xref: /netbsd-src/sys/arch/powerpc/fpu/fpu_implode.c (revision 8e6ab8837d8d6b9198e67c1c445300b483e2f304)
1 /*	$NetBSD: fpu_implode.c,v 1.2 2003/07/15 02:54:43 lukem 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/cdefs.h>
53 __KERNEL_RCSID(0, "$NetBSD: fpu_implode.c,v 1.2 2003/07/15 02:54:43 lukem Exp $");
54 
55 #include <sys/types.h>
56 #include <sys/systm.h>
57 
58 #include <machine/ieee.h>
59 #include <powerpc/instr.h>
60 #include <machine/reg.h>
61 #include <machine/fpu.h>
62 
63 #include <powerpc/fpu/fpu_arith.h>
64 #include <powerpc/fpu/fpu_emu.h>
65 #include <powerpc/fpu/fpu_extern.h>
66 
67 static int round(struct fpemu *, struct fpn *);
68 static int toinf(struct fpemu *, int);
69 
70 /*
71  * Round a number (algorithm from Motorola MC68882 manual, modified for
72  * our internal format).  Set inexact exception if rounding is required.
73  * Return true iff we rounded up.
74  *
75  * After rounding, we discard the guard and round bits by shifting right
76  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
77  * This saves effort later.
78  *
79  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
80  * responsibility to fix this if necessary.
81  */
82 static int
83 round(struct fpemu *fe, struct fpn *fp)
84 {
85 	u_int m0, m1, m2, m3;
86 	int gr, s;
87 	FPU_DECL_CARRY;
88 
89 	m0 = fp->fp_mant[0];
90 	m1 = fp->fp_mant[1];
91 	m2 = fp->fp_mant[2];
92 	m3 = fp->fp_mant[3];
93 	gr = m3 & 3;
94 	s = fp->fp_sticky;
95 
96 	/* mant >>= FP_NG */
97 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
98 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
99 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
100 	m0 >>= FP_NG;
101 
102 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
103 		goto rounddown;
104 
105 	fe->fe_cx |= FPSCR_XX|FPSCR_FI;	/* inexact */
106 
107 	/* Go to rounddown to round down; break to round up. */
108 	switch ((fe->fe_fpscr) & FPSCR_RN) {
109 
110 	case FSR_RD_RN:
111 	default:
112 		/*
113 		 * Round only if guard is set (gr & 2).  If guard is set,
114 		 * but round & sticky both clear, then we want to round
115 		 * but have a tie, so round to even, i.e., add 1 iff odd.
116 		 */
117 		if ((gr & 2) == 0)
118 			goto rounddown;
119 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
120 			break;
121 		goto rounddown;
122 
123 	case FSR_RD_RZ:
124 		/* Round towards zero, i.e., down. */
125 		goto rounddown;
126 
127 	case FSR_RD_RM:
128 		/* Round towards -Inf: up if negative, down if positive. */
129 		if (fp->fp_sign)
130 			break;
131 		goto rounddown;
132 
133 	case FSR_RD_RP:
134 		/* Round towards +Inf: up if positive, down otherwise. */
135 		if (!fp->fp_sign)
136 			break;
137 		goto rounddown;
138 	}
139 
140 	/* Bump low bit of mantissa, with carry. */
141 	fe->fe_cx |= FPSCR_FR;
142 
143 	FPU_ADDS(m3, m3, 1);
144 	FPU_ADDCS(m2, m2, 0);
145 	FPU_ADDCS(m1, m1, 0);
146 	FPU_ADDC(m0, m0, 0);
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 (1);
152 
153 rounddown:
154 	fp->fp_mant[0] = m0;
155 	fp->fp_mant[1] = m1;
156 	fp->fp_mant[2] = m2;
157 	fp->fp_mant[3] = m3;
158 	return (0);
159 }
160 
161 /*
162  * For overflow: return true if overflow is to go to +/-Inf, according
163  * to the sign of the overflowing result.  If false, overflow is to go
164  * to the largest magnitude value instead.
