1 2 /*============================================================================ 3 4 This C source file is part of the SoftFloat IEC/IEEE Floating-point Arithmetic 5 Package, Release 2b. 6 7 Written by John R. Hauser. This work was made possible in part by the 8 International Computer Science Institute, located at Suite 600, 1947 Center 9 Street, Berkeley, California 94704. Funding was partially provided by the 10 National Science Foundation under grant MIP-9311980. The original version 11 of this code was written as part of a project to build a fixed-point vector 12 processor in collaboration with the University of California at Berkeley, 13 overseen by Profs. Nelson Morgan and John Wawrzynek. More information 14 is available through the Web page `http://www.cs.berkeley.edu/~jhauser/ 15 arithmetic/SoftFloat.html'. 16 17 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort has 18 been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT TIMES 19 RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO PERSONS 20 AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ALL LOSSES, 21 COSTS, OR OTHER PROBLEMS THEY INCUR DUE TO THE SOFTWARE, AND WHO FURTHERMORE 22 EFFECTIVELY INDEMNIFY JOHN HAUSER AND THE INTERNATIONAL COMPUTER SCIENCE 23 INSTITUTE (possibly via similar legal warning) AGAINST ALL LOSSES, COSTS, OR 24 OTHER PROBLEMS INCURRED BY THEIR CUSTOMERS AND CLIENTS DUE TO THE SOFTWARE. 25 26 Derivative works are acceptable, even for commercial purposes, so long as 27 (1) the source code for the derivative work includes prominent notice that 28 the work is derivative, and (2) the source code includes prominent notice with 29 these four paragraphs for those parts of this code that are retained. 30 31 =============================================================================*/ 32 33 /* FIXME: Flush-To-Zero only effects results. Denormal inputs should also 34 be flushed to zero. */ 35 #include "softfloat.h" 36 37 /*---------------------------------------------------------------------------- 38 | Primitive arithmetic functions, including multi-word arithmetic, and 39 | division and square root approximations. (Can be specialized to target if 40 | desired.) 41 *----------------------------------------------------------------------------*/ 42 #include "softfloat-macros.h" 43 44 /*---------------------------------------------------------------------------- 45 | Functions and definitions to determine: (1) whether tininess for underflow 46 | is detected before or after rounding by default, (2) what (if anything) 47 | happens when exceptions are raised, (3) how signaling NaNs are distinguished 48 | from quiet NaNs, (4) the default generated quiet NaNs, and (5) how NaNs 49 | are propagated from function inputs to output. These details are target- 50 | specific. 51 *----------------------------------------------------------------------------*/ 52 #include "softfloat-specialize.h" 53 54 void set_float_rounding_mode(int val STATUS_PARAM) 55 { 56 STATUS(float_rounding_mode) = val; 57 } 58 59 void set_float_exception_flags(int val STATUS_PARAM) 60 { 61 STATUS(float_exception_flags) = val; 62 } 63 64 #ifdef FLOATX80 65 void set_floatx80_rounding_precision(int val STATUS_PARAM) 66 { 67 STATUS(floatx80_rounding_precision) = val; 68 } 69 #endif 70 71 /*---------------------------------------------------------------------------- 72 | Takes a 64-bit fixed-point value `absZ' with binary point between bits 6 73 | and 7, and returns the properly rounded 32-bit integer corresponding to the 74 | input. If `zSign' is 1, the input is negated before being converted to an 75 | integer. Bit 63 of `absZ' must be zero. Ordinarily, the fixed-point input 76 | is simply rounded to an integer, with the inexact exception raised if the 77 | input cannot be represented exactly as an integer. However, if the fixed- 78 | point input is too large, the invalid exception is raised and the largest 79 | positive or negative integer is returned. 