1 /* Software floating-point emulation. Common operations. 2 Copyright (C) 1997,1998,1999 Free Software Foundation, Inc. 3 This file is part of the GNU C Library. 4 Contributed by Richard Henderson (rth@cygnus.com), 5 Jakub Jelinek (jj@ultra.linux.cz), 6 David S. Miller (davem@redhat.com) and 7 Peter Maydell (pmaydell@chiark.greenend.org.uk). 8 9 The GNU C Library is free software; you can redistribute it and/or 10 modify it under the terms of the GNU Library General Public License as 11 published by the Free Software Foundation; either version 2 of the 12 License, or (at your option) any later version. 13 14 The GNU C Library is distributed in the hope that it will be useful, 15 but WITHOUT ANY WARRANTY; without even the implied warranty of 16 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 17 Library General Public License for more details. 18 19 You should have received a copy of the GNU Library General Public 20 License along with the GNU C Library; see the file COPYING.LIB. If 21 not, write to the Free Software Foundation, Inc., 22 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */ 23 24 #ifndef __MATH_EMU_OP_COMMON_H__ 25 #define __MATH_EMU_OP_COMMON_H__ 26 27 #define _FP_DECL(wc, X) \ 28 _FP_I_TYPE X##_c=0, X##_s=0, X##_e=0; \ 29 _FP_FRAC_DECL_##wc(X) 30 31 /* 32 * Finish truly unpacking a native fp value by classifying the kind 33 * of fp value and normalizing both the exponent and the fraction. 34 */ 35 36 #define _FP_UNPACK_CANONICAL(fs, wc, X) \ 37 do { \ 38 switch (X##_e) \ 39 { \ 40 default: \ 41 _FP_FRAC_HIGH_RAW_##fs(X) |= _FP_IMPLBIT_##fs; \ 42 _FP_FRAC_SLL_##wc(X, _FP_WORKBITS); \ 43 X##_e -= _FP_EXPBIAS_##fs; \ 44 X##_c = FP_CLS_NORMAL; \ 45 break; \ 46 \ 47 case 0: \ 48 if (_FP_FRAC_ZEROP_##wc(X)) \ 49 X##_c = FP_CLS_ZERO; \ 50 else \ 51 { \ 52 /* a denormalized number */ \ 53 _FP_I_TYPE _shift; \ 54 _FP_FRAC_CLZ_##wc(_shift, X); \ 55 _shift -= _FP_FRACXBITS_##fs; \ 56 _FP_FRAC_SLL_##wc(X, (_shift+_FP_WORKBITS)); \ 57 X##_e -= _FP_EXPBIAS_##fs - 1 + _shift; \ 58 X##_c = FP_CLS_NORMAL; \ 59 FP_SET_EXCEPTION(FP_EX_DENORM); \ 60 if (FP_DENORM_ZERO) \ 61 { \ 62 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 63 X##_c = FP_CLS_ZERO; \ 64 } \ 65 } \ 66 break; \ 67 \ 68 case _FP_EXPMAX_##fs: \ 69 if (_FP_FRAC_ZEROP_##wc(X)) \ 70 X##_c = FP_CLS_INF; \ 71 else \ 72 { \ 73 X##_c = FP_CLS_NAN; \ 74 /* Check for signaling NaN */ \ 75 if (!(_FP_FRAC_HIGH_RAW_##fs(X) & _FP_QNANBIT_##fs)) \ 76 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_SNAN); \ 77 } \ 78 break; \ 79 } \ 80 } while (0) 81 82 /* 83 * Before packing the bits back into the native fp result, take care 84 * of such mundane things as rounding and overflow. Also, for some 85 * kinds of fp values, the original parts may not have been fully 86 * extracted -- but that is ok, we can regenerate them now. 87 */ 88 89 #define _FP_PACK_CANONICAL(fs, wc, X) \ 90 do { \ 91 switch (X##_c) \ 92 { \ 93 case FP_CLS_NORMAL: \ 94 X##_e += _FP_EXPBIAS_##fs; \ 