xref: /freebsd/sys/powerpc/fpu/fpu_implode.c (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 /*	$NetBSD: fpu_implode.c,v 1.6 2005/12/11 12:18:42 christos Exp $ */
2 
3 /*-
4  * SPDX-License-Identifier: BSD-3-Clause
5  *
6  * Copyright (c) 1992, 1993
7  *	The Regents of the University of California.  All rights reserved.
8  *
9  * This software was developed by the Computer Systems Engineering group
10  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
11  * contributed to Berkeley.
12  *
13  * All advertising materials mentioning features or use of this software
14  * must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Lawrence Berkeley Laboratory.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)fpu_implode.c	8.1 (Berkeley) 6/11/93
43  */
44 
45 /*
46  * FPU subroutines: `implode' internal format numbers into the machine's
47  * `packed binary' format.
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 #include <sys/types.h>
54 #include <sys/systm.h>
55 
56 #include <machine/fpu.h>
57 #include <machine/ieee.h>
58 #include <machine/ieeefp.h>
59 
60 #include <powerpc/fpu/fpu_arith.h>
61 #include <powerpc/fpu/fpu_emu.h>
62 #include <powerpc/fpu/fpu_extern.h>
63 #include <powerpc/fpu/fpu_instr.h>
64 
65 static int round(struct fpemu *, struct fpn *);
66 static int toinf(struct fpemu *, int);
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
81 round(struct fpemu *fe, struct fpn *fp)
82 {
83 	u_int m0, m1, m2, m3;
84 	int gr, s;
85 	FPU_DECL_CARRY;
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 |= FPSCR_XX|FPSCR_FI;	/* inexact */
104 
105 	/* Go to rounddown to round down; break to round up. */
106 	switch ((fe->fe_fpscr) & FPSCR_RN) {
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 	default:
171 	case FP_RN:		/* the nearest value is always Inf */
172 		inf = 1;
173 		break;
174 
175 	case FP_RZ:		/* toward 0 => never towards Inf */
176 		inf = 0;
177 		break;
178 
179 	case FP_RP:		/* toward +Inf iff positive */
180 		inf = sign == 0;
181 		break;
182 
183 	case FP_RM:		/* toward -Inf iff negative */
184 		inf = sign;
185 		break;
186 	}
187 	if (inf)
188 		fe->fe_cx |= FPSCR_OX;
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(struct fpemu *fe, struct fpn *fp)
200 {
201 	u_int i;
202 	int sign, exp;
203 
204 	sign = fp->fp_sign;
205 	switch (fp->fp_class) {
206 	case FPC_ZERO:
207 		return (0);
208 
209 	case FPC_NUM:
210 		/*
211 		 * If exp >= 2^32, overflow.  Otherwise shift value right
212 		 * into last mantissa word (this will not exceed 0xffffffff),
213 		 * shifting any guard and round bits out into the sticky
214 		 * bit.  Then ``round'' towards zero, i.e., just set an
215 		 * inexact exception if sticky is set (see round()).
216 		 * If the result is > 0x80000000, or is positive and equals
217 		 * 0x80000000, overflow; otherwise the last fraction word
218 		 * is the result.
219 		 */
220 		if ((exp = fp->fp_exp) >= 32)
221 			break;
222 		/* NB: the following includes exp < 0 cases */
223 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
224 			fe->fe_cx |= FPSCR_UX;
225 		i = fp->fp_mant[3];
226 		if (i >= ((u_int)0x80000000 + sign))
227 			break;
228 		return (sign ? -i : i);
229 
230 	default:		/* Inf, qNaN, sNaN */
231 		break;
232 	}
233 	/* overflow: replace any inexact exception with invalid */
234 	fe->fe_cx |= FPSCR_VXCVI;
235 	return (0x7fffffff + sign);
236 }
237 
238 /*
239  * fpn -> extended int (high bits of int value returned as return value).
240  *
241  * N.B.: this conversion always rounds towards zero (this is a peculiarity
242  * of the SPARC instruction set).
243  */
244 u_int
245 fpu_ftox(struct fpemu *fe, struct fpn *fp, u_int *res)
246 {
247 	u_int64_t i;
248 	int sign, exp;
249 
250 	sign = fp->fp_sign;
251 	switch (fp->fp_class) {
252 	case FPC_ZERO:
253 		res[1] = 0;
254 		return (0);
255 
256 	case FPC_NUM:
257 		/*
258 		 * If exp >= 2^64, overflow.  Otherwise shift value right
259 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
260 		 * shifting any guard and round bits out into the sticky
261 		 * bit.  Then ``round'' towards zero, i.e., just set an
262 		 * inexact exception if sticky is set (see round()).
263 		 * If the result is > 0x8000000000000000, or is positive and equals
264 		 * 0x8000000000000000, overflow; otherwise the last fraction word
265 		 * is the result.
266 		 */
267 		if ((exp = fp->fp_exp) >= 64)
268 			break;
269 		/* NB: the following includes exp < 0 cases */
270 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
271 			fe->fe_cx |= FPSCR_UX;
272 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
273 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
274 			break;
275 		return (sign ? -i : i);
276 
277 	default:		/* Inf, qNaN, sNaN */
278 		break;
279 	}
280 	/* overflow: replace any inexact exception with invalid */
281 	fe->fe_cx |= FPSCR_VXCVI;
282 	return (0x7fffffffffffffffLL + sign);
283 }
284 
285 /*
286  * fpn -> single (32 bit single returned as return value).
