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 /* Fall through */ \ 312 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 313 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 314 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 315 _FP_FRAC_COPY_##wc(R, X); \ 316 R##_s = X##_s; \ 317 R##_c = X##_c; \ 318 break; \ 319 \ 320 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 321 R##_e = Y##_e; \ 322 /* Fall through */ \ 323 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 324 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 325 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 326 _FP_FRAC_COPY_##wc(R, Y); \ 327 R##_s = Y##_s; \ 328 R##_c = Y##_c; \ 329 break; \ 330 \ 331 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 332 if (X##_s != Y##_s) \ 333 { \ 334 /* +INF + -INF => NAN */ \ 335 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 336 R##_s = _FP_NANSIGN_##fs; \ 337 R##_c = FP_CLS_NAN; \ 338 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_ISI); \ 339 break; \ 340 } \ 341 /* FALLTHRU */ \ 342 \ 343 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 344 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 345 R##_s = X##_s; \ 346 R##_c = FP_CLS_INF; \ 347 break; \ 348 \ 349 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 350 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 351 R##_s = Y##_s; \ 352 R##_c = FP_CLS_INF; \ 353 break; \ 354 \ 355 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 356 /* make sure the sign is correct */ \ 357 if (FP_ROUNDMODE == FP_RND_MINF) \ 358 R##_s = X##_s | Y##_s; \ 359 else \ 360 R##_s = X##_s & Y##_s; \ 361 R##_c = FP_CLS_ZERO; \ 362 break; \ 363 \ 364 default: \ 365 abort(); \ 366 } \ 367 } while (0) 368 369 #define _FP_ADD(fs, wc, R, X, Y) _FP_ADD_INTERNAL(fs, wc, R, X, Y, '+') 370 #define _FP_SUB(fs, wc, R, X, Y) \ 371 do { \ 372 if (Y##_c != FP_CLS_NAN) Y##_s ^= 1; \ 373 _FP_ADD_INTERNAL(fs, wc, R, X, Y, '-'); \ 374 } while (0) 375 376 377 /* 378 * Main negation routine. FIXME -- when we care about setting exception 379 * bits reliably, this will not do. We should examine all of the fp classes. 380 */ 381 382 #define _FP_NEG(fs, wc, R, X) \ 383 do { \ 384 _FP_FRAC_COPY_##wc(R, X); \ 385 R##_c = X##_c; \ 386 R##_e = X##_e; \ 387 R##_s = 1 ^ X##_s; \ 388 } while (0) 389 390 391 /* 392 * Main multiplication routine. The input values should be cooked. 393 */ 394 395 #define _FP_MUL(fs, wc, R, X, Y) \ 396 do { \ 397 R##_s = X##_s ^ Y##_s; \ 398 switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \ 399 { \ 400 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \ 401 R##_c = FP_CLS_NORMAL; \ 402 R##_e = X##_e + Y##_e + 1; \ 403 \ 404 _FP_MUL_MEAT_##fs(R,X,Y); \ 405 \ 406 if (_FP_FRAC_OVERP_##wc(fs, R)) \ 407 _FP_FRAC_SRS_##wc(R, 1, _FP_WFRACBITS_##fs); \ 408 else \ 409 R##_e--; \ 410 break; \ 411 \ 412 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \ 413 _FP_CHOOSENAN(fs, wc, R, X, Y, '*'); \ 414 break; \ 415 \ 416 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 417 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 418 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 419 R##_s = X##_s; \ 420 /* Fall through */ \ 421 \ 422 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 423 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 424 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 425 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 426 _FP_FRAC_COPY_##wc(R, X); \ 427 R##_c = X##_c; \ 428 break; \ 429 \ 430 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 431 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 432 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 433 R##_s = Y##_s; \ 434 /* Fall through */ \ 435 \ 436 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 437 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \ 438 _FP_FRAC_COPY_##wc(R, Y); \ 439 R##_c = Y##_c; \ 440 break; \ 441 \ 442 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 443 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 444 R##_s = _FP_NANSIGN_##fs; \ 445 R##_c = FP_CLS_NAN; \ 446 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 447 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_IMZ);\ 448 break; \ 449 \ 450 default: \ 451 abort(); \ 452 } \ 453 } while (0) 454 455 456 /* 457 * Main division routine. The input values should be cooked. 458 */ 459 460 #define _FP_DIV(fs, wc, R, X, Y) \ 461 do { \ 462 R##_s = X##_s ^ Y##_s; \ 463 switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \ 464 { \ 465 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \ 466 R##_c = FP_CLS_NORMAL; \ 467 R##_e = X##_e - Y##_e; \ 468 \ 469 _FP_DIV_MEAT_##fs(R,X,Y); \ 470 break; \ 471 \ 472 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \ 473 _FP_CHOOSENAN(fs, wc, R, X, Y, '/'); \ 474 break; \ 475 \ 476 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \ 477 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \ 478 case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \ 479 R##_s = X##_s; \ 480 _FP_FRAC_COPY_##wc(R, X); \ 481 R##_c = X##_c; \ 482 break; \ 483 \ 484 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \ 485 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \ 486 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \ 487 R##_s = Y##_s; \ 488 _FP_FRAC_COPY_##wc(R, Y); \ 489 R##_c = Y##_c; \ 490 break; \ 491 \ 492 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \ 493 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \ 494 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \ 495 R##_c = FP_CLS_ZERO; \ 496 break; \ 497 \ 498 case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \ 499 FP_SET_EXCEPTION(FP_EX_DIVZERO); \ 500 /* Fall through */ \ 501 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \ 502 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \ 503 R##_c = FP_CLS_INF; \ 504 break; \ 505 \ 506 case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \ 507 R##_s = _FP_NANSIGN_##fs; \ 508 R##_c = FP_CLS_NAN; \ 509 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 510 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_IDI);\ 511 break; \ 512 \ 513 case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \ 514 R##_s = _FP_NANSIGN_##fs; \ 515 R##_c = FP_CLS_NAN; \ 516 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 517 FP_SET_EXCEPTION(FP_EX_INVALID | FP_EX_INVALID_ZDZ);\ 518 break; \ 519 \ 520 default: \ 521 abort(); \ 522 } \ 523 } while (0) 524 525 526 /* 527 * Main differential comparison routine. The inputs should be raw not 528 * cooked. The return is -1,0,1 for normal values, 2 otherwise. 