xref: /openbsd/lib/libc/arch/sparc64/fpu/fpu_implode.c (revision 274d7c50)
1 /*	$OpenBSD: fpu_implode.c,v 1.6 2019/03/15 05:42:38 kevlo 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 /*
49  * FPU subroutines: `implode' internal format numbers into the machine's
50  * `packed binary' format.
51  */
52 
53 #include <sys/param.h>
54 
55 #include <machine/frame.h>
56 #include <machine/fsr.h>
57 #include <machine/ieee.h>
58 #include <machine/instr.h>
59 
60 #include "fpu_arith.h"
61 #include "fpu_emu.h"
62 #include "fpu_extern.h"
63 
64 static int fpround(struct fpemu *, struct fpn *);
65 static int toinf(struct fpemu *, int);
66 
67 #define	FSR_GET_RD(fsr)		(((fsr) >> FSR_RD_SHIFT) & FSR_RD_MASK)
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 fpround(struct fpemu *fe, struct fpn *fp)
83 {
84 	u_int m0, m1, m2, m3;
85 	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 (FSR_GET_RD(fe->fe_fsr)) {
107 	case FSR_RD_RN:
108 	default:
109 		/*
110 		 * Round only if guard is set (gr & 2).  If guard is set,
111 		 * but round & sticky both clear, then we want to round
112 		 * but have a tie, so round to even, i.e., add 1 iff odd.
113 		 */
114 		if ((gr & 2) == 0)
115 			goto rounddown;
116 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
117 			break;
118 		goto rounddown;
119 
120 	case FSR_RD_RZ:
121 		/* Round towards zero, i.e., down. */
122 		goto rounddown;
123 
124 	case FSR_RD_RM:
125 		/* Round towards -Inf: up if negative, down if positive. */
126 		if (fp->fp_sign)
127 			break;
128 		goto rounddown;
129 
130 	case FSR_RD_RP:
131 		/* Round towards +Inf: up if positive, down otherwise. */
132 		if (!fp->fp_sign)
133 			break;
134 		goto rounddown;
135 	}
136 
137 	/* Bump low bit of mantissa, with carry. */
138 	FPU_ADDS(m3, m3, 1);
139 	FPU_ADDCS(m2, m2, 0);
140 	FPU_ADDCS(m1, m1, 0);
141 	FPU_ADDC(m0, m0, 0);
142 	fp->fp_mant[0] = m0;
143 	fp->fp_mant[1] = m1;
144 	fp->fp_mant[2] = m2;
145 	fp->fp_mant[3] = m3;
146 	return (1);
147 
148 rounddown:
149 	fp->fp_mant[0] = m0;
150 	fp->fp_mant[1] = m1;
151 	fp->fp_mant[2] = m2;
152 	fp->fp_mant[3] = m3;
153 	return (0);
154 }
155 
156 /*
157  * For overflow: return true if overflow is to go to +/-Inf, according
158  * to the sign of the overflowing result.  If false, overflow is to go
159  * to the largest magnitude value instead.
160  */
161 static int
162 toinf(struct fpemu *fe, int sign)
163 {
164 	int inf;
165 
166 	/* look at rounding direction */
167 	switch (FSR_GET_RD(fe->fe_fsr)) {
168 	default:
169 	case FSR_RD_RN:		/* the nearest value is always Inf */
170 		inf = 1;
171 		break;
172 
173 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
174 		inf = 0;
175 		break;
176 
177 	case FSR_RD_RP:	/* toward +Inf iff positive */
178 		inf = sign == 0;
179 		break;
180 
181 	case FSR_RD_RM:	/* toward -Inf iff negative */
182 		inf = sign;
183 		break;
184 	}
185 	return (inf);
186 }
187 
188 /*
189  * fpn -> int (int value returned as return value).
190  *
191  * N.B.: this conversion always rounds towards zero (this is a peculiarity
192  * of the SPARC instruction set).
193  */
194 u_int
195 __fpu_ftoi(fe, fp)
196 	struct fpemu *fe;
197 	struct fpn *fp;
198 {
199 	u_int i;
200 	int sign, exp;
201 
202 	sign = fp->fp_sign;
203 	switch (fp->fp_class) {
204 
205 	case FPC_ZERO:
206 		return (0);
207 
208 	case FPC_NUM:
209 		/*
210 		 * If exp >= 2^32, overflow.  Otherwise shift value right
211 		 * into last mantissa word (this will not exceed 0xffffffff),
212 		 * shifting any guard and round bits out into the sticky
213 		 * bit.  Then ``round'' towards zero, i.e., just set an
214 		 * inexact exception if sticky is set (see fpround()).