165  */
166 static int
167 toinf(struct fpemu *fe, int sign)
168 {
169 	int inf;
170 
171 	/* look at rounding direction */
172 	switch ((fe->fe_fpscr) & FPSCR_RN) {
173 
174 	default:
175 	case FSR_RD_RN:		/* the nearest value is always Inf */
176 		inf = 1;
177 		break;
178 
179 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
180 		inf = 0;
181 		break;
182 
183 	case FSR_RD_RP:		/* toward +Inf iff positive */
184 		inf = sign == 0;
185 		break;
186 
187 	case FSR_RD_RM:		/* toward -Inf iff negative */
188 		inf = sign;
189 		break;
190 	}
191 	if (inf) fe->fe_cx |= FPSCR_OX;
192 	return (inf);
193 }
194 
195 /*
196  * fpn -> int (int value returned as return value).
197  *
198  * N.B.: this conversion always rounds towards zero (this is a peculiarity
199  * of the SPARC instruction set).
200  */
201 u_int
202 fpu_ftoi(struct fpemu *fe, struct fpn *fp)
203 {
204 	u_int i;
205 	int sign, exp;
206 
207 	sign = fp->fp_sign;
208 	switch (fp->fp_class) {
209 
210 	case FPC_ZERO:
211 		return (0);
212 
213 	case FPC_NUM:
214 		/*
215 		 * If exp >= 2^32, overflow.  Otherwise shift value right
216 		 * into last mantissa word (this will not exceed 0xffffffff),
217 		 * shifting any guard and round bits out into the sticky
218 		 * bit.  Then ``round'' towards zero, i.e., just set an
219 		 * inexact exception if sticky is set (see round()).
220 		 * If the result is > 0x80000000, or is positive and equals
221 		 * 0x80000000, overflow; otherwise the last fraction word
222 		 * is the result.
223 		 */
224 		if ((exp = fp->fp_exp) >= 32)
225 			break;
226 		/* NB: the following includes exp < 0 cases */
227 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
228 			fe->fe_cx |= FPSCR_UX;
229 		i = fp->fp_mant[3];
230 		if (i >= ((u_int)0x80000000 + sign))
231 			break;
232 		return (sign ? -i : i);
233 
234 	default:		/* Inf, qNaN, sNaN */
235 		break;
236 	}
237 	/* overflow: replace any inexact exception with invalid */
238 	fe->fe_cx |= FPSCR_VXCVI;
239 	return (0x7fffffff + sign);
240 }
241 
242 /*
243  * fpn -> extended int (high bits of int value returned as return value).
244  *
245  * N.B.: this conversion always rounds towards zero (this is a peculiarity
246  * of the SPARC instruction set).
247  */
248 u_int
249 fpu_ftox(struct fpemu *fe, struct fpn *fp, u_int *res)
250 {
251 	u_int64_t i;
252 	int sign, exp;
253 
254 	sign = fp->fp_sign;
255 	switch (fp->fp_class) {
256 
257 	case FPC_ZERO:
258 		res[1] = 0;
259 		return (0);
260 
261 	case FPC_NUM:
262 		/*
263 		 * If exp >= 2^64, overflow.  Otherwise shift value right
264 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
265 		 * shifting any guard and round bits out into the sticky
266 		 * bit.  Then ``round'' towards zero, i.e., just set an
267 		 * inexact exception if sticky is set (see round()).
268 		 * If the result is > 0x8000000000000000, or is positive and equals
269 		 * 0x8000000000000000, overflow; otherwise the last fraction word
270 		 * is the result.