80 *----------------------------------------------------------------------------*/ 81 82 static int32 roundAndPackInt32( flag zSign, bits64 absZ STATUS_PARAM) 83 { 84 int8 roundingMode; 85 flag roundNearestEven; 86 int8 roundIncrement, roundBits; 87 int32 z; 88 89 roundingMode = STATUS(float_rounding_mode); 90 roundNearestEven = ( roundingMode == float_round_nearest_even ); 91 roundIncrement = 0x40; 92 if ( ! roundNearestEven ) { 93 if ( roundingMode == float_round_to_zero ) { 94 roundIncrement = 0; 95 } 96 else { 97 roundIncrement = 0x7F; 98 if ( zSign ) { 99 if ( roundingMode == float_round_up ) roundIncrement = 0; 100 } 101 else { 102 if ( roundingMode == float_round_down ) roundIncrement = 0; 103 } 104 } 105 } 106 roundBits = absZ & 0x7F; 107 absZ = ( absZ + roundIncrement )>>7; 108 absZ &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven ); 109 z = absZ; 110 if ( zSign ) z = - z; 111 if ( ( absZ>>32 ) || ( z && ( ( z < 0 ) ^ zSign ) ) ) { 112 float_raise( float_flag_invalid STATUS_VAR); 113 return zSign ? (sbits32) 0x80000000 : 0x7FFFFFFF; 114 } 115 if ( roundBits ) STATUS(float_exception_flags) |= float_flag_inexact; 116 return z; 117 118 } 119 120 /*---------------------------------------------------------------------------- 121 | Takes the 128-bit fixed-point value formed by concatenating `absZ0' and 122 | `absZ1', with binary point between bits 63 and 64 (between the input words), 123 | and returns the properly rounded 64-bit integer corresponding to the input. 124 | If `zSign' is 1, the input is negated before being converted to an integer. 125 | Ordinarily, the fixed-point input is simply rounded to an integer, with 126 | the inexact exception raised if the input cannot be represented exactly as 127 | an integer. However, if the fixed-point input is too large, the invalid 128 | exception is raised and the largest positive or negative integer is 129 | returned. 130 *----------------------------------------------------------------------------*/ 131 132 static int64 roundAndPackInt64( flag zSign, bits64 absZ0, bits64 absZ1 STATUS_PARAM) 133 { 134 int8 roundingMode; 135 flag roundNearestEven, increment; 136 int64 z; 137 138 roundingMode = STATUS(float_rounding_mode); 139 roundNearestEven = ( roundingMode == float_round_nearest_even ); 140 increment = ( (sbits64) absZ1 < 0 ); 141 if ( ! roundNearestEven ) { 142 if ( roundingMode == float_round_to_zero ) { 143 increment = 0; 144 } 145 else { 146 if ( zSign ) { 147 increment = ( roundingMode == float_round_down ) && absZ1; 148 } 149 else { 150 increment = ( roundingMode == float_round_up ) && absZ1; 151 } 152 } 153 } 154 if ( increment ) { 155 ++absZ0; 156 if ( absZ0 == 0 ) goto overflow; 157 absZ0 &= ~ ( ( (bits64) ( absZ1<<1 ) == 0 ) & roundNearestEven ); 158 } 159 z = absZ0; 160 if ( zSign ) z = - z; 161 if ( z && ( ( z < 0 ) ^ zSign ) ) { 162 overflow: 163 float_raise( float_flag_invalid STATUS_VAR); 164 return 165 zSign ? (sbits64) LIT64( 0x8000000000000000 ) 166 : LIT64( 0x7FFFFFFFFFFFFFFF ); 167 } 168 if ( absZ1 ) STATUS(float_exception_flags) |= float_flag_inexact; 169 return z; 170 171 } 172 173 /*---------------------------------------------------------------------------- 174 | Returns the fraction bits of the single-precision floating-point value `a'. 