529 */ 530 531 #define _FP_CMP(fs, wc, ret, X, Y, un) \ 532 do { \ 533 /* NANs are unordered */ \ 534 if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \ 535 || (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \ 536 { \ 537 ret = un; \ 538 } \ 539 else \ 540 { \ 541 int __is_zero_x; \ 542 int __is_zero_y; \ 543 \ 544 __is_zero_x = (!X##_e && _FP_FRAC_ZEROP_##wc(X)) ? 1 : 0; \ 545 __is_zero_y = (!Y##_e && _FP_FRAC_ZEROP_##wc(Y)) ? 1 : 0; \ 546 \ 547 if (__is_zero_x && __is_zero_y) \ 548 ret = 0; \ 549 else if (__is_zero_x) \ 550 ret = Y##_s ? 1 : -1; \ 551 else if (__is_zero_y) \ 552 ret = X##_s ? -1 : 1; \ 553 else if (X##_s != Y##_s) \ 554 ret = X##_s ? -1 : 1; \ 555 else if (X##_e > Y##_e) \ 556 ret = X##_s ? -1 : 1; \ 557 else if (X##_e < Y##_e) \ 558 ret = X##_s ? 1 : -1; \ 559 else if (_FP_FRAC_GT_##wc(X, Y)) \ 560 ret = X##_s ? -1 : 1; \ 561 else if (_FP_FRAC_GT_##wc(Y, X)) \ 562 ret = X##_s ? 1 : -1; \ 563 else \ 564 ret = 0; \ 565 } \ 566 } while (0) 567 568 569 /* Simplification for strict equality. */ 570 571 #define _FP_CMP_EQ(fs, wc, ret, X, Y) \ 572 do { \ 573 /* NANs are unordered */ \ 574 if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \ 575 || (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \ 576 { \ 577 ret = 1; \ 578 } \ 579 else \ 580 { \ 581 ret = !(X##_e == Y##_e \ 582 && _FP_FRAC_EQ_##wc(X, Y) \ 583 && (X##_s == Y##_s || !X##_e && _FP_FRAC_ZEROP_##wc(X))); \ 584 } \ 585 } while (0) 586 587 /* 588 * Main square root routine. The input value should be cooked. 589 */ 590 591 #define _FP_SQRT(fs, wc, R, X) \ 592 do { \ 593 _FP_FRAC_DECL_##wc(T); _FP_FRAC_DECL_##wc(S); \ 594 _FP_W_TYPE q; \ 595 switch (X##_c) \ 596 { \ 597 case FP_CLS_NAN: \ 598 _FP_FRAC_COPY_##wc(R, X); \ 599 R##_s = X##_s; \ 600 R##_c = FP_CLS_NAN; \ 601 break; \ 602 case FP_CLS_INF: \ 603 if (X##_s) \ 604 { \ 605 R##_s = _FP_NANSIGN_##fs; \ 606 R##_c = FP_CLS_NAN; /* NAN */ \ 607 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 608 FP_SET_EXCEPTION(FP_EX_INVALID); \ 609 } \ 610 else \ 611 { \ 612 R##_s = 0; \ 613 R##_c = FP_CLS_INF; /* sqrt(+inf) = +inf */ \ 614 } \ 615 break; \ 616 case FP_CLS_ZERO: \ 617 R##_s = X##_s; \ 618 R##_c = FP_CLS_ZERO; /* sqrt(+-0) = +-0 */ \ 619 break; \ 620 case FP_CLS_NORMAL: \ 621 R##_s = 0; \ 622 if (X##_s) \ 623 { \ 624 R##_c = FP_CLS_NAN; /* sNAN */ \ 625 R##_s = _FP_NANSIGN_##fs; \ 626 _FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \ 627 FP_SET_EXCEPTION(FP_EX_INVALID); \ 628 break; \ 629 } \ 630 R##_c = FP_CLS_NORMAL; \ 631 if (X##_e & 1) \ 632 _FP_FRAC_SLL_##wc(X, 1); \ 633 R##_e = X##_e >> 1; \ 634 _FP_FRAC_SET_##wc(S, _FP_ZEROFRAC_##wc); \ 635 _FP_FRAC_SET_##wc(R, _FP_ZEROFRAC_##wc); \ 636 q = _FP_OVERFLOW_##fs >> 1; \ 637 _FP_SQRT_MEAT_##wc(R, S, T, X, q); \ 638 } \ 639 } while (0) 640 641 /* 642 * Convert from FP to integer 643 */ 644 645 /* RSIGNED can have following values: 646 * 0: the number is required to be 0..