Annotation of sys/arch/m68k/fpe/fpu_implode.c, Revision 1.1.1.1
1.1 nbrk 1: /* $OpenBSD: fpu_implode.c,v 1.6 2006/06/11 20:43:28 miod Exp $ */
2: /* $NetBSD: fpu_implode.c,v 1.8 2003/10/23 15:07:30 kleink Exp $ */
3:
4: /*
5: * Copyright (c) 1992, 1993
6: * The Regents of the University of California. All rights reserved.
7: *
8: * This software was developed by the Computer Systems Engineering group
9: * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
10: * contributed to Berkeley.
11: *
12: * All advertising materials mentioning features or use of this software
13: * must display the following acknowledgement:
14: * This product includes software developed by the University of
15: * California, Lawrence Berkeley Laboratory.
16: *
17: * Redistribution and use in source and binary forms, with or without
18: * modification, are permitted provided that the following conditions
19: * are met:
20: * 1. Redistributions of source code must retain the above copyright
21: * notice, this list of conditions and the following disclaimer.
22: * 2. Redistributions in binary form must reproduce the above copyright
23: * notice, this list of conditions and the following disclaimer in the
24: * documentation and/or other materials provided with the distribution.
25: * 3. Neither the name of the University nor the names of its contributors
26: * may be used to endorse or promote products derived from this software
27: * without specific prior written permission.
28: *
29: * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30: * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31: * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32: * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33: * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34: * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35: * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36: * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37: * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38: * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39: * SUCH DAMAGE.
40: *
41: * @(#)fpu_implode.c 8.1 (Berkeley) 6/11/93
42: */
43:
44: /*
45: * FPU subroutines: `implode' internal format numbers into the machine's
46: * `packed binary' format.
47: */
48:
49: #include <sys/types.h>
50: #include <sys/systm.h>
51:
52: #include <machine/ieee.h>
53: #include <machine/reg.h>
54:
55: #include <m68k/fpe/fpu_emulate.h>
56: #include <m68k/fpe/fpu_arith.h>
57:
58: /* Conversion from internal format -- note asymmetry. */
59: u_int fpu_ftoi(struct fpemu *fe, struct fpn *fp);
60: u_int fpu_ftos(struct fpemu *fe, struct fpn *fp);
61: u_int fpu_ftod(struct fpemu *fe, struct fpn *fp, u_int *);
62: u_int fpu_ftox(struct fpemu *fe, struct fpn *fp, u_int *);
63:
64: int toinf(struct fpemu *fe, int sign);
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: int
79: fpu_round(struct fpemu *fe, struct fpn *fp)
80: {
81: u_int m0, m1, m2;
82: int gr, s;
83:
84: m0 = fp->fp_mant[0];
85: m1 = fp->fp_mant[1];
86: m2 = fp->fp_mant[2];
87: gr = m2 & 3;
88: s = fp->fp_sticky;
89:
90: /* mant >>= FP_NG */
91: m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG));
92: m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG));
93: m0 >>= FP_NG;
94:
95: if ((gr | s) == 0) /* result is exact: no rounding needed */
96: goto rounddown;
97:
98: fe->fe_fpsr |= FPSR_INEX2; /* inexact */
99:
100: /* Go to rounddown to round down; break to round up. */
101: switch (fe->fe_fpcr & FPCR_ROUND) {
102:
103: case FPCR_NEAR:
104: default:
105: /*
106: * Round only if guard is set (gr & 2). If guard is set,
107: * but round & sticky both clear, then we want to round
108: * but have a tie, so round to even, i.e., add 1 iff odd.
109: */
110: if ((gr & 2) == 0)
111: goto rounddown;
112: if ((gr & 1) || fp->fp_sticky || (m2 & 1))
113: break;
114: goto rounddown;
115:
116: case FPCR_ZERO:
117: /* Round towards zero, i.e., down. */
118: goto rounddown;
119:
120: case FPCR_MINF:
121: /* Round towards -Inf: up if negative, down if positive. */
122: if (fp->fp_sign)
123: break;
124: goto rounddown;
125:
126: case FPCR_PINF:
127: /* Round towards +Inf: up if positive, down otherwise. */
128: if (!fp->fp_sign)
129: break;
130: goto rounddown;
131: }
132:
133: /* Bump low bit of mantissa, with carry. */
134: if (++m2 == 0 && ++m1 == 0)
135: m0++;
136: fp->fp_sticky = 0;
137: fp->fp_mant[0] = m0;
138: fp->fp_mant[1] = m1;
139: fp->fp_mant[2] = m2;
140: return (1);
141:
142: rounddown:
143: fp->fp_sticky = 0;
144: fp->fp_mant[0] = m0;
145: fp->fp_mant[1] = m1;
146: fp->fp_mant[2] = m2;
147: return (0);
148: }
149:
150: /*
151: * For overflow: return true if overflow is to go to +/-Inf, according
152: * to the sign of the overflowing result. If false, overflow is to go
153: * to the largest magnitude value instead.
