1/* 2 * Copyright (c) 1992, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This software was developed by the Computer Systems Engineering group 6 * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and 7 * contributed to Berkeley. 8 * 9 * All advertising materials mentioning features or use of this software 10 * must display the following acknowledgement: 11 * This product includes software developed by the University of 12 * California, Lawrence Berkeley Laboratory. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. 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 * $NetBSD: fpu_implode.c,v 1.8 2001/08/26 05:44:46 eeh Exp $ 44 */ 45 46#include <sys/cdefs.h>
|
47__FBSDID("$FreeBSD: head/lib/libc/sparc64/fpu/fpu_implode.c 92986 2002-03-22 21:53:29Z obrien $");
|
47__FBSDID("$FreeBSD: head/lib/libc/sparc64/fpu/fpu_implode.c 95587 2002-04-27 21:56:28Z jake $"); |
48 49/* 50 * FPU subroutines: `implode' internal format numbers into the machine's 51 * `packed binary' format. 52 */ 53 54#include <sys/param.h> 55 56#include <machine/frame.h> 57#include <machine/fp.h> 58#include <machine/fsr.h> 59#include <machine/ieee.h> 60#include <machine/instr.h> 61 62#include "fpu_arith.h" 63#include "fpu_emu.h" 64#include "fpu_extern.h"
|
65#include "__sparc_utrap_private.h" |
66 67static int round(struct fpemu *, struct fpn *); 68static int toinf(struct fpemu *, int); 69 70/* 71 * Round a number (algorithm from Motorola MC68882 manual, modified for 72 * our internal format). Set inexact exception if rounding is required. 73 * Return true iff we rounded up. 74 * 75 * After rounding, we discard the guard and round bits by shifting right 76 * 2 bits (a la fpu_shr(), but we do not bother with fp->fp_sticky). 77 * This saves effort later. 78 * 79 * Note that we may leave the value 2.0 in fp->fp_mant; it is the caller's 80 * responsibility to fix this if necessary. 81 */ 82static int 83round(struct fpemu *fe, struct fpn *fp) 84{ 85 u_int m0, m1, m2, m3; 86 int gr, s; 87 88 m0 = fp->fp_mant[0]; 89 m1 = fp->fp_mant[1]; 90 m2 = fp->fp_mant[2]; 91 m3 = fp->fp_mant[3]; 92 gr = m3 & 3; 93 s = fp->fp_sticky; 94 95 /* mant >>= FP_NG */ 96 m3 = (m3 >> FP_NG) | (m2 << (32 - FP_NG)); 97 m2 = (m2 >> FP_NG) | (m1 << (32 - FP_NG)); 98 m1 = (m1 >> FP_NG) | (m0 << (32 - FP_NG)); 99 m0 >>= FP_NG; 100 101 if ((gr | s) == 0) /* result is exact: no rounding needed */ 102 goto rounddown; 103 104 fe->fe_cx |= FSR_NX; /* inexact */ 105 106 /* Go to rounddown to round down; break to round up. */ 107 switch (FSR_GET_RD(fe->fe_fsr)) { 108 case FSR_RD_N: 109 default: 110 /* 111 * Round only if guard is set (gr & 2). If guard is set, 112 * but round & sticky both clear, then we want to round 113 * but have a tie, so round to even, i.e., add 1 iff odd. 114 */ 115 if ((gr & 2) == 0) 116 goto rounddown; 117 if ((gr & 1) || fp->fp_sticky || (m3 & 1)) 118 break; 119 goto rounddown; 120 121 case FSR_RD_Z: 122 /* Round towards zero, i.e., down. */ 123 goto rounddown; 124 125 case FSR_RD_NINF: 126 /* Round towards -Inf: up if negative, down if positive. */ 127 if (fp->fp_sign) 128 break; 129 goto rounddown; 130 131 case FSR_RD_PINF: 132 /* Round towards +Inf: up if positive, down otherwise. */ 133 if (!fp->fp_sign) 134 break; 135 goto rounddown; 136 } 137 138 /* Bump low bit of mantissa, with carry. */ 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 149rounddown: 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 */ 162static int 163toinf(struct fpemu *fe, int sign) 164{ 165 int inf; 166 167 /* look at rounding direction */ 168 switch (FSR_GET_RD(fe->fe_fsr)) { 169 default: 170 case FSR_RD_N: /* the nearest value is always Inf */ 171 inf = 1; 172 break; 173 174 case FSR_RD_Z: /* toward 0 => never towards Inf */ 175 inf = 0; 176 break; 177 178 case FSR_RD_PINF: /* toward +Inf iff positive */ 179 inf = sign == 0; 180 break; 181 182 case FSR_RD_NINF: /* toward -Inf iff negative */ 183 inf = sign; 184 break; 185 } 186 return (inf); 187} 188 189/* 190 * fpn -> int (int value returned as return value). 191 * 192 * N.B.: this conversion always rounds towards zero (this is a peculiarity 193 * of the SPARC instruction set). 194 */ 195u_int 196__fpu_ftoi(fe, fp) 197 struct fpemu *fe; 198 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 |= FSR_NX; 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 = (fe->fe_cx & ~FSR_NX) | FSR_NV; 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 */ 244u_int 245__fpu_ftox(fe, fp, res) 246 struct fpemu *fe; 247 struct fpn *fp; 248 u_int *res; 249{ 250 u_int64_t i; 251 int sign, exp; 252 253 sign = fp->fp_sign; 254 switch (fp->fp_class) { 255 256 case FPC_ZERO: 257 res[1] = 0; 258 return (0); 259 260 case FPC_NUM: 261 /* 262 * If exp >= 2^64, overflow. Otherwise shift value right 263 * into last mantissa word (this will not exceed 0xffffffffffffffff), 264 * shifting any guard and round bits out into the sticky 265 * bit. Then ``round'' towards zero, i.e., just set an 266 * inexact exception if sticky is set (see round()). 267 * If the result is > 0x8000000000000000, or is positive and equals 268 * 0x8000000000000000, overflow; otherwise the last fraction word 269 * is the result. 270 */ 271 if ((exp = fp->fp_exp) >= 64) 272 break; 273 /* NB: the following includes exp < 0 cases */ 274 if (__fpu_shr(fp, FP_NMANT - 1 - exp) != 0) 275 fe->fe_cx |= FSR_NX; 276 i = ((u_int64_t)fp->fp_mant[2]<<32)|fp->fp_mant[3]; 277 if (i >= ((u_int64_t)0x8000000000000000LL + sign)) 278 break; 279 if (sign) 280 i = -1; 281 res[1] = (int)i; 282 return (i >> 32); 283 284 default: /* Inf, qNaN, sNaN */ 285 break; 286 } 287 /* overflow: replace any inexact exception with invalid */ 288 fe->fe_cx = (fe->fe_cx & ~FSR_NX) | FSR_NV; 289 return (0x7fffffffffffffffLL + sign); 290} 291 292/* 293 * fpn -> single (32 bit single returned as return value). 294 * We assume <= 29 bits in a single-precision fraction (1.f part). 295 */ 296u_int 297__fpu_ftos(fe, fp) 298 struct fpemu *fe; 299 struct fpn *fp; 300{ 301 u_int sign = fp->fp_sign << 31; 302 int exp; 303 304#define SNG_EXP(e) ((e) << SNG_FRACBITS) /* makes e an exponent */ 305#define SNG_MASK (SNG_EXP(1) - 1) /* mask for fraction */ 306 307 /* Take care of non-numbers first. */ 308 if (ISNAN(fp)) { 309 /* 310 * Preserve upper bits of NaN, per SPARC V8 appendix N. 311 * Note that fp->fp_mant[0] has the quiet bit set, 312 * even if it is classified as a signalling NaN. 313 */ 314 (void) __fpu_shr(fp, FP_NMANT - 1 - SNG_FRACBITS); 315 exp = SNG_EXP_INFNAN; 316 goto done; 317 } 318 if (ISINF(fp)) 319 return (sign | SNG_EXP(SNG_EXP_INFNAN)); 320 if (ISZERO(fp)) 321 return (sign); 322 323 /* 324 * Normals (including subnormals). Drop all the fraction bits 325 * (including the explicit ``implied'' 1 bit) down into the 326 * single-precision range. If the number is subnormal, move 327 * the ``implied'' 1 into the explicit range as well, and shift 328 * right to introduce leading zeroes. Rounding then acts 329 * differently for normals and subnormals: the largest subnormal 330 * may round to the smallest normal (1.0 x 2^minexp), or may 331 * remain subnormal. In the latter case, signal an underflow 332 * if the result was inexact or if underflow traps are enabled. 333 * 334 * Rounding a normal, on the other hand, always produces another 335 * normal (although either way the result might be too big for 336 * single precision, and cause an overflow). If rounding a 337 * normal produces 2.0 in the fraction, we need not adjust that 338 * fraction at all, since both 1.0 and 2.0 are zero under the 339 * fraction mask. 340 * 341 * Note that the guard and round bits vanish from the number after 342 * rounding. 343 */ 344 if ((exp = fp->fp_exp + SNG_EXP_BIAS) <= 0) { /* subnormal */ 345 /* -NG for g,r; -SNG_FRACBITS-exp for fraction */ 346 (void) __fpu_shr(fp, FP_NMANT - FP_NG - SNG_FRACBITS - exp); 347 if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(1)) 348 return (sign | SNG_EXP(1) | 0); 349 if ((fe->fe_cx & FSR_NX) || 350 (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT))) 351 fe->fe_cx |= FSR_UF; 352 return (sign | SNG_EXP(0) | fp->fp_mant[3]); 353 } 354 /* -FP_NG for g,r; -1 for implied 1; -SNG_FRACBITS for fraction */ 355 (void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - SNG_FRACBITS); 356#ifdef DIAGNOSTIC 357 if ((fp->fp_mant[3] & SNG_EXP(1 << FP_NG)) == 0)
|
357 __fpu_panic("fpu_ftos");
|
358 __utrap_panic("fpu_ftos"); |
359#endif 360 if (round(fe, fp) && fp->fp_mant[3] == SNG_EXP(2)) 361 exp++; 362 if (exp >= SNG_EXP_INFNAN) { 363 /* overflow to inf or to max single */ 364 fe->fe_cx |= FSR_OF | FSR_NX; 365 if (toinf(fe, sign)) 366 return (sign | SNG_EXP(SNG_EXP_INFNAN)); 367 return (sign | SNG_EXP(SNG_EXP_INFNAN - 1) | SNG_MASK); 368 } 369done: 370 /* phew, made it */ 371 return (sign | SNG_EXP(exp) | (fp->fp_mant[3] & SNG_MASK)); 372} 373 374/* 375 * fpn -> double (32 bit high-order result returned; 32-bit low order result 376 * left in res[1]). Assumes <= 61 bits in double precision fraction. 377 * 378 * This code mimics fpu_ftos; see it for comments. 379 */ 380u_int 381__fpu_ftod(fe, fp, res) 382 struct fpemu *fe; 383 struct fpn *fp; 384 u_int *res; 385{ 386 u_int sign = fp->fp_sign << 31; 387 int exp; 388 389#define DBL_EXP(e) ((e) << (DBL_FRACBITS & 31)) 390#define DBL_MASK (DBL_EXP(1) - 1) 391 392 if (ISNAN(fp)) { 393 (void) __fpu_shr(fp, FP_NMANT - 1 - DBL_FRACBITS); 394 exp = DBL_EXP_INFNAN; 395 goto done; 396 } 397 if (ISINF(fp)) { 398 sign |= DBL_EXP(DBL_EXP_INFNAN); 399 goto zero; 400 } 401 if (ISZERO(fp)) { 402zero: res[1] = 0; 403 return (sign); 404 } 405 406 if ((exp = fp->fp_exp + DBL_EXP_BIAS) <= 0) { 407 (void) __fpu_shr(fp, FP_NMANT - FP_NG - DBL_FRACBITS - exp); 408 if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(1)) { 409 res[1] = 0; 410 return (sign | DBL_EXP(1) | 0); 411 } 412 