xref: /freebsd/sys/powerpc/fpu/fpu_implode.c (revision 6ae1554a5d9b318f8ad53ccc39fa5a961403da73)
1 /*	$NetBSD: fpu_implode.c,v 1.6 2005/12/11 12:18:42 christos 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 __FBSDID("$FreeBSD$");
50 
51 #include <sys/types.h>
52 #include <sys/systm.h>
53 
54 #include <machine/fpu.h>
55 #include <machine/ieee.h>
56 #include <machine/ieeefp.h>
57 #include <machine/reg.h>
58 
59 #include <powerpc/fpu/fpu_arith.h>
60 #include <powerpc/fpu/fpu_emu.h>
61 #include <powerpc/fpu/fpu_extern.h>
62 #include <powerpc/fpu/fpu_instr.h>
63 
64 static int round(struct fpemu *, struct fpn *);
65 static int toinf(struct fpemu *, int);
66 
67 /*
68  * Round a number (algorithm from Motorola MC68882 manual, modified for
69  * our internal format).  Set inexact exception if rounding is required.
70  * Return true iff we rounded up.
71  *
72  * After rounding, we discard the guard and round bits by shifting right
73  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
74  * This saves effort later.
75  *
76  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
77  * responsibility to fix this if necessary.
78  */
79 static int
80 round(struct fpemu *fe, struct fpn *fp)
81 {
82 	u_int m0, m1, m2, m3;
83 	int gr, s;
84 	FPU_DECL_CARRY;
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 |= FPSCR_XX|FPSCR_FI;	/* inexact */
103 
104 	/* Go to rounddown to round down; break to round up. */
105 	switch ((fe->fe_fpscr) & FPSCR_RN) {
106 
107 	case FP_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 FP_RZ:
121 		/* Round towards zero, i.e., down. */
122 		goto rounddown;
123 
124 	case FP_RM:
125 		/* Round towards -Inf: up if negative, down if positive. */
126 		if (fp->fp_sign)
127 			break;
128 		goto rounddown;
129 
130 	case FP_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 	fe->fe_cx |= FPSCR_FR;
139 
140 	FPU_ADDS(m3, m3, 1);
141 	FPU_ADDCS(m2, m2, 0);
142 	FPU_ADDCS(m1, m1, 0);
143 	FPU_ADDC(m0, m0, 0);
144 	fp->fp_mant[0] = m0;
145 	fp->fp_mant[1] = m1;
146 	fp->fp_mant[2] = m2;
147 	fp->fp_mant[3] = m3;
148 	return (1);
149 
150 rounddown:
151 	fp->fp_mant[0] = m0;
152 	fp->fp_mant[1] = m1;
153 	fp->fp_mant[2] = m2;
154 	fp->fp_mant[3] = m3;
155 	return (0);
156 }
157 
158 /*
159  * For overflow: return true if overflow is to go to +/-Inf, according
160  * to the sign of the overflowing result.  If false, overflow is to go
161  * to the largest magnitude value instead.
162  */
163 static int
164 toinf(struct fpemu *fe, int sign)
165 {
166 	int inf;
167 
168 	/* look at rounding direction */
169 	switch ((fe->fe_fpscr) & FPSCR_RN) {
170 
171 	default:
172 	case FP_RN:		/* the nearest value is always Inf */
173 		inf = 1;
174 		break;
175 
176 	case FP_RZ:		/* toward 0 => never towards Inf */
177 		inf = 0;
178 		break;
179 
180 	case FP_RP:		/* toward +Inf iff positive */
181 		inf = sign == 0;
182 		break;
183 
184 	case FP_RM:		/* toward -Inf iff negative */
185 		inf = sign;
186 		break;
187 	}
188 	if (inf)
189 		fe->fe_cx |= FPSCR_OX;
190 	return (inf);
191 }
192 
193 /*
194  * fpn -> int (int value returned as return value).
