xref: /freebsd/sys/powerpc/fpu/fpu_implode.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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 #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 
58 #include <powerpc/fpu/fpu_arith.h>
59 #include <powerpc/fpu/fpu_emu.h>
60 #include <powerpc/fpu/fpu_extern.h>
61 #include <powerpc/fpu/fpu_instr.h>
62 
63 static int round(struct fpemu *, struct fpn *);
64 static int toinf(struct fpemu *, int);
65 
66 /*
67  * Round a number (algorithm from Motorola MC68882 manual, modified for
68  * our internal format).  Set inexact exception if rounding is required.
69  * Return true iff we rounded up.
70  *
71  * After rounding, we discard the guard and round bits by shifting right
72  * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky).
73  * This saves effort later.
74  *
75  * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's
76  * responsibility to fix this if necessary.
77  */
78 static int
79 round(struct fpemu *fe, struct fpn *fp)
80 {
81 	u_int m0, m1, m2, m3;
82 	int gr, s;
83 	FPU_DECL_CARRY;
84 
85 	m0 = fp->fp_mant[0];
86 	m1 = fp->fp_mant[1];
87 	m2 = fp->fp_mant[2];
88 	m3 = fp->fp_mant[3];
89 	gr = m3 & 3;
90 	s = fp->fp_sticky;
91 
92 	/* mant >>= FP_NG */
93 	m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG));
94 	m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
95 	m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
96 	m0 >>= FP_NG;
97 
98 	if ((gr | s) == 0)	/* result is exact: no rounding needed */
99 		goto rounddown;
100 
101 	fe->fe_cx |= FPSCR_XX|FPSCR_FI;	/* inexact */
102 
103 	/* Go to rounddown to round down; break to round up. */
104 	switch ((fe->fe_fpscr) & FPSCR_RN) {
105 	case FP_RN:
106 	default:
107 		/*
108 		 * Round only if guard is set (gr & 2).  If guard is set,
109 		 * but round & sticky both clear, then we want to round
110 		 * but have a tie, so round to even, i.e., add 1 iff odd.
111 		 */
112 		if ((gr & 2) == 0)
113 			goto rounddown;
114 		if ((gr & 1) || fp->fp_sticky || (m3 & 1))
115 			break;
116 		goto rounddown;
117 
118 	case FP_RZ:
119 		/* Round towards zero, i.e., down. */
120 		goto rounddown;
121 
122 	case FP_RM:
123 		/* Round towards -Inf: up if negative, down if positive. */
124 		if (fp->fp_sign)
125 			break;
126 		goto rounddown;
127 
128 	case FP_RP:
129 		/* Round towards +Inf: up if positive, down otherwise. */
130 		if (!fp->fp_sign)
131 			break;
132 		goto rounddown;
133 	}
134 
135 	/* Bump low bit of mantissa, with carry. */
136 	fe->fe_cx |= FPSCR_FR;
137 
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 ((fe->fe_fpscr) & FPSCR_RN) {
168 	default:
169 	case FP_RN:		/* the nearest value is always Inf */
170 		inf = 1;
171 		break;
172 
173 	case FP_RZ:		/* toward 0 => never towards Inf */
174 		inf = 0;
175 		break;
176 
177 	case FP_RP:		/* toward +Inf iff positive */
178 		inf = sign == 0;
179 		break;
180 
181 	case FP_RM:		/* toward -Inf iff negative */
182 		inf = sign;
183 		break;
184 	}
185 	if (inf)
186 		fe->fe_cx |= FPSCR_OX;
187 	return (inf);
188 }
189 
190 /*
191  * fpn -> int (int value returned as return value).
192  *
193  * N.B.: this conversion always rounds towards zero (this is a peculiarity
194  * of the SPARC instruction set).
195  */
196 u_int
197 fpu_ftoi(struct fpemu *fe, struct fpn *fp)
198 {
199 	u_int i;
200 	int sign, exp;
201 
202 	sign = fp->fp_sign;
203 	switch (fp->fp_class) {
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 round()).
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 |= FPSCR_UX;
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 |= FPSCR_VXCVI;
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
243 fpu_ftox(struct fpemu *fe, struct fpn *fp, u_int *res)
244 {
245 	u_int64_t i;
246 	int sign, exp;
247 
248 	sign = fp->fp_sign;
249 	switch (fp->fp_class) {
250 	case FPC_ZERO:
251 		res[1] = 0;
252 		return (0);
253 
254 	case FPC_NUM:
255 		/*
256 		 * If exp >= 2^64, overflow.  Otherwise shift value right
257 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
258 		 * shifting any guard and round bits out into the sticky
259 		 * bit.  Then ``round'' towards zero, i.e., just set an
260 		 * inexact exception if sticky is set (see round()).
261 		 * If the result is > 0x8000000000000000, or is positive and equals
262 		 * 0x8000000000000000, overflow; otherwise the last fraction word
263 		 * is the result.
264 		 */
265 		if ((exp = fp->fp_exp) >= 64)
266 			break;
267 		/* NB: the following includes exp < 0 cases */
268 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
269 			fe->fe_cx |= FPSCR_UX;
270 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
271 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
272 			break;
273 		return (sign ? -i : i);
274 
275 	default:		/* Inf, qNaN, sNaN */
276 		break;
277 	}
278 	/* overflow: replace any inexact exception with invalid */
279 	fe->fe_cx |= FPSCR_VXCVI;
280 	return (0x7fffffffffffffffLL + sign);
281 }
282 
283 /*
284  * fpn -> single (32 bit single returned as return value).