(2^rsize)-1, if not, NV is set plus 647 * the result is either 0 or (2^rsize)-1 depending on the sign in such case. 648 * 1: the number is required to be -(2^(rsize-1))..(2^(rsize-1))-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 * 2: the number is required to be -(2^(rsize-1))..(2^(rsize-1))-1, if not, NV is 652 * set plus the result is truncated to fit into destination. 653 * -1: the number is required to be -(2^(rsize-1))..(2^rsize)-1, if not, NV is 654 * set plus the result is either -(2^(rsize-1)) or (2^(rsize-1))-1 depending 655 * on the sign in such case. 656 */ 657 #define _FP_TO_INT(fs, wc, r, X, rsize, rsigned) \ 658 do { \ 659 switch (X##_c) \ 660 { \ 661 case FP_CLS_NORMAL: \ 662 if (X##_e < 0) \ 663 { \ 664 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 665 case FP_CLS_ZERO: \ 666 r = 0; \ 667 } \ 668 else if (X##_e >= rsize - (rsigned > 0 || X##_s) \ 669 || (!rsigned && X##_s)) \ 670 { /* overflow */ \ 671 case FP_CLS_NAN: \ 672 case FP_CLS_INF: \ 673 if (rsigned == 2) \ 674 { \ 675 if (X##_c != FP_CLS_NORMAL \ 676 || X##_e >= rsize - 1 + _FP_WFRACBITS_##fs) \ 677 r = 0; \ 678 else \ 679 { \ 680 _FP_FRAC_SLL_##wc(X, (X##_e - _FP_WFRACBITS_##fs + 1)); \ 681 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 682 } \ 683 } \ 684 else if (rsigned) \ 685 { \ 686 r = 1; \ 687 r <<= rsize - 1; \ 688 r -= 1 - X##_s; \ 689 } \ 690 else \ 691 { \ 692 r = 0; \ 693 if (!X##_s) \ 694 r = ~r; \ 695 } \ 696 FP_SET_EXCEPTION(FP_EX_INVALID); \ 697 } \ 698 else \ 699 { \ 700 if (_FP_W_TYPE_SIZE*wc < rsize) \ 701 { \ 702 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 703 r <<= X##_e - _FP_WFRACBITS_##fs; \ 704 } \ 705 else \ 706 { \ 707 if (X##_e >= _FP_WFRACBITS_##fs) \ 708 _FP_FRAC_SLL_##wc(X, (X##_e - _FP_WFRACBITS_##fs + 1)); \ 709 else if (X##_e < _FP_WFRACBITS_##fs - 1) \ 710 { \ 711 _FP_FRAC_SRS_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 2), \ 712 _FP_WFRACBITS_##fs); \ 713 if (_FP_FRAC_LOW_##wc(X) & 1) \ 714 FP_SET_EXCEPTION(FP_EX_INEXACT); \ 715 _FP_FRAC_SRL_##wc(X, 1); \ 716 } \ 717 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 718 } \ 719 if (rsigned && X##_s) \ 720 r = -r; \ 721 } \ 722 break; \ 723 } \ 724 } while (0) 725 726 #define _FP_TO_INT_ROUND(fs, wc, r, X, rsize, rsigned) \ 727 do { \ 728 r = 0; \ 729 switch (X##_c) \ 730 { \ 731 case FP_CLS_NORMAL: \ 732 if (X##_e >= _FP_FRACBITS_##fs - 1) \ 733 { \ 734 if (X##_e < rsize - 1 + _FP_WFRACBITS_##fs) \ 735 { \ 736 if (X##_e >= _FP_WFRACBITS_##fs - 1) \ 737 { \ 738 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 739 r <<= X##_e - _FP_WFRACBITS_##fs + 1; \ 740 } \ 741 else \ 742 { \ 743 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS - X##_e \ 744 + _FP_FRACBITS_##fs - 1); \ 745 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 746 } \ 747 } \ 748 } \ 749 else \ 750 { \ 751 int _lz0, _lz1; \ 752 if (X##_e <= -_FP_WORKBITS - 1) \ 753 _FP_FRAC_SET_##wc(X, _FP_MINFRAC_##wc); \ 754 else \ 755 _FP_FRAC_SRS_##wc(X, _FP_FRACBITS_##fs - 1 - X##_e, \ 756 _FP_WFRACBITS_##fs); \ 757 _FP_FRAC_CLZ_##wc(_lz0, X); \ 758 _FP_ROUND(wc, X); \ 759 _FP_FRAC_CLZ_##wc(_lz1, X); \ 