95 if (X##_e > 0) \ 96 { \ 97 _FP_ROUND(wc, X); \ 98 if (_FP_FRAC_OVERP_##wc(fs, X)) \ 99 { \ 100 _FP_FRAC_CLEAR_OVERP_##wc(fs, X); \ 101 X##_e++; \ 102 } \ 103 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS); \ 104 if (X##_e >= _FP_EXPMAX_##fs) \ 105 { \ 106 /* overflow */ \ 107 switch (FP_ROUNDMODE) \ 108 { \ 109 case FP_RND_NEAREST: \ 110 X##_c = FP_CLS_INF; \ 111 break; \ 112 case FP_RND_PINF: \ 113 if (!X##_s) X##_c = FP_CLS_INF; \ 114 break; \ 115 case FP_RND_MINF: \ 116 if (X##_s) X##_c = FP_CLS_INF; \ 117 break; \ 118 } \ 119 if (X##_c == FP_CLS_INF) \ 120 { \ 121 /* Overflow to infinity */ \ 122 X##_e = _FP_EXPMAX_##fs; \ 123 _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \ 124 } \ 125 else \ 126 { \ 127 /* Overflow to maximum normal */ \ 128 X##_e = _FP_EXPMAX_##fs - 1; \ 129 _FP_FRAC_SET_##wc(X, _FP_MAXFRAC_##wc); \ 130 } \ 131 FP_SET_EXCEPTION(FP_EX_OVERFLOW); \ 132 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 133 } \ 134 } \ 135 else \ 136 { \ 137 /* we've got a denormalized number */ \ 138 X##_e = -X##_e + 1; \ 139 if (X##_e <= _FP_WFRACBITS_##fs) \ 140 { \ 141 _FP_FRAC_SRS_##wc(X, X##_e, _FP_WFRACBITS_##fs); \ 142 if (_FP_FRAC_HIGH_##fs(X) \ 143 & (_FP_OVERFLOW_##fs >> 1)) \ 144 { \ 145 X##_e = 1; \ 146 _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \ 147 } \ 148 else \ 149 { \ 150 _FP_ROUND(wc, X); \ 151 if (_FP_FRAC_HIGH_##fs(X) \ 152 & (_FP_OVERFLOW_##fs >> 1)) \ 153 { \ 154 X##_e = 1; \ 155 _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \ 156 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 157 } \ 158 else \ 159 { \ 160 X##_e = 0; \ 161 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS); \ 162 } \ 163 } \ 164 if ((FP_CUR_EXCEPTIONS & FP_EX_INEXACT) || \ 165 (FP_TRAPPING_EXCEPTIONS & FP_EX_UNDERFLOW)) \ 166 FP_SET_EXCEPTION(FP_EX_UNDERFLOW); \ 167 } \ 168 else \ 169 { \ 170 /* underflow to zero */ \ 171 X##_e = 0; \ 172 if (!_FP_FRAC_ZEROP_##wc(X)) \ 173 { \ 174 _FP_FRAC_SET_##wc(X, _FP_MINFRAC_##wc); \ 175 _FP_ROUND(wc, X); \ 176 _FP_FRAC_LOW_##wc(X) >>= (_FP_WORKBITS); \ 177 } \ 178 FP_SET_EXCEPTION(FP_EX_UNDERFLOW); \ 179 } \ 180 } \ 181 break; \ 182 \ 183 case FP_CLS_ZERO: \ 184 X##_e = 0; \ 185 _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \ 186 break; \ 187 \ 188 case FP_CLS_INF: \ 189 X##_e = _FP_EXPMAX_##fs; \ 190 _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \ 191 break; \ 192 \ 193 case FP_CLS_NAN: \ 194 X##_e = _FP_EXPMAX_##fs; \ 195 if (!_FP_KEEPNANFRACP) \ 196 { \ 197 _FP_FRAC_SET_##wc(X, _FP_NANFRAC_##fs); \ 198 X##_s = _FP_NANSIGN_##fs; \ 199 } \ 200 else \ 201 _FP_FRAC_HIGH_RAW_##fs(X) |= _FP_QNANBIT_##fs; \ 202 break; \ 203 } \ 204 } while (0) 205 206 /* This one accepts raw argument and not cooked, returns 207 * 1 if X is a signaling NaN. 208 */ 209 #define _FP_ISSIGNAN(fs, wc, X) \ 210 ({ \ 211 int __ret = 0; \ 212 if (X##_e == _FP_EXPMAX_##fs) \ 213 { \ 214 if (!_FP_FRAC_ZEROP_##wc(X) \ 215 && !