175 *----------------------------------------------------------------------------*/ 176 177 INLINE bits32 extractFloat32Frac( float32 a ) 178 { 179 180 return float32_val(a) & 0x007FFFFF; 181 182 } 183 184 /*---------------------------------------------------------------------------- 185 | Returns the exponent bits of the single-precision floating-point value `a'. 186 *----------------------------------------------------------------------------*/ 187 188 INLINE int16 extractFloat32Exp( float32 a ) 189 { 190 191 return ( float32_val(a)>>23 ) & 0xFF; 192 193 } 194 195 /*---------------------------------------------------------------------------- 196 | Returns the sign bit of the single-precision floating-point value `a'. 197 *----------------------------------------------------------------------------*/ 198 199 INLINE flag extractFloat32Sign( float32 a ) 200 { 201 202 return float32_val(a)>>31; 203 204 } 205 206 /*---------------------------------------------------------------------------- 207 | Normalizes the subnormal single-precision floating-point value represented 208 | by the denormalized significand `aSig'. The normalized exponent and 209 | significand are stored at the locations pointed to by `zExpPtr' and 210 | `zSigPtr', respectively. 211 *----------------------------------------------------------------------------*/ 212 213 static void 214 normalizeFloat32Subnormal( bits32 aSig, int16 *zExpPtr, bits32 *zSigPtr ) 215 { 216 int8 shiftCount; 217 218 shiftCount = countLeadingZeros32( aSig ) - 8; 219 *zSigPtr = aSig<<shiftCount; 220 *zExpPtr = 1 - shiftCount; 221 222 } 223 224 /*---------------------------------------------------------------------------- 225 | Packs the sign `zSign', exponent `zExp', and significand `zSig' into a 226 | single-precision floating-point value, returning the result. After being 227 | shifted into the proper positions, the three fields are simply added 228 | together to form the result. This means that any integer portion of `zSig' 229 | will be added into the exponent. Since a properly normalized significand 230 | will have an integer portion equal to 1, the `zExp' input should be 1 less 231 | than the desired result exponent whenever `zSig' is a complete, normalized 232 | significand. 233 *----------------------------------------------------------------------------*/ 234 235 INLINE float32 packFloat32( flag zSign, int16 zExp, bits32 zSig ) 236 { 237 238 return make_float32( 239 ( ( (bits32) zSign )<<31 ) + ( ( (bits32) zExp )<<23 ) + zSig); 240 241 } 242 243 /*---------------------------------------------------------------------------- 244 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 245 | and significand `zSig', and returns the proper single-precision floating- 246 | point value corresponding to the abstract input. Ordinarily, the abstract 247 | value is simply rounded and packed into the single-precision format, with 248 | the inexact exception raised if the abstract input cannot be represented 249 | exactly. However, if the abstract value is too large, the overflow and 250 | inexact exceptions are raised and an infinity or maximal finite value is 251 | returned. If the abstract value is too small, the input value is rounded to 252 | a subnormal number, and the underflow and inexact exceptions are raised if 253 | the abstract input cannot be represented exactly as a subnormal single- 254 | precision floating-point number. 255 | The input significand `zSig' has its binary point between bits 30 256 | and 29, which is 7 bits to the left of the usual location. This shifted 257 | significand must be normalized or smaller. If `zSig' is not normalized, 258 | `zExp' must be 0; in that case, the result returned is a subnormal number, 259 | and it must not require rounding. In the usual case that `zSig' is 260 | normalized, `zExp' must be 1 less than the ``true'' floating-point exponent. 