154: */
155: int
156: toinf(struct fpemu *fe, int sign)
157: {
158: int inf;
159:
160: /* look at rounding direction */
161: switch (fe->fe_fpcr & FPCR_ROUND) {
162:
163: default:
164: case FPCR_NEAR: /* the nearest value is always Inf */
165: inf = 1;
166: break;
167:
168: case FPCR_ZERO: /* toward 0 => never towards Inf */
169: inf = 0;
170: break;
171:
172: case FPCR_PINF: /* toward +Inf iff positive */
173: inf = (sign == 0);
174: break;
175:
176: case FPCR_MINF: /* toward -Inf iff negative */
177: inf = sign;
178: break;
179: }
180: return (inf);
181: }
182:
183: /*
184: * fpn -> int (int value returned as return value).
185: *
186: * N.B.: this conversion always rounds towards zero (this is a peculiarity
187: * of the SPARC instruction set).
188: */
189: u_int
190: fpu_ftoi(fe, fp)
191: struct fpemu *fe;
192: struct fpn *fp;
193: {
194: u_int i;
195: int sign, exp;
196:
197: sign = fp->fp_sign;
198: switch (fp->fp_class) {
199:
200: case FPC_ZERO:
201: return (0);
202:
203: case FPC_NUM:
204: /*
205: * If exp >= 2^32, overflow. Otherwise shift value right
206: * into last mantissa word (this will not exceed 0xffffffff),
207: * shifting any guard and round bits out into the sticky
208: * bit. Then ``round'' towards zero, i.e., just set an
209: * inexact exception if sticky is set (see fpu_round()).
210: * If the result is > 0x80000000, or is positive and equals
211: * 0x80000000, overflow; otherwise the last fraction word
212: * is the result.
213: */
214: if ((exp = fp->fp_exp) >= 32)
215: break;
216: /* NB: the following includes exp < 0 cases */
217: if (fpu_shr(fp, FP_NMANT - 1 - FP_NG - exp) != 0)
218: /* m68881/2 do not underflow when
219: converting to integer */;
220: fpu_round(fe, fp);
221: i = fp->fp_mant[2];
222: if (i >= ((u_int)0x80000000 + sign))
223: break;
224: return (sign ? -i : i);
225:
226: default: /* Inf, qNaN, sNaN */
227: break;
228: }
229: /* overflow: replace any inexact exception with invalid */
230: fe->fe_fpsr = (fe->fe_fpsr & ~FPSR_INEX2) | FPSR_OPERR;
231: return (0x7fffffff + sign);
232: }
233:
234: /*
235: * fpn -> single (32 bit single returned as return value).
236: * We assume <= 29 bits in a single-precision fraction (1.f part).
237: */
238: u_int
239: fpu_ftos(fe, fp)
240: struct fpemu *fe;
241: struct fpn *fp;
242: {
243: u_int sign = fp->fp_sign << 31;
244: int exp;
245:
246: #define SNG_EXP(e) ((e) << SNG_FRACBITS) /* makes e an exponent */
247: #define SNG_MASK (SNG_EXP(1) - 1) /* mask for fraction */
248:
249: /* Take care of non-numbers first. */
250: if (ISNAN(fp)) {
251: /*
252: * Preserve upper bits of NaN, per SPARC V8 appendix N.
253: * Note that fp->fp_mant[0] has the quiet bit set,
254: * even if it is classified as a signalling NaN.