(_FP_FRAC_HIGH_RAW_##fs(X) & _FP_QNANBIT_##fs)) \ 216 __ret = 1; \ 217 } \ 218 __ret; \ 219 }) 220 221 222 223 224 225 /* 226 * Main addition routine. The input values should be cooked. 227 */ 228 229 #define _FP_ADD_INTERNAL(fs, wc, R, X, Y, OP) \ 230 do { \ 231 switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \ 232 { \ 233 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \ 234 { \ 235 /* shift the smaller number so that its exponent matches the larger */ \ 236 _FP_I_TYPE diff = X##_e - Y##_e; \ 237 \ 238 if (diff < 0) \ 239 { \ 240 diff = -diff; \ 241 if (diff <= _FP_WFRACBITS_##fs) \ 242 _FP_FRAC_SRS_##wc(X, diff, _FP_WFRACBITS_##fs); \ 243 else if (!_FP_FRAC_ZEROP_##wc(X)) \ 244 _FP_FRAC_SET_##wc(X, _FP_MINFRAC_##wc); \ 245 R##_e = Y##_e; \ 246 } \ 247 else \ 248 { \ 249 if (diff > 0) \ 250 { \ 251 if (diff <= _FP_WFRACBITS_##fs) \ 252 _FP_FRAC_SRS_##wc(Y, diff, _FP_WFRACBITS_##fs); \ 253 else if (!_FP_FRAC_ZEROP_##wc(Y)) \ 254 _FP_FRAC_SET_##wc(Y, _FP_MINFRAC_##wc); \ 255 } \ 256 R##_e = X##_e; \ 257 } \ 258 \ 259 R##_c = FP_CLS_NORMAL; \ 260 \ 261 if (X##_s == Y##_s) \ 262 { \ 263 R##_s = X##_s; \ 264 _FP_FRAC_ADD_##wc(R, X, Y); \ 265 if (_FP_FRAC_OVERP_##wc(fs, R)) \ 266 { \ 267 _FP_FRAC_SRS_##wc(R, 1, _FP_WFRACBITS_##fs); \ 268 R##_e++; \ 269 } \ 270 } \ 271 else \ 272 { \ 273 R##_s = X##_s; \ 274 _FP_FRAC_SUB_##wc(R, X, Y); \ 275 if (_FP_FRAC_ZEROP_##wc(R)) \ 276 { \ 277 /* return an exact zero */ \ 278 if (FP_ROUNDMODE == FP_RND_MINF) \ 279 R##_s |= Y##_s; \ 280 else \ 281 R##_s &= Y##_s; \ 282 R##_c = FP_CLS_ZERO; \ 283 } \ 284 else \ 285 { \ 286 if (_FP_FRAC_NEGP_##wc(R)) \ 287 { \ 288 _FP_FRAC_SUB_##wc(R, Y, X); \ 289 R##_s = Y##_s; \ 290 } \ 291 \ 292 /* renormalize after subtraction */ \ 293 _FP_FRAC_CLZ_##wc(diff, R); \ 294 diff -= _FP_WFRACXBITS_##fs; \ 295 if (diff) \ 296 { \ 297 R##_e -= diff; \ 298 _FP_FRAC_SLL_##wc(R, diff); \ 299 } \ 300 } \ 301 } \ 302 break; \ 303 } \ 304 \ 305 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \ 306 _FP_CHOOSENAN(fs, wc, R, X, Y, OP); \ 307 break; \ 308 \ 309 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \ 310 R##_e = X##_e; \ 311 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 312 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 313 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 314 _FP_FRAC_COPY_##wc(R, X); \ 315 R##_s = X##_s; \ 316 R##_c = X##_c; \ 317 break; \ 318 \ 319 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 320 R##_e = Y##_e; \ 321 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 322 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 323 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 324 _FP_FRAC_COPY_##wc(R, Y); \ 325 R##_s = Y##_s; \ 326 R##_c = Y##_c; \ 327 break; \ 328 \ 329 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 330 if (X##_s != Y##_s) \ 331 { \ 332 /* +INF + -INF => NAN */ \ 333 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 334 R##_s = _FP_NANSIGN_##fs; \ 335 R##_c = FP_CLS_NAN; \ 336 