261 | The handling of underflow and overflow follows the IEC/IEEE Standard for 262 | Binary Floating-Point Arithmetic. 263 *----------------------------------------------------------------------------*/ 264 265 static float32 roundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig STATUS_PARAM) 266 { 267 int8 roundingMode; 268 flag roundNearestEven; 269 int8 roundIncrement, roundBits; 270 flag isTiny; 271 272 roundingMode = STATUS(float_rounding_mode); 273 roundNearestEven = ( roundingMode == float_round_nearest_even ); 274 roundIncrement = 0x40; 275 if ( ! roundNearestEven ) { 276 if ( roundingMode == float_round_to_zero ) { 277 roundIncrement = 0; 278 } 279 else { 280 roundIncrement = 0x7F; 281 if ( zSign ) { 282 if ( roundingMode == float_round_up ) roundIncrement = 0; 283 } 284 else { 285 if ( roundingMode == float_round_down ) roundIncrement = 0; 286 } 287 } 288 } 289 roundBits = zSig & 0x7F; 290 if ( 0xFD <= (bits16) zExp ) { 291 if ( ( 0xFD < zExp ) 292 || ( ( zExp == 0xFD ) 293 && ( (sbits32) ( zSig + roundIncrement ) < 0 ) ) 294 ) { 295 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 296 return packFloat32( zSign, 0xFF, - ( roundIncrement == 0 )); 297 } 298 if ( zExp < 0 ) { 299 if ( STATUS(flush_to_zero) ) return packFloat32( zSign, 0, 0 ); 300 isTiny = 301 ( STATUS(float_detect_tininess) == float_tininess_before_rounding ) 302 || ( zExp < -1 ) 303 || ( zSig + roundIncrement < 0x80000000 ); 304 shift32RightJamming( zSig, - zExp, &zSig ); 305 zExp = 0; 306 roundBits = zSig & 0x7F; 307 if ( isTiny && roundBits ) float_raise( float_flag_underflow STATUS_VAR); 308 } 309 } 310 if ( roundBits ) STATUS(float_exception_flags) |= float_flag_inexact; 311 zSig = ( zSig + roundIncrement )>>7; 312 zSig &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven ); 313 if ( zSig == 0 ) zExp = 0; 314 return packFloat32( zSign, zExp, zSig ); 315 316 } 317 318 /*---------------------------------------------------------------------------- 319 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 320 | and significand `zSig', and returns the proper single-precision floating- 321 | point value corresponding to the abstract input. This routine is just like 322 | `roundAndPackFloat32' except that `zSig' does not have to be normalized. 323 | Bit 31 of `zSig' must be zero, and `zExp' must be 1 less than the ``true'' 324 | floating-point exponent. 325 *----------------------------------------------------------------------------*/ 326 327 static float32 328 normalizeRoundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig STATUS_PARAM) 329 { 330 int8 shiftCount; 331 332 shiftCount = countLeadingZeros32( zSig ) - 1; 333 return roundAndPackFloat32( zSign, zExp - shiftCount, zSig<<shiftCount STATUS_VAR); 334 335 } 336 337 /*---------------------------------------------------------------------------- 338 | Returns the fraction bits of the double-precision floating-point value `a'. 339 *----------------------------------------------------------------------------*/ 340 341 INLINE bits64 extractFloat64Frac( float64 a ) 342 { 343 344 return float64_val(a) & LIT64( 0x000FFFFFFFFFFFFF ); 345 346 } 347 348 /*---------------------------------------------------------------------------- 349 | Returns the exponent bits of the double-precision floating-point value `a'. 350 *----------------------------------------------------------------------------*/ 351 352 INLINE int16 extractFloat64Exp( float64 a ) 353 { 354 355 return ( float64_val(a)>>52 ) & 0x7FF; 356 357 } 358 359 /*---------------------------------------------------------------------------- 360 | Returns the sign bit of the double-precision floating-point value `a'. 