255: */
256: (void) fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS);
257: exp = SNG_EXP_INFNAN;
258: goto done;
259: }
260: if (ISINF(fp))
261: return (sign | SNG_EXP(SNG_EXP_INFNAN));
262: if (ISZERO(fp))
263: return (sign);
264:
265: /*
266: * Normals (including subnormals). Drop all the fraction bits
267: * (including the explicit ``implied'' 1 bit) down into the
268: * single-precision range. If the number is subnormal, move
269: * the ``implied'' 1 into the explicit range as well, and shift
270: * right to introduce leading zeroes. Rounding then acts
271: * differently for normals and subnormals: the largest subnormal
272: * may round to the smallest normal (1.0 x 2^minexp), or may
273: * remain subnormal. In the latter case, signal an underflow
274: * if the result was inexact or if underflow traps are enabled.
275: *
276: * Rounding a normal, on the other hand, always produces another
277: * normal (although either way the result might be too big for
278: * single precision, and cause an overflow). If rounding a
279: * normal produces 2.0 in the fraction, we need not adjust that
280: * fraction at all, since both 1.0 and 2.0 are zero under the
281: * fraction mask.
282: *
283: * Note that the guard and round bits vanish from the number after
284: * rounding.
285: */
286: if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) { /* subnormal */
287: fe->fe_fpsr |= FPSR_UNFL;
288: /* -NG for g,r; -SNG_FRACBITS-exp for fraction */
289: (void) fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp);
290: if (fpu_round(fe, fp) && fp->fp_mant[2] == SNG_EXP(1))
291: return (sign | SNG_EXP(1) | 0);
292: if (fe->fe_fpsr & FPSR_INEX2)
293: fe->fe_fpsr |= FPSR_UNFL
294: /* mc68881/2 don't underflow when converting */;
295: return (sign | SNG_EXP(0) | fp->fp_mant[2]);
296: }
297: /* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */
298: (void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS);
299: #ifdef DIAGNOSTIC
300: if ((fp->fp_mant[2] & SNG_EXP(1 << FP_NG)) == 0)
301: panic("fpu_ftos");
302: #endif
303: if (fpu_round(fe, fp) && fp->fp_mant[2] == SNG_EXP(2))
304: exp++;
305: if (exp >= SNG_EXP_INFNAN) {
306: /* overflow to inf or to max single */
307: fe->fe_fpsr |= FPSR_OPERR | FPSR_INEX2 | FPSR_OVFL;
308: if (toinf(fe, sign))
309: return (sign | SNG_EXP(SNG_EXP_INFNAN));
310: return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK);
311: }
312: done:
313: /* phew, made it */
314: return (sign | SNG_EXP(exp) | (fp->fp_mant[2] & SNG_MASK));
315: }
316:
317: /*
318: * fpn -> double (32 bit high-order result returned; 32-bit low order result
319: * left in res[1]). Assumes <= 61 bits in double precision fraction.
320: *
321: * This code mimics fpu_ftos; see it for comments.
322: */
323: u_int
324: fpu_ftod(fe, fp, res)
325: struct fpemu *fe;
326: struct fpn *fp;
327: u_int *res;
328: {
329: u_int sign = fp->fp_sign << 31;
330: int exp;
331:
332: #define DBL_EXP(e) ((e) << (DBL_FRACBITS & 31))
333: #define DBL_MASK (DBL_EXP(1) - 1)
334:
335: if (ISNAN(fp)) {
336: (void) fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS);
337: exp = DBL_EXP_INFNAN;
338: goto done;
339: }
340: if (ISINF(fp)) {
341: sign |= DBL_EXP(DBL_EXP_INFNAN);
342: res[1] = 0;
343: return (sign);
344: }
345: if (ISZERO(fp)) {
346: res[1] = 0;
347: return (sign);
348: }
349:
350: if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) {
351: fe->fe_fpsr |= FPSR_UNFL;
352: (void) fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp);
353: if (fpu_round(fe, fp) && fp->fp_mant[1] == DBL_EXP(1)) {
354: res[1] = 0;
355: return (sign | DBL_EXP(1) | 0);
356: }
357: if (fe->fe_fpsr & FPSR_INEX2)
358: fe->fe_fpsr |= FPSR_UNFL
359: /* mc68881/2 don't underflow when converting */;
360: exp = 0;
361: goto done;
362: }
363: (void) fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS);
364: if (fpu_round(fe, fp) && fp->fp_mant[1] == DBL_EXP(2))
365: exp++;
366: if (exp >= DBL_EXP_INFNAN) {
367: fe->fe_fpsr |= FPSR_OPERR | FPSR_INEX2 | FPSR_OVFL;
368: if (toinf(fe, sign)) {
369: res[1] = 0;
370: return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0);
371: }
372: res[1] = ~0;
373: return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK);
374: }
375: done:
376: res[1] = fp->fp_mant[2];
377: return (sign | DBL_EXP(exp) | (fp->fp_mant[1] & DBL_MASK));
378: }
379:
380: /*
381: * fpn -> 68k extended (32 bit high-order result returned; two 32-bit low
382: * order result left in res[1] & res[2]). Assumes == 64 bits in extended
383: * precision fraction.