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_ISI); \ 337 break; \ 338 } \ 339 /* FALLTHRU */ \ 340 \ 341 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 342 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 343 R##_s = X##_s; \ 344 R##_c = FP_CLS_INF; \ 345 break; \ 346 \ 347 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 348 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 349 R##_s = Y##_s; \ 350 R##_c = FP_CLS_INF; \ 351 break; \ 352 \ 353 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 354 /* make sure the sign is correct */ \ 355 if (FP_ROUNDMODE == FP_RND_MINF) \ 356 R##_s = X##_s | Y##_s; \ 357 else \ 358 R##_s = X##_s & Y##_s; \ 359 R##_c = FP_CLS_ZERO; \ 360 break; \ 361 \ 362 default: \ 363 abort(); \ 364 } \ 365 } while (0) 366 367 #define _FP_ADD(fs, wc, R, X, Y) _FP_ADD_INTERNAL(fs, wc, R, X, Y, '+') 368 #define _FP_SUB(fs, wc, R, X, Y) \ 369 do { \ 370 if (Y##_c != FP_CLS_NAN) Y##_s ^= 1; \ 371 _FP_ADD_INTERNAL(fs, wc, R, X, Y, '-'); \ 372 } while (0) 373 374 375 /* 376 * Main negation routine. FIXME -- when we care about setting exception 377 * bits reliably, this will not do. We should examine all of the fp classes. 378 */ 379 380 #define _FP_NEG(fs, wc, R, X) \ 381 do { \ 382 _FP_FRAC_COPY_##wc(R, X); \ 383 R##_c = X##_c; \ 384 R##_e = X##_e; \ 385 R##_s = 1 ^ X##_s; \ 386 } while (0) 387 388 389 /* 390 * Main multiplication routine. The input values should be cooked. 391 */ 392 393 #define _FP_MUL(fs, wc, R, X, Y) \ 394 do { \ 395 R##_s = X##_s ^ Y##_s; \ 396 switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \ 397 { \ 398 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \ 399 R##_c = FP_CLS_NORMAL; \ 400 R##_e = X##_e + Y##_e + 1; \ 401 \ 402 _FP_MUL_MEAT_##fs(R,X,Y); \ 403 \ 404 if (_FP_FRAC_OVERP_##wc(fs, R)) \ 405 _FP_FRAC_SRS_##wc(R, 1, _FP_WFRACBITS_##fs); \ 406 else \ 407 R##_e--; \ 408 break; \ 409 \ 410 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \ 411 _FP_CHOOSENAN(fs, wc, R, X, Y, '*'); \ 412 break; \ 413 \ 414 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 415 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 416 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 417 R##_s = X##_s; \ 418 \ 419 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 420 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 421 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 422 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 423 _FP_FRAC_COPY_##wc(R, X); \ 424 R##_c = X##_c; \ 425 break; \ 426 \ 427 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 428 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 429 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 430 R##_s = Y##_s; \ 431 \ 432 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 433 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \ 434 _FP_FRAC_COPY_##wc(R, Y); \ 435 R##_c = Y##_c; \ 436 break; \ 437 \ 438 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 439 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 440 R##_s = _FP_NANSIGN_##fs; \ 441 R##_c = FP_CLS_NAN; \ 442 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 443 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_IMZ);\ 444 break; \ 445 \ 446 default: \ 447 abort(); \ 448 } \ 449 } while (0) 450 451 452 /* 453 * Main division routine. The input values should be cooked. 