361 *----------------------------------------------------------------------------*/ 362 363 INLINE flag extractFloat64Sign( float64 a ) 364 { 365 366 return float64_val(a)>>63; 367 368 } 369 370 /*---------------------------------------------------------------------------- 371 | Normalizes the subnormal double-precision floating-point value represented 372 | by the denormalized significand `aSig'. The normalized exponent and 373 | significand are stored at the locations pointed to by `zExpPtr' and 374 | `zSigPtr', respectively. 375 *----------------------------------------------------------------------------*/ 376 377 static void 378 normalizeFloat64Subnormal( bits64 aSig, int16 *zExpPtr, bits64 *zSigPtr ) 379 { 380 int8 shiftCount; 381 382 shiftCount = countLeadingZeros64( aSig ) - 11; 383 *zSigPtr = aSig<<shiftCount; 384 *zExpPtr = 1 - shiftCount; 385 386 } 387 388 /*---------------------------------------------------------------------------- 389 | Packs the sign `zSign', exponent `zExp', and significand `zSig' into a 390 | double-precision floating-point value, returning the result. After being 391 | shifted into the proper positions, the three fields are simply added 392 | together to form the result. This means that any integer portion of `zSig' 393 | will be added into the exponent. Since a properly normalized significand 394 | will have an integer portion equal to 1, the `zExp' input should be 1 less 395 | than the desired result exponent whenever `zSig' is a complete, normalized 396 | significand. 397 *----------------------------------------------------------------------------*/ 398 399 INLINE float64 packFloat64( flag zSign, int16 zExp, bits64 zSig ) 400 { 401 402 return make_float64( 403 ( ( (bits64) zSign )<<63 ) + ( ( (bits64) zExp )<<52 ) + zSig); 404 405 } 406 407 /*---------------------------------------------------------------------------- 408 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 409 | and significand `zSig', and returns the proper double-precision floating- 410 | point value corresponding to the abstract input. Ordinarily, the abstract 411 | value is simply rounded and packed into the double-precision format, with 412 | the inexact exception raised if the abstract input cannot be represented 413 | exactly. However, if the abstract value is too large, the overflow and 414 | inexact exceptions are raised and an infinity or maximal finite value is 415 | returned. If the abstract value is too small, the input value is rounded 416 | to a subnormal number, and the underflow and inexact exceptions are raised 417 | if the abstract input cannot be represented exactly as a subnormal double- 418 | precision floating-point number. 419 | The input significand `zSig' has its binary point between bits 62 420 | and 61, which is 10 bits to the left of the usual location. This shifted 421 | significand must be normalized or smaller. If `zSig' is not normalized, 422 | `zExp' must be 0; in that case, the result returned is a subnormal number, 423 | and it must not require rounding. In the usual case that `zSig' is 424 | normalized, `zExp' must be 1 less than the ``true'' floating-point exponent. 425 | The handling of underflow and overflow follows the IEC/IEEE Standard for 426 | Binary Floating-Point Arithmetic. 