271 		 */
272 		if ((exp = fp->fp_exp) >= 64)
273 			break;
274 		/* NB: the following includes exp < 0 cases */
275 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
276 			fe->fe_cx |= FPSCR_UX;
277 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
278 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
279 			break;
280 		return (sign ? -i : i);
281 
282 	default:		/* Inf, qNaN, sNaN */
283 		break;
284 	}
285 	/* overflow: replace any inexact exception with invalid */
286 	fe->fe_cx |= FPSCR_VXCVI;
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
295 fpu_ftos(struct fpemu *fe, struct fpn *fp)
296 {
297 	u_int sign = fp->fp_sign << 31;
298 	int exp;
299 
300 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
301 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
302 
303 	/* Take care of non-numbers first. */
304 	if (ISNAN(fp)) {
305 		/*
306 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
307 		 * Note that fp->fp_mant[0] has the quiet bit set,
308 		 * even if it is classified as a signalling NaN.
309 		 */
310 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
311 		exp = SNG_EXP_INFNAN;
312 		goto done;
313 	}
314 	if (ISINF(fp))
315 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
316 	if (ISZERO(fp))
317 		return (sign);
318 
319 	/*
320 	 * Normals (including subnormals).  Drop all the fraction bits
321 	 * (including the explicit ``implied'' 1 bit) down into the
322 	 * single-precision range.  If the number is subnormal, move
323 	 * the ``implied'' 1 into the explicit range as well, and shift
324 	 * right to introduce leading zeroes.  Rounding then acts
325 	 * differently for normals and subnormals: the largest subnormal
326 	 * may round to the smallest normal (1.0 x 2^minexp), or may
327 	 * remain subnormal.  In the latter case, signal an underflow
328 	 * if the result was inexact or if underflow traps are enabled.
329 	 *
330 	 * Rounding a normal, on the other hand, always produces another
331 	 * normal (although either way the result might be too big for
332 	 * single precision, and cause an overflow).  If rounding a
333 	 * normal produces 2.0 in the fraction, we need not adjust that
334 	 * fraction at all, since both 1.0 and 2.0 are zero under the
335 	 * fraction mask.
336 	 *
337 	 * Note that the guard and round bits vanish from the number after
338 	 * rounding.
339 	 */
340 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
341 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
342 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
343 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
344 			return (sign | SNG_EXP(1) | 0);
345 		if ((fe->fe_cx & FPSCR_FI) ||
346 		    (fe->fe_fpscr & FPSCR_UX))
347 			fe->fe_cx |= FPSCR_UX;
348 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
349 	}
350 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
351 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
352 #ifdef DIAGNOSTIC
353 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
354 		panic("fpu_ftos");
355 #endif
356 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
357 		exp++;
358 	if (exp >= SNG_EXP_INFNAN) {
359 		/* overflow to inf or to max single */
360 		if (toinf(fe, sign))
361 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
362 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
363 	}
364 done:
365 	/* phew, made it */
366 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
367 }
368 
369 /*
370  * fpn -> double (32 bit high-order result returned; 32-bit low order result
371  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
372  *
373  * This code mimics fpu_ftos; see it for comments.
374  */
375 u_int
376 fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *res)
377 {
378 	u_int sign = fp->fp_sign << 31;
379 	int exp;
380 
381 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
382 #define	DBL_MASK	(DBL_EXP(1) - 1)
383 
384 	if (ISNAN(fp)) {
385 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
386 		exp = DBL_EXP_INFNAN;
387 		goto done;
388 	}
389 	if (ISINF(fp)) {
390 		sign |= DBL_EXP(DBL_EXP_INFNAN);
391 		goto zero;
392 	}
393 	if (ISZERO(fp)) {
394 zero:		res[1] = 0;
395 		return (sign);
396 	}
397 
398 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
399 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
400 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
401 			res[1] = 0;
402 			return (sign | DBL_EXP(1) | 0);
403 		}
404 		if ((fe->fe_cx & FPSCR_FI) ||
405 		    (fe->fe_fpscr & FPSCR_UX))
406 			fe->fe_cx |= FPSCR_UX;
407 		exp = 0;
408 		goto done;
409 	}
410 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
411 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
412 		exp++;
413 	if (exp >= DBL_EXP_INFNAN) {
414 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
415 		if (toinf(fe, sign)) {
416 			res[1] = 0;
417 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
418 		}
419 		res[1] = ~0;
420 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
421 	}
422 done:
423 	res[1] = fp->fp_mant[3];
424 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
425 }
426 
427 /*
428  * fpn -> extended (32 bit high-order result returned; low-order fraction
429  * words left in res[1]..res[3]).  Like ftod, which is like ftos ... but
430  * our internal format *is* extended precision, plus 2 bits for guard/round,
431  * so we can avoid a small bit of work.
