xref: /openbsd-src/lib/libc/arch/sparc64/fpu/fpu_implode.c (revision 6244ddcc7bcc151947d91dad33fcee9e7428fcc3)
1 /*	$OpenBSD: fpu_implode.c,v 1.8 2024/03/29 21:02:11 miod 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/types.h>
54 
55 #include <machine/fsr.h>
56 #include <machine/ieee.h>
57 #include <machine/instr.h>
58 
59 #include "fpu_arith.h"
60 #include "fpu_emu.h"
61 #include "fpu_extern.h"
62 
63 static int fpround(struct fpemu *, struct fpn *);
64 static int toinf(struct fpemu *, int);
65 
66 #define	FSR_GET_RD(fsr)		(((fsr) >> FSR_RD_SHIFT) & FSR_RD_MASK)
67 
68 /*
69  * Round a number (algorithm from Motorola MC68882 manual, modified for
70  * our internal format).  Set inexact exception if rounding is required.
71  * Return true iff we rounded up.
72  *
73  * After rounding, we discard the guard and round bits by shifting right
74  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
75  * This saves effort later.
76  *
77  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
78  * responsibility to fix this if necessary.
79  */
80 static int
fpround(struct fpemu * fe,struct fpn * fp)81 fpround(struct fpemu *fe, struct fpn *fp)
82 {
83 	u_int m0, m1, m2, m3;
84 	int gr, s;
85 
86 	m0 = fp->fp_mant[0];
87 	m1 = fp->fp_mant[1];
88 	m2 = fp->fp_mant[2];
89 	m3 = fp->fp_mant[3];
90 	gr = m3 & 3;
91 	s = fp->fp_sticky;
92 
93 	/* mant >>= FP_NG */
94 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
95 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
96 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
97 	m0 >>= FP_NG;
98 
99 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
100 		goto rounddown;
101 
102 	fe->fe_cx |= FSR_NX;	/* inexact */
103 
104 	/* Go to rounddown to round down; break to round up. */
105 	switch (FSR_GET_RD(fe->fe_fsr)) {
106 	case FSR_RD_RN:
107 	default:
108 		/*
109 		 * Round only if guard is set (gr & 2).  If guard is set,
110 		 * but round & sticky both clear, then we want to round
111 		 * but have a tie, so round to even, i.e., add 1 iff odd.
112 		 */
113 		if ((gr & 2) == 0)
114 			goto rounddown;
115 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
116 			break;
117 		goto rounddown;
118 
119 	case FSR_RD_RZ:
120 		/* Round towards zero, i.e., down. */
121 		goto rounddown;
122 
123 	case FSR_RD_RM:
124 		/* Round towards -Inf: up if negative, down if positive. */
125 		if (fp->fp_sign)
126 			break;
127 		goto rounddown;
128 
129 	case FSR_RD_RP:
130 		/* Round towards +Inf: up if positive, down otherwise. */
131 		if (!fp->fp_sign)
132 			break;
133 		goto rounddown;
134 	}
135 
136 	/* Bump low bit of mantissa, with carry. */
137 	FPU_ADDS(m3, m3, 1);
138 	FPU_ADDCS(m2, m2, 0);
139 	FPU_ADDCS(m1, m1, 0);
140 	FPU_ADDC(m0, m0, 0);
141 	fp->fp_mant[0] = m0;
142 	fp->fp_mant[1] = m1;
143 	fp->fp_mant[2] = m2;
144 	fp->fp_mant[3] = m3;
145 	return (1);
146 
147 rounddown:
148 	fp->fp_mant[0] = m0;
149 	fp->fp_mant[1] = m1;
150 	fp->fp_mant[2] = m2;
151 	fp->fp_mant[3] = m3;
152 	return (0);
153 }
154 
155 /*
156  * For overflow: return true if overflow is to go to +/-Inf, according
157  * to the sign of the overflowing result.  If false, overflow is to go
158  * to the largest magnitude value instead.
159  */
160 static int
toinf(struct fpemu * fe,int sign)161 toinf(struct fpemu *fe, int sign)
162 {
163 	int inf;
164 
165 	/* look at rounding direction */
166 	switch (FSR_GET_RD(fe->fe_fsr)) {
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
__fpu_ftoi(fe,fp)194 __fpu_ftoi(fe, fp)
195 	struct fpemu *fe;
196 	struct fpn *fp;
197 {
198 	u_int i;
199 	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 fpround()).
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 -> extended int (high bits of int value returned as return value).
238  *
239  * N.B.: this conversion always rounds towards zero (this is a peculiarity
240  * of the SPARC instruction set).