427 *----------------------------------------------------------------------------*/ 428 429 static float64 roundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig STATUS_PARAM) 430 { 431 int8 roundingMode; 432 flag roundNearestEven; 433 int16 roundIncrement, roundBits; 434 flag isTiny; 435 436 roundingMode = STATUS(float_rounding_mode); 437 roundNearestEven = ( roundingMode == float_round_nearest_even ); 438 roundIncrement = 0x200; 439 if ( ! roundNearestEven ) { 440 if ( roundingMode == float_round_to_zero ) { 441 roundIncrement = 0; 442 } 443 else { 444 roundIncrement = 0x3FF; 445 if ( zSign ) { 446 if ( roundingMode == float_round_up ) roundIncrement = 0; 447 } 448 else { 449 if ( roundingMode == float_round_down ) roundIncrement = 0; 450 } 451 } 452 } 453 roundBits = zSig & 0x3FF; 454 if ( 0x7FD <= (bits16) zExp ) { 455 if ( ( 0x7FD < zExp ) 456 || ( ( zExp == 0x7FD ) 457 && ( (sbits64) ( zSig + roundIncrement ) < 0 ) ) 458 ) { 459 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 460 return packFloat64( zSign, 0x7FF, - ( roundIncrement == 0 )); 461 } 462 if ( zExp < 0 ) { 463 if ( STATUS(flush_to_zero) ) return packFloat64( zSign, 0, 0 ); 464 isTiny = 465 ( STATUS(float_detect_tininess) == float_tininess_before_rounding ) 466 || ( zExp < -1 ) 467 || ( zSig + roundIncrement < LIT64( 0x8000000000000000 ) ); 468 shift64RightJamming( zSig, - zExp, &zSig ); 469 zExp = 0; 470 roundBits = zSig & 0x3FF; 471 if ( isTiny && roundBits ) float_raise( float_flag_underflow STATUS_VAR); 472 } 473 } 474 if ( roundBits ) STATUS(float_exception_flags) |= float_flag_inexact; 475 zSig = ( zSig + roundIncrement )>>10; 476 zSig &= ~ ( ( ( roundBits ^ 0x200 ) == 0 ) & roundNearestEven ); 477 if ( zSig == 0 ) zExp = 0; 478 return packFloat64( zSign, zExp, zSig ); 479 480 } 481 482 /*---------------------------------------------------------------------------- 483 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 484 | and significand `zSig', and returns the proper double-precision floating- 485 | point value corresponding to the abstract input. This routine is just like 486 | `roundAndPackFloat64' except that `zSig' does not have to be normalized. 487 | Bit 63 of `zSig' must be zero, and `zExp' must be 1 less than the ``true'' 488 | floating-point exponent. 489 *----------------------------------------------------------------------------*/ 490 491 static float64 492 normalizeRoundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig STATUS_PARAM) 493 { 494 int8 shiftCount; 495 496 shiftCount = countLeadingZeros64( zSig ) - 1; 497 return roundAndPackFloat64( zSign, zExp - shiftCount, zSig<<shiftCount STATUS_VAR); 498 499 } 500 501 #ifdef FLOATX80 502 503 /*---------------------------------------------------------------------------- 504 | Returns the fraction bits of the extended double-precision floating-point 505 | value `a'. 506 *----------------------------------------------------------------------------*/ 507 508 INLINE bits64 extractFloatx80Frac( floatx80 a ) 509 { 510 511 return a.low; 512 513 } 514 515 /*---------------------------------------------------------------------------- 516 | Returns the exponent bits of the extended double-precision floating-point 517 | value `a'. 518 *----------------------------------------------------------------------------*/ 519 520 INLINE int32 extractFloatx80Exp( floatx80 a ) 521 { 522 523 return a.high & 0x7FFF; 524 525 } 526 527 /*---------------------------------------------------------------------------- 528 | Returns the sign bit of the extended double-precision floating-point value 529 | `a'. 530 *----------------------------------------------------------------------------*/ 531 532 INLINE flag extractFloatx80Sign( floatx80 a ) 533 { 534 535 return a.high>>15; 536 537 } 538 539 /*---------------------------------------------------------------------------- 540 | Normalizes the subnormal extended double-precision floating-point value 541 | represented by the denormalized significand `aSig'. The normalized exponent 542 | and significand are stored at the locations pointed to by `zExpPtr' and 543 | `zSigPtr', respectively. 