285  * We assume <= 29 bits in a single-precision fraction (1.f part).
286  */
287 u_int
288 fpu_ftos(struct fpemu *fe, struct fpn *fp)
289 {
290 	u_int sign = fp->fp_sign << 31;
291 	int exp;
292 
293 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
294 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
295 
296 	/* Take care of non-numbers first. */
297 	if (ISNAN(fp)) {
298 		/*
299 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
300 		 * Note that fp->fp_mant[0] has the quiet bit set,
301 		 * even if it is classified as a signalling NaN.
302 		 */
303 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
304 		exp = SNG_EXP_INFNAN;
305 		goto done;
306 	}
307 	if (ISINF(fp))
308 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
309 	if (ISZERO(fp))
310 		return (sign);
311 
312 	/*
313 	 * Normals (including subnormals).  Drop all the fraction bits
314 	 * (including the explicit ``implied'' 1 bit) down into the
315 	 * single-precision range.  If the number is subnormal, move
316 	 * the ``implied'' 1 into the explicit range as well, and shift
317 	 * right to introduce leading zeroes.  Rounding then acts
318 	 * differently for normals and subnormals: the largest subnormal
319 	 * may round to the smallest normal (1.0 x 2^minexp), or may
320 	 * remain subnormal.  In the latter case, signal an underflow
321 	 * if the result was inexact or if underflow traps are enabled.
322 	 *
323 	 * Rounding a normal, on the other hand, always produces another
324 	 * normal (although either way the result might be too big for
325 	 * single precision, and cause an overflow).  If rounding a
326 	 * normal produces 2.0 in the fraction, we need not adjust that
327 	 * fraction at all, since both 1.0 and 2.0 are zero under the
328 	 * fraction mask.
329 	 *
330 	 * Note that the guard and round bits vanish from the number after
331 	 * rounding.
332 	 */
333 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
334 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
335 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
336 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
337 			return (sign | SNG_EXP(1) | 0);
338 		if ((fe->fe_cx & FPSCR_FI) ||
339 		    (fe->fe_fpscr & FPSCR_UX))
340 			fe->fe_cx |= FPSCR_UX;
341 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
342 	}
343 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
344 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
345 #ifdef DIAGNOSTIC
346 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
347 		panic("fpu_ftos");
348 #endif
349 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
350 		exp++;
351 	if (exp >= SNG_EXP_INFNAN) {
352 		/* overflow to inf or to max single */
353 		if (toinf(fe, sign))
354 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
355 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
356 	}
357 done:
358 	/* phew, made it */
359 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
360 }
361 
362 /*
363  * fpn -> double (32 bit high-order result returned; 32-bit low order result
364  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
365  *
366  * This code mimics fpu_ftos; see it for comments.
367  */
368 u_int
369 fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *res)
370 {
371 	u_int sign = fp->fp_sign << 31;
372 	int exp;
373 
374 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
375 #define	DBL_MASK	(DBL_EXP(1) - 1)
376 
377 	if (ISNAN(fp)) {
378 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
379 		exp = DBL_EXP_INFNAN;
380 		goto done;
381 	}
382 	if (ISINF(fp)) {
383 		sign |= DBL_EXP(DBL_EXP_INFNAN);
384 		goto zero;
385 	}
386 	if (ISZERO(fp)) {
387 zero:		res[1] = 0;
388 		return (sign);
389 	}
390 
391 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
392 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
393 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
394 			res[1] = 0;
395 			return (sign | DBL_EXP(1) | 0);
396 		}
397 		if ((fe->fe_cx & FPSCR_FI) ||
398 		    (fe->fe_fpscr & FPSCR_UX))
399 			fe->fe_cx |= FPSCR_UX;
400 		exp = 0;
401 		goto done;
402 	}
403 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
404 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
405 		exp++;
406 	if (exp >= DBL_EXP_INFNAN) {
407 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
408 		if (toinf(fe, sign)) {
409 			res[1] = 0;
410 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
411 		}
412 		res[1] = ~0;
413 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
414 	}
415 done:
416 	res[1] = fp->fp_mant[3];
417 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
418 }
419 
420 /*
421  * Implode an fpn, writing the result into the given space.
422  */
423 void
424 fpu_implode(struct fpemu *fe, struct fpn *fp, int type, u_int *space)
425 {
426 
427 	switch (type) {
428 	case FTYPE_LNG:
429 		space[0] = fpu_ftox(fe, fp, space);
430 		DPRINTF(FPE_REG, ("fpu_implode: long %x %x\n",
431 			space[0], space[1]));
432 		break;
433 
434 	case FTYPE_INT:
435 		space[0] = 0;
436 		space[1] = fpu_ftoi(fe, fp);
437 		DPRINTF(FPE_REG, ("fpu_implode: int %x\n",
438 			space[1]));
439 		break;
440 
441 	case FTYPE_SNG:
442 		space[0] = fpu_ftos(fe, fp);
443 		DPRINTF(FPE_REG, ("fpu_implode: single %x\n",
444 			space[0]));
445 		break;
446 
447 	case FTYPE_DBL:
448 		space[0] = fpu_ftod(fe, fp, space);
449 		DPRINTF(FPE_REG, ("fpu_implode: double %x %x\n",
450 			space[0], space[1]));
451 		break;		break;
452 
453 	default:
454 		panic("fpu_implode: invalid type %d", type);
455 	}
456 }
457