241  */
242 u_int
__fpu_ftox(fe,fp,res)243 __fpu_ftox(fe, fp, res)
244 	struct fpemu *fe;
245 	struct fpn *fp;
246 	u_int *res;
247 {
248 	u_int64_t i;
249 	int sign, exp;
250 
251 	sign = fp->fp_sign;
252 	switch (fp->fp_class) {
253 
254 	case FPC_ZERO:
255 		res[1] = 0;
256 		return (0);
257 
258 	case FPC_NUM:
259 		/*
260 		 * If exp >= 2^64, overflow.  Otherwise shift value right
261 		 * into last mantissa word (this will not exceed
262 		 * 0xffffffffffffffff), shifting any guard and round bits out
263 		 * into the sticky bit.  Then ``round'' towards zero, i.e.,
264 		 * just set an inexact exception if sticky is set (see
265 		 * fpround()).  If the result is > 0x8000000000000000, or is
266 		 * positive and equals 0x8000000000000000, overflow;
267 		 * otherwise the last fraction word is the result.
268 		 */
269 		if ((exp = fp->fp_exp) >= 64)
270 			break;
271 		/* NB: the following includes exp < 0 cases */
272 		if (__fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
273 			fe->fe_cx |= FSR_NX;
274 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
275 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
276 			break;
277 		if (sign)
278 			i = -i;
279 		res[1] = (int)i;
280 		return (i >> 32);
281 
282 	default:		/* Inf, qNaN, sNaN */
283 		break;
284 	}
285 	/* overflow: replace any inexact exception with invalid */
286 	fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV;
287 	return (0x7fffffffffffffffLL + sign);
288 }
289 
290 /*
291  * fpn -> single (32 bit single returned as return value).
292  * We assume <= 29 bits in a single-precision fraction (1.f part).
293  */
294 u_int
__fpu_ftos(fe,fp)295 __fpu_ftos(fe, fp)
296 	struct fpemu *fe;
297 	struct fpn *fp;
298 {
299 	u_int sign = fp->fp_sign << 31;
300 	int exp;
301 
302 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
303 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
304 
305 	/* Take care of non-numbers first. */
306 	if (ISNAN(fp)) {
307 		/*
308 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
309 		 * Note that fp->fp_mant[0] has the quiet bit set,
310 		 * even if it is classified as a signalling NaN.
311 		 */
312 		(void) __fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
313 		exp = SNG_EXP_INFNAN;
314 		goto done;
315 	}
316 	if (ISINF(fp))
317 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
318 	if (ISZERO(fp))
319 		return (sign);
320 
321 	/*
322 	 * Normals (including subnormals).  Drop all the fraction bits
323 	 * (including the explicit ``implied'' 1 bit) down into the
324 	 * single-precision range.  If the number is subnormal, move
325 	 * the ``implied'' 1 into the explicit range as well, and shift
326 	 * right to introduce leading zeroes.  Rounding then acts
327 	 * differently for normals and subnormals: the largest subnormal
328 	 * may round to the smallest normal (1.0 x 2^minexp), or may
329 	 * remain subnormal.  In the latter case, signal an underflow
330 	 * if the result was inexact or if underflow traps are enabled.
331 	 *
332 	 * Rounding a normal, on the other hand, always produces another
333 	 * normal (although either way the result might be too big for
334 	 * single precision, and cause an overflow).  If rounding a
335 	 * normal produces 2.0 in the fraction, we need not adjust that
336 	 * fraction at all, since both 1.0 and 2.0 are zero under the
337 	 * fraction mask.
338 	 *
339 	 * Note that the guard and round bits vanish from the number after
340 	 * rounding.
341 	 */
342 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
343 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
344 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
345 		if (fpround(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
346 			return (sign | SNG_EXP(1) | 0);
347 		if ((fe->fe_cx & FSR_NX) ||
348 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
349 			fe->fe_cx |= FSR_UF;
350 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
351 	}
352 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
353 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
354 #ifdef DIAGNOSTIC
355 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
356 		__utrap_panic("fpu_ftos");
357 #endif
358 	if (fpround(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
359 		exp++;
360 	if (exp >= SNG_EXP_INFNAN) {
361 		/* overflow to inf or to max single */
362 		fe->fe_cx |= FSR_OF | FSR_NX;
363 		if (toinf(fe, sign))
364 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
365 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
366 	}
367 done:
368 	/* phew, made it */
369 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
370 }
371 
372 /*
373  * fpn -> double (32 bit high-order result returned; 32-bit low order result
374  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
375  *
376  * This code mimics fpu_ftos; see it for comments.