432  */
433 u_int
434 fpu_ftoq(struct fpemu *fe, struct fpn *fp, u_int *res)
435 {
436 	u_int sign = fp->fp_sign << 31;
437 	int exp;
438 
439 #define	EXT_EXP(e)	((e) << (EXT_FRACBITS & 31))
440 #define	EXT_MASK	(EXT_EXP(1) - 1)
441 
442 	if (ISNAN(fp)) {
443 		(void) fpu_shr(fp, 2);	/* since we are not rounding */
444 		exp = EXT_EXP_INFNAN;
445 		goto done;
446 	}
447 	if (ISINF(fp)) {
448 		sign |= EXT_EXP(EXT_EXP_INFNAN);
449 		goto zero;
450 	}
451 	if (ISZERO(fp)) {
452 zero:		res[1] = res[2] = res[3] = 0;
453 		return (sign);
454 	}
455 
456 	if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) {
457 		(void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
458 		if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) {
459 			res[1] = res[2] = res[3] = 0;
460 			return (sign | EXT_EXP(1) | 0);
461 		}
462 		if ((fe->fe_cx & FPSCR_FI) ||
463 		    (fe->fe_fpscr & FPSCR_UX))
464 			fe->fe_cx |= FPSCR_UX;
465 		exp = 0;
466 		goto done;
467 	}
468 	/* Since internal == extended, no need to shift here. */
469 	if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2))
470 		exp++;
471 	if (exp >= EXT_EXP_INFNAN) {
472 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
473 		if (toinf(fe, sign)) {
474 			res[1] = res[2] = res[3] = 0;
475 			return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
476 		}
477 		res[1] = res[2] = res[3] = ~0;
478 		return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
479 	}
480 done:
481 	res[1] = fp->fp_mant[1];
482 	res[2] = fp->fp_mant[2];
483 	res[3] = fp->fp_mant[3];
484 	return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK));
485 }
486 
487 /*
488  * Implode an fpn, writing the result into the given space.
489  */
490 void
491 fpu_implode(struct fpemu *fe, struct fpn *fp, int type, u_int *space)
492 {
493 
494 	switch (type) {
495 
496 	case FTYPE_LNG:
497 		space[0] = fpu_ftox(fe, fp, space);
498 		DPRINTF(FPE_REG, ("fpu_implode: long %x %x\n",
499 			space[0], space[1]));
500 		break;
501 
502 	case FTYPE_INT:
503 		space[0] = 0;
504 		space[1] = fpu_ftoi(fe, fp);
505 		DPRINTF(FPE_REG, ("fpu_implode: int %x\n",
506 			space[1]));
507 		break;
508 
509 	case FTYPE_SNG:
510 		space[0] = fpu_ftos(fe, fp);
511 		DPRINTF(FPE_REG, ("fpu_implode: single %x\n",
512 			space[0]));
513 		break;
514 
515 	case FTYPE_DBL:
516 		space[0] = fpu_ftod(fe, fp, space);
517 		DPRINTF(FPE_REG, ("fpu_implode: double %x %x\n",
518 			space[0], space[1]));
519 		break;		break;
520 
521 	case FTYPE_EXT:
522 		/* funky rounding precision options ?? */
523 		space[0] = fpu_ftoq(fe, fp, space);
524 		DPRINTF(FPE_REG, ("fpu_implode: long double %x %x %x %x\n",
525 			space[0], space[1], space[2], space[3]));
526 		break;		break;
527 
528 	default:
529 		panic("fpu_implode: invalid type %d", type);
530 	}
531 }
532