xref: /netbsd-src/sys/arch/powerpc/fpu/fpu_implode.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: fpu_implode.c,v 1.4 2003/10/27 04:32:02 simonb 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 __KERNEL_RCSID(0, "$NetBSD: fpu_implode.c,v 1.4 2003/10/27 04:32:02 simonb Exp $");
50 
51 #include <sys/types.h>
52 #include <sys/systm.h>
53 
54 #include <machine/ieee.h>
55 #include <powerpc/instr.h>
56 #include <machine/reg.h>
57 #include <machine/fpu.h>
58 
59 #include <powerpc/fpu/fpu_arith.h>
60 #include <powerpc/fpu/fpu_emu.h>
61 #include <powerpc/fpu/fpu_extern.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 
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 	fe->fe_cx |= FPSCR_FR;
138 
139 	FPU_ADDS(m3, m3, 1);
140 	FPU_ADDCS(m2, m2, 0);
141 	FPU_ADDCS(m1, m1, 0);
142 	FPU_ADDC(m0, m0, 0);
143 	fp->fp_mant[0] = m0;
144 	fp->fp_mant[1] = m1;
145 	fp->fp_mant[2] = m2;
146 	fp->fp_mant[3] = m3;
147 	return (1);
148 
149 rounddown:
150 	fp->fp_mant[0] = m0;
151 	fp->fp_mant[1] = m1;
152 	fp->fp_mant[2] = m2;
153 	fp->fp_mant[3] = m3;
154 	return (0);
155 }
156 
157 /*
158  * For overflow: return true if overflow is to go to +/-Inf, according
159  * to the sign of the overflowing result.  If false, overflow is to go
160  * to the largest magnitude value instead.
161  */
162 static int
163 toinf(struct fpemu *fe, int sign)
164 {
165 	int inf;
166 
167 	/* look at rounding direction */
168 	switch ((fe->fe_fpscr) & FPSCR_RN) {
169 
170 	default:
171 	case FSR_RD_RN:		/* the nearest value is always Inf */
172 		inf = 1;
173 		break;
174 
175 	case FSR_RD_RZ:		/* toward 0 => never towards Inf */
176 		inf = 0;
177 		break;
178 
179 	case FSR_RD_RP:		/* toward +Inf iff positive */
180 		inf = sign == 0;
181 		break;
182 
183 	case FSR_RD_RM:		/* toward -Inf iff negative */
184 		inf = sign;
185 		break;
186 	}
187 	if (inf) fe->fe_cx |= FPSCR_OX;
188 	return (inf);
189 }
190 
191 /*
192  * fpn -> int (int value returned as return value).
193  *
194  * N.B.: this conversion always rounds towards zero (this is a peculiarity
195  * of the SPARC instruction set).
196  */
197 u_int
198 fpu_ftoi(struct fpemu *fe, struct fpn *fp)
199 {
200 	u_int i;
201 	int sign, exp;
202 
203 	sign = fp->fp_sign;
204 	switch (fp->fp_class) {
205 
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 
253 	case FPC_ZERO:
254 		res[1] = 0;
255 		return (0);
256 
257 	case FPC_NUM:
258 		/*
259 		 * If exp >= 2^64, overflow.  Otherwise shift value right
260 		 * into last mantissa word (this will not exceed 0xffffffffffffffff),
261 		 * shifting any guard and round bits out into the sticky
262 		 * bit.  Then ``round'' towards zero, i.e., just set an
263 		 * inexact exception if sticky is set (see round()).
264 		 * If the result is > 0x8000000000000000, or is positive and equals
265 		 * 0x8000000000000000, overflow; otherwise the last fraction word
266 		 * is the result.
267 		 */
268 		if ((exp = fp->fp_exp) >= 64)
269 			break;
270 		/* NB: the following includes exp < 0 cases */
271 		if (fpu_shr(fp, FP_NMANT - 1 - exp) != 0)
272 			fe->fe_cx |= FPSCR_UX;
273 		i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3];
274 		if (i >= ((u_int64_t)0x8000000000000000LL + sign))
275 			break;
276 		return (sign ? -i : i);
277 
278 	default:		/* Inf, qNaN, sNaN */
279 		break;
280 	}
281 	/* overflow: replace any inexact exception with invalid */
282 	fe->fe_cx |= FPSCR_VXCVI;
283 	return (0x7fffffffffffffffLL + sign);
284 }
285 
286 /*
287  * fpn -> single (32 bit single returned as return value).
288  * We assume <= 29 bits in a single-precision fraction (1.f part).
289  */
290 u_int
291 fpu_ftos(struct fpemu *fe, struct fpn *fp)
292 {
293 	u_int sign = fp->fp_sign << 31;
294 	int exp;
295 
296 #define	SNG_EXP(e)	((e) << SNG_FRACBITS)	/* makes e an exponent */
297 #define	SNG_MASK	(SNG_EXP(1) - 1)	/* mask for fraction */
298 
299 	/* Take care of non-numbers first. */
300 	if (ISNAN(fp)) {
301 		/*
302 		 * Preserve upper bits of NaN, per SPARC V8 appendix N.
303 		 * Note that fp->fp_mant[0] has the quiet bit set,
304 		 * even if it is classified as a signalling NaN.