384: *
385: * This code mimics fpu_ftos; see it for comments.
386: */
387: u_int
388: fpu_ftox(fe, fp, res)
389: struct fpemu *fe;
390: struct fpn *fp;
391: u_int *res;
392: {
393: u_int sign = fp->fp_sign << 31;
394: int exp;
395:
396: #define EXT_EXP(e) ((e) << 16)
397: /*
398: * on m68k extended prec, significand does not share the same long
399: * word with exponent
400: */
401: #define EXT_MASK 0
402: #define EXT_EXPLICIT1 (1UL << (63 & 31))
403: #define EXT_EXPLICIT2 (1UL << (64 & 31))
404:
405: if (ISNAN(fp)) {
406: (void) fpu_shr(fp, FP_NMANT - EXT_FRACBITS);
407: exp = EXT_EXP_INFNAN;
408: goto done;
409: }
410: if (ISINF(fp)) {
411: sign |= EXT_EXP(EXT_EXP_INFNAN);
412: res[1] = res[2] = 0;
413: return (sign);
414: }
415: if (ISZERO(fp)) {
416: res[1] = res[2] = 0;
417: return (sign);
418: }
419:
420: if ((exp = fp->fp_exp + EXT_EXP_BIAS) < 0) {
421: fe->fe_fpsr |= FPSR_UNFL;
422: /* I'm not sure about this <=... exp==0 doesn't mean
423: it's a denormal in extended format */
424: (void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp);
425: if (fpu_round(fe, fp) && fp->fp_mant[1] == EXT_EXPLICIT1) {
426: res[1] = res[2] = 0;
427: return (sign | EXT_EXP(1) | 0);
428: }
429: if (fe->fe_fpsr & FPSR_INEX2)
430: fe->fe_fpsr |= FPSR_UNFL
431: /* mc68881/2 don't underflow */;
432: exp = 0;
433: goto done;
434: }
435: #if (FP_NMANT - FP_NG - EXT_FRACBITS) > 0
436: (void) fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS);
437: #endif
438: if (fpu_round(fe, fp) && fp->fp_mant[0] == EXT_EXPLICIT2)
439: exp++;
440: if (exp >= EXT_EXP_INFNAN) {
441: fe->fe_fpsr |= FPSR_OPERR | FPSR_INEX2 | FPSR_OVFL;
442: if (toinf(fe, sign)) {
443: res[1] = res[2] = 0;
444: return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0);
445: }
446: res[1] = res[2] = ~0;
447: return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK);
448: }
449: done:
450: res[1] = fp->fp_mant[1];
451: res[2] = fp->fp_mant[2];
452: return (sign | EXT_EXP(exp));
453: }
454:
455: /*
456: * Implode an fpn, writing the result into the given space.
457: */
458: void
459: fpu_implode(fe, fp, type, space)
460: struct fpemu *fe;
461: struct fpn *fp;
462: int type;
463: u_int *space;
464: {
465: /* XXX Dont delete exceptions set here: fe->fe_fpsr &= ~FPSR_EXCP; */
466:
467: switch (type) {
468: case FTYPE_LNG:
469: space[0] = fpu_ftoi(fe, fp);
470: break;
471:
472: case FTYPE_SNG:
473: space[0] = fpu_ftos(fe, fp);
474: break;
475:
476: case FTYPE_DBL:
477: space[0] = fpu_ftod(fe, fp, space);
478: break;
479:
480: case FTYPE_EXT:
481: /* funky rounding precision options ?? */
482: space[0] = fpu_ftox(fe, fp, space);
483: break;
484:
485: default:
486: panic("fpu_implode");
487: }
488: }
CVSweb