195  *
196  * N.B.: this conversion always rounds towards zero (this is a peculiarity
197  * of the SPARC instruction set).
198  */
199 u_int
200 fpu_ftoi(struct fpemu *fe, struct fpn *fp)
201 {
202 	u_int i;
203 	int sign, exp;
204 
205 	sign = fp->fp_sign;
206 	switch (fp->fp_class) {
207 
208 	case FPC_ZERO:
209 		return (0);
210 
211 	case FPC_NUM:
212 		/*
213 		 * If exp >= 2^32, overflow.  Otherwise shift value right
214 		 * into last mantissa word (this will not exceed 0xffffffff),
215 		 * shifting any guard and round bits out into the sticky
216 		 * bit.  Then ``round'' towards zero, i.e., just set an
217 		 * inexact exception if sticky is set (see round()).
218 		 * If the result is > 0x80000000, or is positive and equals
219 		 * 0x80000000, overflow; otherwise the last fraction word
220 		 * is the result.
221 		 */
222 		if ((exp = fp->fp_exp) >= 32)
223 			break;
224 		/* NB: the following includes exp < 0 cases */
225 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
226 			fe->fe_cx |= FPSCR_UX;
227 		i = fp->fp_mant[3];
228 		if (i >= ((u_int)0x80000000 + sign))
229 			break;
230 		return (sign ? -i : i);
231 
232 	default:		/* Inf, qNaN, sNaN */
233 		break;
234 	}
235 	/* overflow: replace any inexact exception with invalid */
236 	fe->fe_cx |= FPSCR_VXCVI;
237 	return (0x7fffffff + sign);
238 }
239 
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(struct fpemu *fe, struct fpn *fp, 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 0xffffffffffffffff),
263 		 * shifting any guard and round bits out into the sticky
264 		 * bit.  Then ``round'' towards zero, i.e., just set an
265 		 * inexact exception if sticky is set (see round()).
266 		 * If the result is > 0x8000000000000000, or is positive and equals
267 		 * 0x8000000000000000, overflow; otherwise the last fraction word
268 		 * 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 |= FPSCR_UX;
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 		return (sign ? -i : i);
279 
280 	default:		/* Inf, qNaN, sNaN */
281 		break;
282 	}
283 	/* overflow: replace any inexact exception with invalid */
284 	fe->fe_cx |= FPSCR_VXCVI;
285 	return (0x7fffffffffffffffLL + sign);
286 }
287 
288 /*
289  * fpn -> single (32 bit single returned as return value).
290  * We assume <= 29 bits in a single-precision fraction (1.f part).
291  */
292 u_int
293 fpu_ftos(struct fpemu *fe, struct fpn *fp)
294 {
295 	u_int sign = fp->fp_sign << 31;
296 	int exp;
297 
298 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
299 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
300 
301 	/* Take care of non-numbers first. */
302 	if (ISNAN(fp)) {
303 		/*
304 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
305 		 * Note that fp->fp_mant[0] has the quiet bit set,
306 		 * even if it is classified as a signalling NaN.
307 		 */
308 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
309 		exp = SNG_EXP_INFNAN;
310 		goto done;
311 	}
312 	if (ISINF(fp))
313 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
314 	if (ISZERO(fp))
315 		return (sign);
316 
317 	/*
318 	 * Normals (including subnormals).  Drop all the fraction bits
319 	 * (including the explicit ``implied'' 1 bit) down into the
320 	 * single-precision range.  If the number is subnormal, move
321 	 * the ``implied'' 1 into the explicit range as well, and shift
322 	 * right to introduce leading zeroes.  Rounding then acts
323 	 * differently for normals and subnormals: the largest subnormal
324 	 * may round to the smallest normal (1.0 x 2^minexp), or may
325 	 * remain subnormal.  In the latter case, signal an underflow
326 	 * if the result was inexact or if underflow traps are enabled.