454 */ 455 456 #define _FP_DIV(fs, wc, R, X, Y) \ 457 do { \ 458 R##_s = X##_s ^ Y##_s; \ 459 switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \ 460 { \ 461 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \ 462 R##_c = FP_CLS_NORMAL; \ 463 R##_e = X##_e - Y##_e; \ 464 \ 465 _FP_DIV_MEAT_##fs(R,X,Y); \ 466 break; \ 467 \ 468 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \ 469 _FP_CHOOSENAN(fs, wc, R, X, Y, '/'); \ 470 break; \ 471 \ 472 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 473 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 474 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 475 R##_s = X##_s; \ 476 _FP_FRAC_COPY_##wc(R, X); \ 477 R##_c = X##_c; \ 478 break; \ 479 \ 480 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 481 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 482 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 483 R##_s = Y##_s; \ 484 _FP_FRAC_COPY_##wc(R, Y); \ 485 R##_c = Y##_c; \ 486 break; \ 487 \ 488 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 489 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 490 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 491 R##_c = FP_CLS_ZERO; \ 492 break; \ 493 \ 494 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \ 495 FP_SET_EXCEPTION(FP_EX_DIVZERO); \ 496 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 497 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 498 R##_c = FP_CLS_INF; \ 499 break; \ 500 \ 501 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 502 R##_s = _FP_NANSIGN_##fs; \ 503 R##_c = FP_CLS_NAN; \ 504 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 505 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_IDI);\ 506 break; \ 507 \ 508 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 509 R##_s = _FP_NANSIGN_##fs; \ 510 R##_c = FP_CLS_NAN; \ 511 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 512 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_ZDZ);\ 513 break; \ 514 \ 515 default: \ 516 abort(); \ 517 } \ 518 } while (0) 519 520 521 /* 522 * Main differential comparison routine. The inputs should be raw not 523 * cooked. The return is -1,0,1 for normal values, 2 otherwise. 524 */ 525 526 #define _FP_CMP(fs, wc, ret, X, Y, un) \ 527 do { \ 528 /* NANs are unordered */ \ 529 if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \ 530 || (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \ 531 { \ 532 ret = un; \ 533 } \ 534 else \ 535 { \ 536 int __is_zero_x; \ 537 int __is_zero_y; \ 538 \ 539 __is_zero_x = (!X##_e && _FP_FRAC_ZEROP_##wc(X)) ? 