287  * We assume <= 29 bits in a single-precision fraction (1.f part).
288  */
289 u_int
290 fpu_ftos(struct fpemu *fe, struct fpn *fp)
291 {
292 	u_int sign = fp->fp_sign << 31;
293 	int exp;
294 
295 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
296 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
297 
298 	/* Take care of non-numbers first. */
299 	if (ISNAN(fp)) {
300 		/*
301 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
302 		 * Note that fp->fp_mant[0] has the quiet bit set,
303 		 * even if it is classified as a signalling NaN.
304 		 */
305 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
306 		exp = SNG_EXP_INFNAN;
307 		goto done;
308 	}
309 	if (ISINF(fp))
310 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
311 	if (ISZERO(fp))
312 		return (sign);
313 
314 	/*
315 	 * Normals (including subnormals).  Drop all the fraction bits
316 	 * (including the explicit ``implied'' 1 bit) down into the
317 	 * single-precision range.  If the number is subnormal, move
318 	 * the ``implied'' 1 into the explicit range as well, and shift
319 	 * right to introduce leading zeroes.  Rounding then acts
320 	 * differently for normals and subnormals: the largest subnormal
321 	 * may round to the smallest normal (1.0 x 2^minexp), or may
322 	 * remain subnormal.  In the latter case, signal an underflow
323 	 * if the result was inexact or if underflow traps are enabled.
324 	 *
325 	 * Rounding a normal, on the other hand, always produces another
326 	 * normal (although either way the result might be too big for
327 	 * single precision, and cause an overflow).  If rounding a
328 	 * normal produces 2.0 in the fraction, we need not adjust that
329 	 * fraction at all, since both 1.0 and 2.0 are zero under the
330 	 * fraction mask.
331 	 *
332 	 * Note that the guard and round bits vanish from the number after
333 	 * rounding.
334 	 */
335 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
336 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
337 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
338 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
339 			return (sign | SNG_EXP(1) | 0);
340 		if ((fe->fe_cx & FPSCR_FI) ||
341 		    (fe->fe_fpscr & FPSCR_UX))
342 			fe->fe_cx |= FPSCR_UX;
343 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
344 	}
345 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
346 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
347 #ifdef DIAGNOSTIC
348 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
349 		panic("fpu_ftos");
350 #endif
351 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
352 		exp++;
353 	if (exp >= SNG_EXP_INFNAN) {
354 		/* overflow to inf or to max single */
355 		if (toinf(fe, sign))
356 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
357 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
358 	}
359 done:
360 	/* phew, made it */
361 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
362 }
363 
364 /*
365  * fpn -> double (32 bit high-order result returned; 32-bit low order result
366  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
367  *
368  * This code mimics fpu_ftos; see it for comments.
369  */
370 u_int
371 fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *res)
372 {
373 	u_int sign = fp->fp_sign << 31;
374 	int exp;
375 
376 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
377 #define	DBL_MASK	(DBL_EXP(1) - 1)
378 
379 	if (ISNAN(fp)) {
380 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
381 		exp = DBL_EXP_INFNAN;
382 		goto done;
383 	}
384 	if (ISINF(fp)) {
385 		sign |= DBL_EXP(DBL_EXP_INFNAN);
386 		goto zero;
387 	}
388 	if (ISZERO(fp)) {
389 zero:		res[1] = 0;
390 		return (sign);
391 	}
392 
393 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
394 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
395 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
396 			res[1] = 0;
397 			return (sign | DBL_EXP(1) | 0);
398 		}
399 		if ((fe->fe_cx & FPSCR_FI) ||
400 		    (fe->fe_fpscr & FPSCR_UX))
401 			fe->fe_cx |= FPSCR_UX;
402 		exp = 0;
403 		goto done;
404 	}
405 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
406 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
407 		exp++;
408 	if (exp >= DBL_EXP_INFNAN) {
409 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
410 		if (toinf(fe, sign)) {
411 			res[1] = 0;
412 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
413 		}
414 		res[1] = ~0;
415 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
416 	}
417 done:
418 	res[1] = fp->fp_mant[3];
419 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
420 }
421 
422 /*
423  * Implode an fpn, writing the result into the given space.
424  */
425 void
426 fpu_implode(struct fpemu *fe, struct fpn *fp, int type, u_int *space)
427 {
428 
429 	switch (type) {
430 	case FTYPE_LNG:
431 		space[0] = fpu_ftox(fe, fp, space);
432 		DPRINTF(FPE_REG, ("fpu_implode: long %x %x\n",
433 			space[0], space[1]));
434 		break;
435 
436 	case FTYPE_INT:
437 		space[0] = 0;
438 		space[1] = fpu_ftoi(fe, fp);
439 		DPRINTF(FPE_REG, ("fpu_implode: int %x\n",
440 			space[1]));
441 		break;
442 
443 	case FTYPE_SNG:
444 		space[0] = fpu_ftos(fe, fp);
445 		DPRINTF(FPE_REG, ("fpu_implode: single %x\n",
446 			space[0]));
447 		break;
448 
449 	case FTYPE_DBL:
450 		space[0] = fpu_ftod(fe, fp, space);
451 		DPRINTF(FPE_REG, ("fpu_implode: double %x %x\n",
452 			space[0], space[1]));
453 		break;		break;
454 
455 	default:
456 		panic("fpu_implode: invalid type %d", type);
457 	}
458 }
459