if (ISINF(fp))
322 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
323 	if (ISZERO(fp))
324 		return (sign);
325 
326 	/*
327 	 * Normals (including subnormals).  Drop all the fraction bits
328 	 * (including the explicit ``implied'' 1 bit) down into the
329 	 * single-precision range.  If the number is subnormal, move
330 	 * the ``implied'' 1 into the explicit range as well, and shift
331 	 * right to introduce leading zeroes.  Rounding then acts
332 	 * differently for normals and subnormals: the largest subnormal
333 	 * may round to the smallest normal (1.0 x 2^minexp), or may
334 	 * remain subnormal.  In the latter case, signal an underflow
335 	 * if the result was inexact or if underflow traps are enabled.
336 	 *
337 	 * Rounding a normal, on the other hand, always produces another
338 	 * normal (although either way the result might be too big for
339 	 * single precision, and cause an overflow).  If rounding a
340 	 * normal produces 2.0 in the fraction, we need not adjust that
341 	 * fraction at all, since both 1.0 and 2.0 are zero under the
342 	 * fraction mask.
343 	 *
344 	 * Note that the guard and round bits vanish from the number after
345 	 * rounding.
346 	 */
347 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
348 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
349 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
350 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
351 			return (sign | SNG_EXP(1) | 0);
352 		if ((fe->fe_cx & FSR_NX) ||
353 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
354 			fe->fe_cx |= FSR_UF;
355 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
356 	}
357 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
358 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
359 #ifdef DIAGNOSTIC
360 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
361 		__utrap_panic("fpu_ftos");
362 #endif
363 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
364 		exp++;
365 	if (exp >= SNG_EXP_INFNAN) {
366 		/* overflow to inf or to max single */
367 		fe->fe_cx |= FSR_OF | FSR_NX;
368 		if (toinf(fe, sign))
369 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
370 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
371 	}
372 done:
373 	/* phew, made it */
374 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
375 }
376 
377 /*
378  * fpn -> double (32 bit high-order result returned; 32-bit low order result
379  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
380  *
381  * This code mimics fpu_ftos; see it for comments.
382  */
383 u_int
384 __fpu_ftod(fe, fp, res)
385 	struct fpemu *fe;
386 	struct fpn *fp;
387 	u_int *res;
388 {
389 	u_int sign = fp->fp_sign << 31;
390 	int exp;
391 
392 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
393 #define	DBL_MASK	(DBL_EXP(1) - 1)
394 
395 	if (ISNAN(fp)) {
396 		(void) __fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
397 		exp = DBL_EXP_INFNAN;
398 		goto done;
399 	}
400 	if (ISINF(fp)) {
401 		sign |= DBL_EXP(DBL_EXP_INFNAN);
402 		goto zero;
403 	}
404 	if (ISZERO(fp)) {
405 zero:		res[1] = 0;
406 		return (sign);
407 	}
408 
409 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
410 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
411 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
412 			res[1] = 0;
413 			return (sign | DBL_EXP(1) | 0);
414 		}
415 		if ((fe->fe_cx & FSR_NX) ||
416 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
417 			fe->fe_cx |= FSR_UF;
418 		exp = 0;
419 		goto done;
420 	}
421 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
422 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
423 		exp++;
424 	if (exp >= DBL_EXP_INFNAN) {
425 		fe->fe_cx |= FSR_OF | FSR_NX;
426 		if (toinf(fe, sign)) {
427 			res[1] = 0;
428 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
429 		}
430 		res[1] = ~0;
431 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
432 	}
433 done:
434 	res[1] = fp->fp_mant[3];
435 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
436 }
437 
438 /*
439  * fpn -> extended (32 bit high-order result returned; low-order fraction
440  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
441  * our internal format *is* extended precision, plus 2 bits for guard/round,
442  * so we can avoid a small bit of work.
443  */
444 u_int
445 __fpu_ftoq(fe, fp, res)
446 	struct fpemu *fe;
447 	struct fpn *fp;
448 	u_int *res;
449 {
450 	u_int sign = fp->fp_sign << 31;
451 	int exp;
452 
453 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
454 #define	EXT_MASK	(EXT_EXP(1) - 1)
455 
456 	if (ISNAN(fp)) {
457 		(void) __fpu_shr(fp, 2);	/* since we are not rounding */
458 		exp = EXT_EXP_INFNAN;
459 		goto done;
460 	}
461 	if (ISINF(fp)) {
462 		sign |= EXT_EXP(EXT_EXP_INFNAN);
463 		goto zero;
464 	}
465 	if (ISZERO(fp)) {
466 zero:		res[1] = res[2] = res[3] = 0;
467 		return (sign);
468 	}
469 
470 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
471 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
472 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
473 			res[1] = res[2] = res[3] = 0;
474 			return (sign | EXT_EXP(1) | 0);
475 		}
476 		if ((fe->fe_cx & FSR_NX) ||
477 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
478 			fe->fe_cx |= FSR_UF;
479 		exp = 0;
480 		goto done;
481 	}
482 	/* Since internal == extended, no need to shift here. */
483 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
484 		exp++;
485 	if (exp >= EXT_EXP_INFNAN) {
486 		fe->fe_cx |= FSR_OF | FSR_NX;
487 		if (toinf(fe, sign)) {
488 			res[1] = res[2] = res[3] = 0;
489 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
490 		}
491 		res[1] = res[2] = res[3] = ~0;
492 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
493 	}
494 done:
495 	res[1] = fp->fp_mant[1];
496 	res[2] = fp->fp_mant[2];
497 	res[3] = fp->fp_mant[3];
498 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
499 }
500 
501 /*
502  * Implode an fpn, writing the result into the given space.
503  */
504 void
505 __fpu_implode(fe, fp, type, space)
506 	struct fpemu *fe;
507 	struct fpn *fp;
508 	int type;
509 	u_int *space;
510 {
511 
512 	switch (type) {
513 
514 	case FTYPE_LNG:
515 		space[0] = __fpu_ftox(fe, fp, space);
516 		break;
517 
518 	case FTYPE_INT:
519 		space[0] = __fpu_ftoi(fe, fp);
520 		break;
521 
522 	case FTYPE_SNG:
523 		space[0] = __fpu_ftos(fe, fp);
524 		break;
525 
526 	case FTYPE_DBL:
527 		space[0] = __fpu_ftod(fe, fp, space);
528 		break;
529 
530 	case FTYPE_EXT:
531 		/* funky rounding precision options ?? */
532 		space[0] = __fpu_ftoq(fe, fp, space);
533 		break;
534 
535 #ifdef DIAGNOSTIC
536 	default:
537 		__utrap_panic("fpu_implode");
538 #endif
539 	}
540 	DPRINTF(FPE_REG, ("fpu_implode: %x %x %x %x\n",
541 		space[0], space[1], space[2], space[3]));
542 }
543