1 : 0; \ 540 __is_zero_y = (!Y##_e && _FP_FRAC_ZEROP_##wc(Y)) ? 1 : 0; \ 541 \ 542 if (__is_zero_x && __is_zero_y) \ 543 ret = 0; \ 544 else if (__is_zero_x) \ 545 ret = Y##_s ? 1 : -1; \ 546 else if (__is_zero_y) \ 547 ret = X##_s ? -1 : 1; \ 548 else if (X##_s != Y##_s) \ 549 ret = X##_s ? -1 : 1; \ 550 else if (X##_e > Y##_e) \ 551 ret = X##_s ? -1 : 1; \ 552 else if (X##_e < Y##_e) \ 553 ret = X##_s ? 1 : -1; \ 554 else if (_FP_FRAC_GT_##wc(X, Y)) \ 555 ret = X##_s ? -1 : 1; \ 556 else if (_FP_FRAC_GT_##wc(Y, X)) \ 557 ret = X##_s ? 1 : -1; \ 558 else \ 559 ret = 0; \ 560 } \ 561 } while (0) 562 563 564 /* Simplification for strict equality. */ 565 566 #define _FP_CMP_EQ(fs, wc, ret, X, Y) \ 567 do { \ 568 /* NANs are unordered */ \ 569 if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \ 570 || (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \ 571 { \ 572 ret = 1; \ 573 } \ 574 else \ 575 { \ 576 ret = !(X##_e == Y##_e \ 577 && _FP_FRAC_EQ_##wc(X, Y) \ 578 && (X##_s == Y##_s || !X##_e && _FP_FRAC_ZEROP_##wc(X))); \ 579 } \ 580 } while (0) 581 582 /* 583 * Main square root routine. The input value should be cooked. 584 */ 585 586 #define _FP_SQRT(fs, wc, R, X) \ 587 do { \ 588 _FP_FRAC_DECL_##wc(T); _FP_FRAC_DECL_##wc(S); \ 589 _FP_W_TYPE q; \ 590 switch (X##_c) \ 591 { \ 592 case FP_CLS_NAN: \ 593 _FP_FRAC_COPY_##wc(R, X); \ 594 R##_s = X##_s; \ 595 R##_c = FP_CLS_NAN; \ 596 break; \ 597 case FP_CLS_INF: \ 598 if (X##_s) \ 599 { \ 600 R##_s = _FP_NANSIGN_##fs; \ 601 R##_c = FP_CLS_NAN; /* NAN */ \ 602 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 603 FP_SET_EXCEPTION(FP_EX_INVALID); \ 604 } \ 605 else \ 606 { \ 607 R##_s = 0; \ 608 R##_c = FP_CLS_INF; /* sqrt(+inf) = +inf */ \ 609 } \ 610 break; \ 611 case FP_CLS_ZERO: \ 612 R##_s = X##_s; \ 613 R##_c = FP_CLS_ZERO; /* sqrt(+-0) = +-0 */ \ 614 break; \ 615 case FP_CLS_NORMAL: \ 616 R##_s = 0; \ 617 if (X##_s) \ 618 { \ 619 R##_c = FP_CLS_NAN; /* sNAN */ \ 620 R##_s = _FP_NANSIGN_##fs; \ 621 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 622 FP_SET_EXCEPTION(FP_EX_INVALID); \ 623 break; \ 624 } \ 625 R##_c = FP_CLS_NORMAL; \ 626 if (X##_e & 1) \ 627 _FP_FRAC_SLL_##wc(X, 1); \ 628 R##_e = X##_e >> 1; \ 629 _FP_FRAC_SET_##wc(S, _FP_ZEROFRAC_##wc); \ 630 _FP_FRAC_SET_##wc(R, _FP_ZEROFRAC_##wc); \ 631 q = _FP_OVERFLOW_##fs >> 1; \ 632 _FP_SQRT_MEAT_##wc(R, S, T, X, q); \ 633 } \ 634 } while (0) 635 636 /* 637 * Convert from FP to integer 638 */ 639 640 /* RSIGNED can have following values: 641 * 0: the number is required to be 0..(2^rsize)-1, if not, NV is set plus 642 * the result is either 0 or (2^rsize)-1 depending on the sign in such case. 643 * 1: the number is required to be -(2^(rsize-1))..(2^(rsize-1))-1, if not, NV is 644 * set plus the result is either -(2^(rsize-1)) or (2^(rsize-1))-1 depending 645 * on the sign in such case. 646 * 2: the number is required to be -(2^(rsize-1))..(2^(rsize-1))-1, if not, NV is 647 * set plus the result is truncated to fit into destination. 