377  */
378 u_int
__fpu_ftod(fe,fp,res)379 __fpu_ftod(fe, fp, res)
380 	struct fpemu *fe;
381 	struct fpn *fp;
382 	u_int *res;
383 {
384 	u_int sign = fp->fp_sign << 31;
385 	int exp;
386 
387 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
388 #define	DBL_MASK	(DBL_EXP(1) - 1)
389 
390 	if (ISNAN(fp)) {
391 		(void) __fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
392 		exp = DBL_EXP_INFNAN;
393 		goto done;
394 	}
395 	if (ISINF(fp)) {
396 		sign |= DBL_EXP(DBL_EXP_INFNAN);
397 		goto zero;
398 	}
399 	if (ISZERO(fp)) {
400 zero:		res[1] = 0;
401 		return (sign);
402 	}
403 
404 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
405 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
406 		if (fpround(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
407 			res[1] = 0;
408 			return (sign | DBL_EXP(1) | 0);
409 		}
410 		if ((fe->fe_cx & FSR_NX) ||
411 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
412 			fe->fe_cx |= FSR_UF;
413 		exp = 0;
414 		goto done;
415 	}
416 	(void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
417 	if (fpround(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
418 		exp++;
419 	if (exp >= DBL_EXP_INFNAN) {
420 		fe->fe_cx |= FSR_OF | FSR_NX;
421 		if (toinf(fe, sign)) {
422 			res[1] = 0;
423 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
424 		}
425 		res[1] = ~0;
426 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
427 	}
428 done:
429 	res[1] = fp->fp_mant[3];
430 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
431 }
432 
433 /*
434  * fpn -> extended (32 bit high-order result returned; low-order fraction
435  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
436  * our internal format *is* extended precision, plus 2 bits for guard/round,
437  * so we can avoid a small bit of work.
438  */
439 u_int
__fpu_ftoq(fe,fp,res)440 __fpu_ftoq(fe, fp, res)
441 	struct fpemu *fe;
442 	struct fpn *fp;
443 	u_int *res;
444 {
445 	u_int sign = fp->fp_sign << 31;
446 	int exp;
447 
448 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
449 #define	EXT_MASK	(EXT_EXP(1) - 1)
450 
451 	if (ISNAN(fp)) {
452 		(void) __fpu_shr(fp, 2);	/* since we are not rounding */
453 		exp = EXT_EXP_INFNAN;
454 		goto done;
455 	}
456 	if (ISINF(fp)) {
457 		sign |= EXT_EXP(EXT_EXP_INFNAN);
458 		goto zero;
459 	}
460 	if (ISZERO(fp)) {
461 zero:		res[1] = res[2] = res[3] = 0;
462 		return (sign);
463 	}
464 
465 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
466 		(void) __fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
467 		if (fpround(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
468 			res[1] = res[2] = res[3] = 0;
469 			return (sign | EXT_EXP(1) | 0);
470 		}
471 		if ((fe->fe_cx & FSR_NX) ||
472 		    (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT)))
473 			fe->fe_cx |= FSR_UF;
474 		exp = 0;
475 		goto done;
476 	}
477 	/* Since internal == extended, no need to shift here. */
478 	if (fpround(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
479 		exp++;
480 	if (exp >= EXT_EXP_INFNAN) {
481 		fe->fe_cx |= FSR_OF | FSR_NX;
482 		if (toinf(fe, sign)) {
483 			res[1] = res[2] = res[3] = 0;
484 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
485 		}
486 		res[1] = res[2] = res[3] = ~0;
487 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
488 	}
489 done:
490 	res[1] = fp->fp_mant[1];
491 	res[2] = fp->fp_mant[2];
492 	res[3] = fp->fp_mant[3];
493 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
494 }
495 
496 /*
497  * Implode an fpn, writing the result into the given space.
498  */
499 void
__fpu_implode(fe,fp,type,space)500 __fpu_implode(fe, fp, type, space)
501 	struct fpemu *fe;
502 	struct fpn *fp;
503 	int type;
504 	u_int *space;
505 {
506 
507 	switch (type) {
508 
509 	case FTYPE_LNG:
510 		space[0] = __fpu_ftox(fe, fp, space);
511 		break;
512 
513 	case FTYPE_INT:
514 		space[0] = __fpu_ftoi(fe, fp);
515 		break;
516 
517 	case FTYPE_SNG:
518 		space[0] = __fpu_ftos(fe, fp);
519 		break;
520 
521 	case FTYPE_DBL:
522 		space[0] = __fpu_ftod(fe, fp, space);
523 		break;
524 
525 	case FTYPE_EXT:
526 		/* funky rounding precision options ?? */
527 		space[0] = __fpu_ftoq(fe, fp, space);
528 		break;
529 
530 #ifdef DIAGNOSTIC
531 	default:
532 		__utrap_panic("fpu_implode");
533 #endif
534 	}
535 	DPRINTF(FPE_REG, ("fpu_implode: %x %x %x %x\n",
536 		space[0], space[1], space[2], space[3]));
537 }
538