327 	 *
328 	 * Rounding a normal, on the other hand, always produces another
329 	 * normal (although either way the result might be too big for
330 	 * single precision, and cause an overflow).  If rounding a
331 	 * normal produces 2.0 in the fraction, we need not adjust that
332 	 * fraction at all, since both 1.0 and 2.0 are zero under the
333 	 * fraction mask.
334 	 *
335 	 * Note that the guard and round bits vanish from the number after
336 	 * rounding.
337 	 */
338 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
339 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
340 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
341 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
342 			return (sign | SNG_EXP(1) | 0);
343 		if ((fe->fe_cx & FPSCR_FI) ||
344 		    (fe->fe_fpscr & FPSCR_UX))
345 			fe->fe_cx |= FPSCR_UX;
346 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
347 	}
348 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
349 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
350 #ifdef DIAGNOSTIC
351 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
352 		panic("fpu_ftos");
353 #endif
354 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
355 		exp++;
356 	if (exp >= SNG_EXP_INFNAN) {
357 		/* overflow to inf or to max single */
358 		if (toinf(fe, sign))
359 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
360 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
361 	}
362 done:
363 	/* phew, made it */
364 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
365 }
366 
367 /*
368  * fpn -> double (32 bit high-order result returned; 32-bit low order result
369  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
370  *
371  * This code mimics fpu_ftos; see it for comments.
372  */
373 u_int
374 fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *res)
375 {
376 	u_int sign = fp->fp_sign << 31;
377 	int exp;
378 
379 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
380 #define	DBL_MASK	(DBL_EXP(1) - 1)
381 
382 	if (ISNAN(fp)) {
383 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
384 		exp = DBL_EXP_INFNAN;
385 		goto done;
386 	}
387 	if (ISINF(fp)) {
388 		sign |= DBL_EXP(DBL_EXP_INFNAN);
389 		goto zero;
390 	}
391 	if (ISZERO(fp)) {
392 zero:		res[1] = 0;
393 		return (sign);
394 	}
395 
396 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
397 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
398 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
399 			res[1] = 0;
400 			return (sign | DBL_EXP(1) | 0);
401 		}
402 		if ((fe->fe_cx & FPSCR_FI) ||
403 		    (fe->fe_fpscr & FPSCR_UX))
404 			fe->fe_cx |= FPSCR_UX;
405 		exp = 0;
406 		goto done;
407 	}
408 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
409 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
410 		exp++;
411 	if (exp >= DBL_EXP_INFNAN) {
412 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
413 		if (toinf(fe, sign)) {
414 			res[1] = 0;
415 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
416 		}
417 		res[1] = ~0;
418 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
419 	}
420 done:
421 	res[1] = fp->fp_mant[3];
422 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
423 }
424 
425 /*
426  * Implode an fpn, writing the result into the given space.
427  */
428 void
429 fpu_implode(struct fpemu *fe, struct fpn *fp, int type, u_int *space)
430 {
431 
432 	switch (type) {
433 
434 	case FTYPE_LNG:
435 		space[0] = fpu_ftox(fe, fp, space);
436 		DPRINTF(FPE_REG, ("fpu_implode: long %x %x\n",
437 			space[0], space[1]));
438 		break;
439 
440 	case FTYPE_INT:
441 		space[0] = 0;
442 		space[1] = fpu_ftoi(fe, fp);
443 		DPRINTF(FPE_REG, ("fpu_implode: int %x\n",
444 			space[1]));
445 		break;
446 
447 	case FTYPE_SNG:
448 		space[0] = fpu_ftos(fe, fp);
449 		DPRINTF(FPE_REG, ("fpu_implode: single %x\n",
450 			space[0]));
451 		break;
452 
453 	case FTYPE_DBL:
454 		space[0] = fpu_ftod(fe, fp, space);
455 		DPRINTF(FPE_REG, ("fpu_implode: double %x %x\n",
456 			space[0], space[1]));
457 		break;		break;
458 
459 	default:
460 		panic("fpu_implode: invalid type %d", type);
461 	}
462 }
463