215 		 * If the result is > 0x80000000, or is positive and equals
216 		 * 0x80000000, overflow; otherwise the last fraction word
217 		 * is the result.
218 		 */
219 		if ((exp = fp->fp_exp) >= 32)
220 			break;
221 		/* NB: the following includes exp < 0 cases */
222 		if (__fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
223 			fe->fe_cx |= FSR_NX;
224 		i = fp->fp_mant[3];
225 		if (i >= ((u_int)0x80000000 + sign))
226 			break;
227 		return (sign ? -i : i);
228 
229 	default:		/* Inf, qNaN, sNaN */
230 		break;
231 	}
232 	/* overflow: replace any inexact exception with invalid */
233 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
234 	return (0x7fffffff + sign);
235 }
236 
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 	struct fpn *fp;
247 	u_int *res;
248 {
249 	u_int64_t i;
250 	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
263 		 * 0xffffffffffffffff), shifting any guard and round bits out
264 		 * into the sticky bit.  Then ``round'' towards zero, i.e.,
265 		 * just set an inexact exception if sticky is set (see
266 		 * fpround()).  If the result is > 0x8000000000000000, or is
267 		 * positive and equals 0x8000000000000000, overflow;
268 		 * otherwise the last fraction word 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 		if (sign)
279 			i = -i;
280 		res[1] = (int)i;
281 		return (i >> 32);
282 
283 	default:		/* Inf, qNaN, sNaN */
284 		break;
285 	}
286 	/* overflow: replace any inexact exception with invalid */
287 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
288 	return (0x7fffffffffffffffLL + sign);
289 }
290 
291 /*
292  * fpn -> single (32 bit single returned as return value).
293  * We assume <= 29 bits in a single-precision fraction (1.f part).
294  */
295 u_int
296 __fpu_ftos(fe, fp)
297 	struct fpemu *fe;
298 	struct fpn *fp;
299 {
300 	u_int sign = fp->fp_sign << 31;
301 	int exp;
302 
303 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
304 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
305 
306 	/* Take care of non-numbers first. */
307 	if (ISNAN(fp)) {
308 		/*
309 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
310 		 * Note that fp->fp_mant[0] has the quiet bit set,
311 		 * even if it is classified as a signalling NaN.
312 		 */
313 		(void) __fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
314 		exp = SNG_EXP_INFNAN;
315 		goto done;
316 	}
317 	if (ISINF(fp))
318 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
319 	if (ISZERO(fp))
320 		return (sign);
321 
322 	/*
323 	 * Normals (including subnormals).  Drop all the fraction bits
324 	 * (including the explicit ``implied'' 1 bit) down into the
325 	 * single-precision range.  If the number is subnormal, move
326 	 * the ``implied'' 1 into the explicit range as well, and shift
327 	 * right to introduce leading zeroes.  Rounding then acts
328 	 * differently for normals and subnormals: the largest subnormal
329 	 * may round to the smallest normal (1.0 x 2^minexp), or may
330 	 * remain subnormal.  In the latter case, signal an underflow
331 	 * if the result was inexact or if underflow traps are enabled.
332 	 *
333 	 * Rounding a normal, on the other hand, always produces another
334 	 * normal (although either way the result might be too big for
335 	 * single precision, and cause an overflow).  If rounding a
336 	 * normal produces 2.0 in the fraction, we need not adjust that
337 	 * fraction at all, since both 1.0 and 2.0 are zero under the
338 	 * fraction mask.
339 	 *
340 	 * Note that the guard and round bits vanish from the number after
341 	 * rounding.
342 	 */
343 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
344 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
345 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
346 		if (fpround(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
347 			return (sign | SNG_EXP(1) | 0);
348 		if ((fe->fe_cx & FSR_NX) ||
349 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
350 			fe->fe_cx |= FSR_UF;
351 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
352 	}
353 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
354 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
355 #ifdef DIAGNOSTIC
356 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
357 		__utrap_panic("fpu_ftos");
358 #endif
359 	if (fpround(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
360 		exp++;
361 	if (exp >= SNG_EXP_INFNAN) {
362 		/* overflow to inf or to max single */
363 		fe->fe_cx |= FSR_OF | FSR_NX;
364 		if (toinf(fe, sign))
365 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
366 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
367 	}
368 done:
369 	/* phew, made it */
370 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
371 }
372 
373 /*
374  * fpn -> double (32 bit high-order result returned; 32-bit low order result
375  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
376  *
377  * This code mimics fpu_ftos; see it for comments.