648 * -1: the number is required to be -(2^(rsize-1))..(2^rsize)-1, if not, NV is 649 * set plus the result is either -(2^(rsize-1)) or (2^(rsize-1))-1 depending 650 * on the sign in such case. 651 */ 652 #define _FP_TO_INT(fs, wc, r, X, rsize, rsigned) \ 653 do { \ 654 switch (X##_c) \ 655 { \ 656 case FP_CLS_NORMAL: \ 657 if (X##_e < 0) \ 658 { \ 659 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 660 case FP_CLS_ZERO: \ 661 r = 0; \ 662 } \ 663 else if (X##_e >= rsize - (rsigned > 0 || X##_s) \ 664 || (!rsigned && X##_s)) \ 665 { /* overflow */ \ 666 case FP_CLS_NAN: \ 667 case FP_CLS_INF: \ 668 if (rsigned == 2) \ 669 { \ 670 if (X##_c != FP_CLS_NORMAL \ 671 || X##_e >= rsize - 1 + _FP_WFRACBITS_##fs) \ 672 r = 0; \ 673 else \ 674 { \ 675 _FP_FRAC_SLL_##wc(X, (X##_e - _FP_WFRACBITS_##fs + 1)); \ 676 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 677 } \ 678 } \ 679 else if (rsigned) \ 680 { \ 681 r = 1; \ 682 r <<= rsize - 1; \ 683 r -= 1 - X##_s; \ 684 } \ 685 else \ 686 { \ 687 r = 0; \ 688 if (!X##_s) \ 689 r = ~r; \ 690 } \ 691 FP_SET_EXCEPTION(FP_EX_INVALID); \ 692 } \ 693 else \ 694 { \ 695 if (_FP_W_TYPE_SIZE*wc < rsize) \ 696 { \ 697 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 698 r <<= X##_e - _FP_WFRACBITS_##fs; \ 699 } \ 700 else \ 701 { \ 702 if (X##_e >= _FP_WFRACBITS_##fs) \ 703 _FP_FRAC_SLL_##wc(X, (X##_e - _FP_WFRACBITS_##fs + 1)); \ 704 else if (X##_e < _FP_WFRACBITS_##fs - 1) \ 705 { \ 706 _FP_FRAC_SRS_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 2), \ 707 _FP_WFRACBITS_##fs); \ 708 if (_FP_FRAC_LOW_##wc(X) & 1) \ 709 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 710 _FP_FRAC_SRL_##wc(X, 1); \ 711 } \ 712 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 713 } \ 714 if (rsigned && X##_s) \ 715 r = -r; \ 716 } \ 717 break; \ 718 } \ 719 } while (0) 720 721 #define _FP_TO_INT_ROUND(fs, wc, r, X, rsize, rsigned) \ 722 do { \ 723 r = 0; \ 724 switch (X##_c) \ 725 { \ 726 case FP_CLS_NORMAL: \ 727 if (X##_e >= _FP_FRACBITS_##fs - 1) \ 728 { \ 729 if (X##_e < rsize - 1 + _FP_WFRACBITS_##fs) \ 730 { \ 731 if (X##_e >= _FP_WFRACBITS_##fs - 1) \ 732 { \ 733 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 734 r <<= X##_e - _FP_WFRACBITS_##fs + 1; \ 735 } \ 736 else \ 737 { \ 738 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS - X##_e \ 739 + _FP_FRACBITS_##fs - 1); \ 740 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 741 } \ 742 } \ 743 } \ 744 else \ 745 { \ 746 int _lz0, _lz1; \ 747 if (X##_e <= -_FP_WORKBITS - 1) \ 748 _FP_FRAC_SET_##wc(X, _FP_MINFRAC_##wc); \ 749 else \ 750 _FP_FRAC_SRS_##wc(X, _FP_FRACBITS_##fs - 1 - X##_e, \ 751 _FP_WFRACBITS_##fs); \ 752 _FP_FRAC_CLZ_##wc(_lz0, X); \ 753 _FP_ROUND(wc, X); \ 754 _FP_FRAC_CLZ_##wc(_lz1, X); \ 755 if (_lz1 < _lz0) \ 756 X##_e++; /* For overflow detection. */ \ 757 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS); \ 758 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 759 } \ 760 if (rsigned && X##_s) \ 761 r = -r; \ 762 if (X##_e >= rsize - (rsigned > 0 || X##_s) \ 763 || (!rsigned && X##_s)) \ 764 { /* overflow */ \ 765 case FP_CLS_NAN: \ 766 case FP_CLS_INF: \ 767 if (!rsigned) \ 768 { \ 769 r = 0; \ 770 if (!X##_s) \ 771 r = ~r; \ 772 } \ 773 else if (rsigned != 2) \ 774 { \ 775 r = 1; \ 776 r <<= rsize - 1; \ 777 r -= 1 - X##_s; \ 778 } \ 779 FP_SET_EXCEPTION(FP_EX_INVALID); \ 780 } \ 781 break; \ 782 case FP_CLS_ZERO: \ 783 break; \ 784 } \ 785 } while (0) 786 787 #define _FP_FROM_INT(fs, wc, X, r, rsize, rtype) \ 788 do { \ 789 if (r) \ 790 { \ 791 unsigned rtype ur_; \ 792 X##_c = FP_CLS_NORMAL; \ 793 \ 794 if ((X##_s = (r < 0))) \ 795 ur_ = (unsigned rtype) -r; \ 796 else \ 797 ur_ = (unsigned rtype) r; \ 798 (void) (((rsize) <= _FP_W_TYPE_SIZE) \ 799 ? ({ __FP_CLZ(X##_e, ur_); }) \ 800 : ({ \ 801 __FP_CLZ_2(X##_e, (_FP_W_TYPE)(ur_ >> _FP_W_TYPE_SIZE), \ 802 (_FP_W_TYPE)ur_); \ 803 })); \ 804 if (rsize < _FP_W_TYPE_SIZE) \ 805 X##_e -= (_FP_W_TYPE_SIZE - rsize); \ 806 X##_e = rsize - X##_e - 1; \ 807 \ 808 if (_FP_FRACBITS_##fs < rsize && _FP_WFRACBITS_##fs <= X##_e) \ 809 __FP_FRAC_SRS_1(ur_, (X##_e - _FP_WFRACBITS_##fs + 1), rsize);\ 810 _FP_FRAC_DISASSEMBLE_##wc(X, ur_, rsize); \ 811 if ((_FP_WFRACBITS_##fs - X##_e - 1) > 0) \ 812 _FP_FRAC_SLL_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 1)); \ 813 } \ 814 else \ 815 { \ 816 X##_c = FP_CLS_ZERO, X##_s = 0; \ 817 } \ 818 } while (0) 819 820 821 #define FP_CONV(dfs,sfs,dwc,swc,D,S) \ 822 do { \ 823 _FP_FRAC_CONV_##dwc##_##swc(dfs, sfs, D, S); \ 824 D##_e = S##_e; \ 825 D##_c = S##_c; \ 826 D##_s = S##_s; \ 827 } while (0) 828 829 /* 830 * Helper primitives. 831 */ 832 833 /* Count leading zeros in a word. */ 834 835 #ifndef __FP_CLZ 836 #if _FP_W_TYPE_SIZE < 64 837 /* this is just to shut the compiler up about shifts > word length -- PMM 02/1998 */ 838 #define __FP_CLZ(r, x) \ 839 do { \ 840 _FP_W_TYPE _t = (x); \ 841 r = _FP_W_TYPE_SIZE - 1; \ 842 if (_t > 0xffff) r -= 16; \ 843 if (_t > 0xffff) _t >>= 16; \ 844 if (_t > 0xff) r -= 8; \ 845 if (_t > 0xff) _t >>= 8; \ 846 if (_t & 0xf0) r -= 4; \ 847 if (_t & 0xf0) _t >>= 4; \ 848 if (_t & 0xc) r -= 2; \ 849 if (_t & 0xc) _t >>= 2; \ 850 if (_t & 0x2) r -= 1; \ 851 } while (0) 852 #else /* not _FP_W_TYPE_SIZE < 64 */ 853 #define __FP_CLZ(r, x) \ 854 do { \ 855 _FP_W_TYPE _t = (x); \ 856 r = _FP_W_TYPE_SIZE - 1; \ 857 if (_t > 0xffffffff) r -= 32; \ 858 if (_t > 0xffffffff) _t >>= 32; \ 859 if (_t > 0xffff) r -= 16; \ 860 if (_t > 0xffff) _t >>= 16; \ 861 if (_t > 0xff) r -= 8; \ 862 if (_t > 0xff) _t >>= 8; \ 863 if (_t & 0xf0) r -= 4; \ 864 if (_t & 0xf0) _t >>= 4; \ 865 if (_t & 0xc) r -= 2; \ 866 if (_t & 0xc) _t >>= 2; \ 867 if (_t & 0x2) r -= 1; \ 868 } while (0) 869 #endif /* not _FP_W_TYPE_SIZE < 64 */ 870 #endif /* ndef __FP_CLZ */ 871 872 #define _FP_DIV_HELP_imm(q, r, n, d) \ 873 do { \ 874 q = n / d, r = n % d; \ 875 } while (0) 876 877 #endif /* __MATH_EMU_OP_COMMON_H__ */ 878