305 		 */
306 		(void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
307 		exp = SNG_EXP_INFNAN;
308 		goto done;
309 	}
310 	if (ISINF(fp))
311 		return (sign | SNG_EXP(SNG_EXP_INFNAN));
312 	if (ISZERO(fp))
313 		return (sign);
314 
315 	/*
316 	 * Normals (including subnormals).  Drop all the fraction bits
317 	 * (including the explicit ``implied'' 1 bit) down into the
318 	 * single-precision range.  If the number is subnormal, move
319 	 * the ``implied'' 1 into the explicit range as well, and shift
320 	 * right to introduce leading zeroes.  Rounding then acts
321 	 * differently for normals and subnormals: the largest subnormal
322 	 * may round to the smallest normal (1.0 x 2^minexp), or may
323 	 * remain subnormal.  In the latter case, signal an underflow
324 	 * if the result was inexact or if underflow traps are enabled.
325 	 *
326 	 * Rounding a normal, on the other hand, always produces another
327 	 * normal (although either way the result might be too big for
328 	 * single precision, and cause an overflow).  If rounding a
329 	 * normal produces 2.0 in the fraction, we need not adjust that
330 	 * fraction at all, since both 1.0 and 2.0 are zero under the
331 	 * fraction mask.
332 	 *
333 	 * Note that the guard and round bits vanish from the number after
334 	 * rounding.
335 	 */
336 	if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) {	/* subnormal */
337 		/* -NG for g,r; -SNG_FRACBITS-exp for fraction */
338 		(void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
339 		if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1))
340 			return (sign | SNG_EXP(1) | 0);
341 		if ((fe->fe_cx & FPSCR_FI) ||
342 		    (fe->fe_fpscr & FPSCR_UX))
343 			fe->fe_cx |= FPSCR_UX;
344 		return (sign | SNG_EXP(0) | fp->fp_mant[3]);
345 	}
346 	/* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
347 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
348 #ifdef DIAGNOSTIC
349 	if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
350 		panic("fpu_ftos");
351 #endif
352 	if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2))
353 		exp++;
354 	if (exp >= SNG_EXP_INFNAN) {
355 		/* overflow to inf or to max single */
356 		if (toinf(fe, sign))
357 			return (sign | SNG_EXP(SNG_EXP_INFNAN));
358 		return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
359 	}
360 done:
361 	/* phew, made it */
362 	return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK));
363 }
364 
365 /*
366  * fpn -> double (32 bit high-order result returned; 32-bit low order result
367  * left in res[1]).  Assumes <= 61 bits in double precision fraction.
368  *
369  * This code mimics fpu_ftos; see it for comments.
370  */
371 u_int
372 fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *res)
373 {
374 	u_int sign = fp->fp_sign << 31;
375 	int exp;
376 
377 #define	DBL_EXP(e)	((e) << (DBL_FRACBITS & 31))
378 #define	DBL_MASK	(DBL_EXP(1) - 1)
379 
380 	if (ISNAN(fp)) {
381 		(void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
382 		exp = DBL_EXP_INFNAN;
383 		goto done;
384 	}
385 	if (ISINF(fp)) {
386 		sign |= DBL_EXP(DBL_EXP_INFNAN);
387 		goto zero;
388 	}
389 	if (ISZERO(fp)) {
390 zero:		res[1] = 0;
391 		return (sign);
392 	}
393 
394 	if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
395 		(void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
396 		if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) {
397 			res[1] = 0;
398 			return (sign | DBL_EXP(1) | 0);
399 		}
400 		if ((fe->fe_cx & FPSCR_FI) ||
401 		    (fe->fe_fpscr & FPSCR_UX))
402 			fe->fe_cx |= FPSCR_UX;
403 		exp = 0;
404 		goto done;
405 	}
406 	(void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
407 	if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2))
408 		exp++;
409 	if (exp >= DBL_EXP_INFNAN) {
410 		fe->fe_cx |= FPSCR_OX | FPSCR_UX;
411 		if (toinf(fe, sign)) {
412 			res[1] = 0;
413 			return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
414 		}
415 		res[1] = ~0;
416 		return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
417 	}
418 done:
419 	res[1] = fp->fp_mant[3];
420 	return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK));
421 }
422 
423 /*
424  * Implode an fpn, writing the result into the given space.
425  */
426 void
427 fpu_implode(struct fpemu *fe, struct fpn *fp, int type, u_int *space)
428 {
429 
430 	switch (type) {
431 
432 	case FTYPE_LNG:
433 		space[0] = fpu_ftox(fe, fp, space);
434 		DPRINTF(FPE_REG, ("fpu_implode: long %x %x\n",
435 			space[0], space[1]));
436 		break;
437 
438 	case FTYPE_INT:
439 		space[0] = 0;
440 		space[1] = fpu_ftoi(fe, fp);
441 		DPRINTF(FPE_REG, ("fpu_implode: int %x\n",
442 			space[1]));
443 		break;
444 
445 	case FTYPE_SNG:
446 		space[0] = fpu_ftos(fe, fp);
447 		DPRINTF(FPE_REG, ("fpu_implode: single %x\n",
448 			space[0]));
449 		break;
450 
451 	case FTYPE_DBL:
452 		space[0] = fpu_ftod(fe, fp, space);
453 		DPRINTF(FPE_REG, ("fpu_implode: double %x %x\n",
454 			space[0], space[1]));
455 		break;		break;
456 
457 	default:
458 		panic("fpu_implode: invalid type %d", type);
459 	}
460 }
461