378  */
379 u_int
380 __fpu_ftod(fe, fp, res)
381 	struct fpemu *fe;
382 	struct fpn *fp;
383 	u_int *res;
384 {
385 	u_int sign = fp->fp_sign << 31;
386 	int exp;
387 
388 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
389 #define	DBL_MASK	(DBL_EXP(1) - 1)
390 
391 	if (ISNAN(fp)) {
392 		(void) __fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
393 		exp = DBL_EXP_INFNAN;
394 		goto done;
395 	}
396 	if (ISINF(fp)) {
397 		sign |= DBL_EXP(DBL_EXP_INFNAN);
398 		goto zero;
399 	}
400 	if (ISZERO(fp)) {
401 zero:		res[1] = 0;
402 		return (sign);
403 	}
404 
405 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
406 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
407 		if (fpround(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
408 			res[1] = 0;
409 			return (sign | DBL_EXP(1) | 0);
410 		}
411 		if ((fe->fe_cx & FSR_NX) ||
412 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
413 			fe->fe_cx |= FSR_UF;
414 		exp = 0;
415 		goto done;
416 	}
417 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
418 	if (fpround(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
419 		exp++;
420 	if (exp >= DBL_EXP_INFNAN) {
421 		fe->fe_cx |= FSR_OF | FSR_NX;
422 		if (toinf(fe, sign)) {
423 			res[1] = 0;
424 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
425 		}
426 		res[1] = ~0;
427 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
428 	}
429 done:
430 	res[1] = fp->fp_mant[3];
431 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
432 }
433 
434 /*
435  * fpn -> extended (32 bit high-order result returned; low-order fraction
436  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
437  * our internal format *is* extended precision, plus 2 bits for guard/round,
438  * so we can avoid a small bit of work.
439  */
440 u_int
441 __fpu_ftoq(fe, fp, res)
442 	struct fpemu *fe;
443 	struct fpn *fp;
444 	u_int *res;
445 {
446 	u_int sign = fp->fp_sign << 31;
447 	int exp;
448 
449 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
450 #define	EXT_MASK	(EXT_EXP(1) - 1)
451 
452 	if (ISNAN(fp)) {
453 		(void) __fpu_shr(fp, 2);	/* since we are not rounding */
454 		exp = EXT_EXP_INFNAN;
455 		goto done;
456 	}
457 	if (ISINF(fp)) {
458 		sign |= EXT_EXP(EXT_EXP_INFNAN);
459 		goto zero;
460 	}
461 	if (ISZERO(fp)) {
462 zero:		res[1] = res[2] = res[3] = 0;
463 		return (sign);
464 	}
465 
466 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
467 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
468 		if (fpround(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
469 			res[1] = res[2] = res[3] = 0;
470 			return (sign | EXT_EXP(1) | 0);
471 		}
472 		if ((fe->fe_cx & FSR_NX) ||
473 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
474 			fe->fe_cx |= FSR_UF;
475 		exp = 0;
476 		goto done;
477 	}
478 	/* Since internal == extended, no need to shift here. */
479 	if (fpround(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
480 		exp++;
481 	if (exp >= EXT_EXP_INFNAN) {
482 		fe->fe_cx |= FSR_OF | FSR_NX;
483 		if (toinf(fe, sign)) {
484 			res[1] = res[2] = res[3] = 0;
485 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
486 		}
487 		res[1] = res[2] = res[3] = ~0;
488 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
489 	}
490 done:
491 	res[1] = fp->fp_mant[1];
492 	res[2] = fp->fp_mant[2];
493 	res[3] = fp->fp_mant[3];
494 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
495 }
496 
497 /*
498  * Implode an fpn, writing the result into the given space.
499  */
500 void
501 __fpu_implode(fe, fp, type, space)
502 	struct fpemu *fe;
503 	struct fpn *fp;
504 	int type;
505 	u_int *space;
506 {
507 
508 	switch (type) {
509 
510 	case FTYPE_LNG:
511 		space[0] = __fpu_ftox(fe, fp, space);
512 		break;
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 #ifdef DIAGNOSTIC
532 	default:
533 		__utrap_panic("fpu_implode");
534 #endif
535 	}
536 	DPRINTF(FPE_REG, ("fpu_implode: %x %x %x %x\n",
537 		space[0], space[1], space[2], space[3]));
538 }
539