if ((fe->fe_cx & FSR_NX) || 413 (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT))) 414 fe->fe_cx |= FSR_UF; 415 exp = 0; 416 goto done; 417 } 418 (void) __fpu_shr(fp, FP_NMANT - FP_NG - 1 - DBL_FRACBITS); 419 if (round(fe, fp) && fp->fp_mant[2] == DBL_EXP(2)) 420 exp++; 421 if (exp >= DBL_EXP_INFNAN) { 422 fe->fe_cx |= FSR_OF | FSR_NX; 423 if (toinf(fe, sign)) { 424 res[1] = 0; 425 return (sign | DBL_EXP(DBL_EXP_INFNAN) | 0); 426 } 427 res[1] = ~0; 428 return (sign | DBL_EXP(DBL_EXP_INFNAN) | DBL_MASK); 429 } 430done: 431 res[1] = fp->fp_mant[3]; 432 return (sign | DBL_EXP(exp) | (fp->fp_mant[2] & DBL_MASK)); 433} 434 435/* 436 * fpn -> extended (32 bit high-order result returned; low-order fraction 437 * words left in res[1]..res[3]). Like ftod, which is like ftos ... but 438 * our internal format *is* extended precision, plus 2 bits for guard/round, 439 * so we can avoid a small bit of work. 440 */ 441u_int 442__fpu_ftoq(fe, fp, res) 443 struct fpemu *fe; 444 struct fpn *fp; 445 u_int *res; 446{ 447 u_int sign = fp->fp_sign << 31; 448 int exp; 449 450#define EXT_EXP(e) ((e) << (EXT_FRACBITS & 31)) 451#define EXT_MASK (EXT_EXP(1) - 1) 452 453 if (ISNAN(fp)) { 454 (void) __fpu_shr(fp, 2); /* since we are not rounding */ 455 exp = EXT_EXP_INFNAN; 456 goto done; 457 } 458 if (ISINF(fp)) { 459 sign |= EXT_EXP(EXT_EXP_INFNAN); 460 goto zero; 461 } 462 if (ISZERO(fp)) { 463zero: res[1] = res[2] = res[3] = 0; 464 return (sign); 465 } 466 467 if ((exp = fp->fp_exp + EXT_EXP_BIAS) <= 0) { 468 (void) __fpu_shr(fp, FP_NMANT - FP_NG - EXT_FRACBITS - exp); 469 if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(1)) { 470 res[1] = res[2] = res[3] = 0; 471 return (sign | EXT_EXP(1) | 0); 472 } 473 if ((fe->fe_cx & FSR_NX) || 474 (fe->fe_fsr & (FSR_UF << FSR_TEM_SHIFT))) 475 fe->fe_cx |= FSR_UF; 476 exp = 0; 477 goto done; 478 } 479 /* Since internal == extended, no need to shift here. */ 480 if (round(fe, fp) && fp->fp_mant[0] == EXT_EXP(2)) 481 exp++; 482 if (exp >= EXT_EXP_INFNAN) { 483 fe->fe_cx |= FSR_OF | FSR_NX; 484 if (toinf(fe, sign)) { 485 res[1] = res[2] = res[3] = 0; 486 return (sign | EXT_EXP(EXT_EXP_INFNAN) | 0); 487 } 488 res[1] = res[2] = res[3] = ~0; 489 return (sign | EXT_EXP(EXT_EXP_INFNAN) | EXT_MASK); 490 } 491done: 492 res[1] = fp->fp_mant[1]; 493 res[2] = fp->fp_mant[2]; 494 res[3] = fp->fp_mant[3]; 495 return (sign | EXT_EXP(exp) | (fp->fp_mant[0] & EXT_MASK)); 496} 497 498/* 499 * Implode an fpn, writing the result into the given space. 500 */ 501void 502__fpu_implode(fe, fp, type, space) 503 struct fpemu *fe; 504 struct fpn *fp; 505 int type; 506 u_int *space; 507{ 508 509 switch (type) { 510 511 case FTYPE_LNG: 512 space[0] = __fpu_ftox(fe, fp, space); 513 break; 514 515 case FTYPE_INT: 516 space[0] = __fpu_ftoi(fe, fp); 517 break; 518 519 case FTYPE_SNG: 520 space[0] = __fpu_ftos(fe, fp); 521 break; 522 523 case FTYPE_DBL: 524 space[0] = __fpu_ftod(fe, fp, space); 525 break; 526 527 case FTYPE_EXT: 528 /* funky rounding precision options ?? */ 529 space[0] = __fpu_ftoq(fe, fp, space); 530 break; 531 532 default:
|
532 __fpu_panic("fpu_implode");
|
533 __utrap_panic("fpu_implode"); |
534 } 535 DPRINTF(FPE_REG, ("fpu_implode: %x %x %x %x\n", 536 space[0], space[1], space[2], space[3])); 537}
|