760 if (_lz1 < _lz0) \ 761 X##_e++; /* For overflow detection. */ \ 762 _FP_FRAC_SRL_##wc(X, _FP_WORKBITS); \ 763 _FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \ 764 } \ 765 if (rsigned && X##_s) \ 766 r = -r; \ 767 if (X##_e >= rsize - (rsigned > 0 || X##_s) \ 768 || (!rsigned && X##_s)) \ 769 { /* overflow */ \ 770 case FP_CLS_NAN: \ 771 case FP_CLS_INF: \ 772 if (!rsigned) \ 773 { \ 774 r = 0; \ 775 if (!X##_s) \ 776 r = ~r; \ 777 } \ 778 else if (rsigned != 2) \ 779 { \ 780 r = 1; \ 781 r <<= rsize - 1; \ 782 r -= 1 - X##_s; \ 783 } \ 784 FP_SET_EXCEPTION(FP_EX_INVALID); \ 785 } \ 786 break; \ 787 case FP_CLS_ZERO: \ 788 break; \ 789 } \ 790 } while (0) 791 792 #define _FP_FROM_INT(fs, wc, X, r, rsize, rtype) \ 793 do { \ 794 if (r) \ 795 { \ 796 unsigned rtype ur_; \ 797 X##_c = FP_CLS_NORMAL; \ 798 \ 799 if ((X##_s = (r < 0))) \ 800 ur_ = (unsigned rtype) -r; \ 801 else \ 802 ur_ = (unsigned rtype) r; \ 803 (void) (((rsize) <= _FP_W_TYPE_SIZE) \ 804 ? ({ __FP_CLZ(X##_e, ur_); }) \ 805 : ({ \ 806 __FP_CLZ_2(X##_e, (_FP_W_TYPE)(ur_ >> _FP_W_TYPE_SIZE), \ 807 (_FP_W_TYPE)ur_); \ 808 })); \ 809 if (rsize < _FP_W_TYPE_SIZE) \ 810 X##_e -= (_FP_W_TYPE_SIZE - rsize); \ 811 X##_e = rsize - X##_e - 1; \ 812 \ 813 if (_FP_FRACBITS_##fs < rsize && _FP_WFRACBITS_##fs <= X##_e) \ 814 __FP_FRAC_SRS_1(ur_, (X##_e - _FP_WFRACBITS_##fs + 1), rsize);\ 815 _FP_FRAC_DISASSEMBLE_##wc(X, ur_, rsize); \ 816 if ((_FP_WFRACBITS_##fs - X##_e - 1) > 0) \ 817 _FP_FRAC_SLL_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 1)); \ 818 } \ 819 else \ 820 { \ 821 X##_c = FP_CLS_ZERO, X##_s = 0; \ 822 } \ 823 } while (0) 824 825 826 #define FP_CONV(dfs,sfs,dwc,swc,D,S) \ 827 do { \ 828 _FP_FRAC_CONV_##dwc##_##swc(dfs, sfs, D, S); \ 829 D##_e = S##_e; \ 830 D##_c = S##_c; \ 831 D##_s = S##_s; \ 832 } while (0) 833 834 /* 835 * Helper primitives. 836 */ 837 838 /* Count leading zeros in a word. */ 839 840 #ifndef __FP_CLZ 841 #if _FP_W_TYPE_SIZE < 64 842 /* this is just to shut the compiler up about shifts > word length -- PMM 02/1998 */ 843 #define __FP_CLZ(r, x) \ 844 do { \ 845 _FP_W_TYPE _t = (x); \ 846 r = _FP_W_TYPE_SIZE - 1; \ 847 if (_t > 0xffff) r -= 16; \ 848 if (_t > 0xffff) _t >>= 16; \ 849 if (_t > 0xff) r -= 8; \ 850 if (_t > 0xff) _t >>= 8; \ 851 if (_t & 0xf0) r -= 4; \ 852 if (_t & 0xf0) _t >>= 4; \ 853 if (_t & 0xc) r -= 2; \ 854 if (_t & 0xc) _t >>= 2; \ 855 if (_t & 0x2) r -= 1; \ 856 } while (0) 857 #else /* not _FP_W_TYPE_SIZE < 64 */ 858 #define __FP_CLZ(r, x) \ 859 do { \ 860 _FP_W_TYPE _t = (x); \ 861 r = _FP_W_TYPE_SIZE - 1; \ 862 if (_t > 0xffffffff) r -= 32; \ 863 if (_t > 0xffffffff) _t >>= 32; \ 864 if (_t > 0xffff) r -= 16; \ 865 if (_t > 0xffff) _t >>= 16; \ 866 if (_t > 0xff) r -= 8; \ 867 if (_t > 0xff) _t >>= 8; \ 868 if (_t & 0xf0) r -= 4; \ 869 if (_t & 0xf0) _t >>= 4; \ 870 if (_t & 0xc) r -= 2; \ 871 if (_t & 0xc) _t >>= 2; \ 872 if (_t & 0x2) r -= 1; \ 873 } while (0) 874 #endif /* not _FP_W_TYPE_SIZE < 64 */ 875 #endif /* ndef __FP_CLZ */ 876 877 #define _FP_DIV_HELP_imm(q, r, n, d) \ 878 do { \ 879 q = n / d, r = n % d; \ 880 } while (0) 881 882 #endif /* __MATH_EMU_OP_COMMON_H__ */ 883