544 *----------------------------------------------------------------------------*/ 545 546 static void 547 normalizeFloatx80Subnormal( bits64 aSig, int32 *zExpPtr, bits64 *zSigPtr ) 548 { 549 int8 shiftCount; 550 551 shiftCount = countLeadingZeros64( aSig ); 552 *zSigPtr = aSig<<shiftCount; 553 *zExpPtr = 1 - shiftCount; 554 555 } 556 557 /*---------------------------------------------------------------------------- 558 | Packs the sign `zSign', exponent `zExp', and significand `zSig' into an 559 | extended double-precision floating-point value, returning the result. 560 *----------------------------------------------------------------------------*/ 561 562 INLINE floatx80 packFloatx80( flag zSign, int32 zExp, bits64 zSig ) 563 { 564 floatx80 z; 565 566 z.low = zSig; 567 z.high = ( ( (bits16) zSign )<<15 ) + zExp; 568 return z; 569 570 } 571 572 /*---------------------------------------------------------------------------- 573 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 574 | and extended significand formed by the concatenation of `zSig0' and `zSig1', 575 | and returns the proper extended double-precision floating-point value 576 | corresponding to the abstract input. Ordinarily, the abstract value is 577 | rounded and packed into the extended double-precision format, with the 578 | inexact exception raised if the abstract input cannot be represented 579 | exactly. However, if the abstract value is too large, the overflow and 580 | inexact exceptions are raised and an infinity or maximal finite value is 581 | returned. If the abstract value is too small, the input value is rounded to 582 | a subnormal number, and the underflow and inexact exceptions are raised if 583 | the abstract input cannot be represented exactly as a subnormal extended 584 | double-precision floating-point number. 585 | If `roundingPrecision' is 32 or 64, the result is rounded to the same 586 | number of bits as single or double precision, respectively. Otherwise, the 587 | result is rounded to the full precision of the extended double-precision 588 | format. 589 | The input significand must be normalized or smaller. If the input 590 | significand is not normalized, `zExp' must be 0; in that case, the result 591 | returned is a subnormal number, and it must not require rounding. The 592 | handling of underflow and overflow follows the IEC/IEEE Standard for Binary 593 | Floating-Point Arithmetic. 594 *----------------------------------------------------------------------------*/ 595 596 static floatx80 597 roundAndPackFloatx80( 598 int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1 599 STATUS_PARAM) 600 { 601 int8 roundingMode; 602 flag roundNearestEven, increment, isTiny; 603 int64 roundIncrement, roundMask, roundBits; 604 605 roundingMode = STATUS(float_rounding_mode); 606 roundNearestEven = ( roundingMode == float_round_nearest_even ); 607 if ( roundingPrecision == 80 ) goto precision80; 608 if ( roundingPrecision == 64 ) { 609 roundIncrement = LIT64( 0x0000000000000400 ); 610 roundMask = LIT64( 0x00000000000007FF ); 611 } 612 else if ( roundingPrecision == 32 ) { 613 roundIncrement = LIT64( 0x0000008000000000 ); 614 roundMask = LIT64( 0x000000FFFFFFFFFF ); 615 } 616 else { 617 goto precision80; 618 } 619 zSig0 |= ( zSig1 != 0 ); 620 if ( ! roundNearestEven ) { 621 if ( roundingMode == float_round_to_zero ) { 622 roundIncrement = 0; 623 } 624 else { 625 roundIncrement = roundMask; 626 if ( zSign ) { 627 if ( roundingMode == float_round_up ) roundIncrement = 0; 628 } 629 else { 630 if ( roundingMode == float_round_down ) roundIncrement = 0; 631 } 632 } 633 } 634 roundBits = zSig0 & roundMask; 635 if ( 0x7FFD <= (bits32) ( zExp - 1 ) ) { 636 if ( ( 0x7FFE < zExp ) 637 || ( ( zExp == 0x7FFE ) && ( zSig0 + roundIncrement < zSig0 ) ) 638 ) { 639 goto overflow; 640 } 641 if ( zExp <= 0 ) { 642 if ( STATUS(flush_to_zero) ) return packFloatx80( zSign, 0, 0 ); 643 isTiny = 644 ( STATUS(float_detect_tininess) == float_tininess_before_rounding ) 645 || ( zExp < 0 ) 646 || ( zSig0 <= zSig0 + roundIncrement ); 647 shift64RightJamming( zSig0, 1 - zExp, &zSig0 ); 648 zExp = 0; 649 roundBits = zSig0 & roundMask; 650 if ( isTiny && roundBits ) float_raise( float_flag_underflow STATUS_VAR); 651 if ( roundBits ) STATUS(float_exception_flags) |= float_flag_inexact; 652 zSig0 += roundIncrement; 653 if ( (sbits64) zSig0 < 0 ) zExp = 1; 654 roundIncrement = roundMask + 1; 655 if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) { 656 roundMask |= roundIncrement; 657 } 658 zSig0 &= ~ roundMask; 659 return packFloatx80( zSign, zExp, zSig0 ); 660 } 661 } 662 if ( roundBits ) STATUS(float_exception_flags) |= float_flag_inexact; 663 zSig0 += roundIncrement; 664 if ( zSig0 < roundIncrement ) { 665 ++zExp; 666 zSig0 = LIT64( 0x8000000000000000 ); 667 } 668 roundIncrement = roundMask + 1; 669 if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) { 670 roundMask |= roundIncrement; 671 } 672 zSig0 &= ~ roundMask; 673 if ( zSig0 == 0 ) zExp = 0; 674 return packFloatx80( zSign, zExp, zSig0 ); 675 precision80: 676 increment = ( (sbits64) zSig1 < 0 ); 677 if ( ! roundNearestEven ) { 678 if ( roundingMode == float_round_to_zero ) { 679 increment = 0; 680 } 681 else { 682 if ( zSign ) { 683 increment = ( roundingMode == float_round_down ) && zSig1; 684 } 685 else { 686 increment = ( roundingMode == float_round_up ) && zSig1; 687 } 688 } 689 } 690 if ( 0x7FFD <= (bits32) ( zExp - 1 ) ) { 691 if ( ( 0x7FFE < zExp ) 692 || ( ( zExp == 0x7FFE ) 693 && ( zSig0 == LIT64( 0xFFFFFFFFFFFFFFFF ) ) 694 && increment 695 ) 696 ) { 697 roundMask = 0; 698 overflow: 699 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 700 if ( ( roundingMode == float_round_to_zero ) 701 || ( zSign && ( roundingMode == float_round_up ) ) 702 || ( ! zSign && ( roundingMode == float_round_down ) ) 703 ) { 704 return packFloatx80( zSign, 0x7FFE, ~ roundMask ); 705 } 706 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 707 } 708 if ( zExp <= 0 ) { 709 isTiny = 710 ( STATUS(float_detect_tininess) == float_tininess_before_rounding ) 711 || ( zExp < 0 ) 712 || ! increment 713 || ( zSig0 < LIT64( 0xFFFFFFFFFFFFFFFF ) ); 714 shift64ExtraRightJamming( zSig0, zSig1, 1 - zExp, &zSig0, &zSig1 ); 715 zExp = 0; 716 if ( isTiny && zSig1 ) float_raise( float_flag_underflow STATUS_VAR); 717 if ( zSig1 ) STATUS(float_exception_flags) |= float_flag_inexact; 718 if ( roundNearestEven ) { 719 increment = ( (sbits64) zSig1 < 0 ); 720 } 721 else { 722 if ( zSign ) { 723 increment = ( roundingMode == float_round_down ) && zSig1; 724 } 725 else { 726 increment = ( roundingMode == float_round_up ) && zSig1; 727 } 728 } 729 if ( increment ) { 730 ++zSig0; 731 zSig0 &= 732 ~ ( ( (bits64) ( zSig1<<1 ) == 0 ) & roundNearestEven ); 733 if ( (sbits64) zSig0 < 0 ) zExp = 1; 734 } 735 return packFloatx80( zSign, zExp, zSig0 ); 736 } 737 } 738 if ( zSig1 ) STATUS(float_exception_flags) |= float_flag_inexact; 739 if ( increment ) { 740 ++zSig0; 741 if ( zSig0 == 0 ) { 742 ++zExp; 743 zSig0 = LIT64( 0x8000000000000000 ); 744 } 745 else { 746 zSig0 &= ~ ( ( (bits64) ( zSig1<<1 ) == 0 ) & roundNearestEven ); 747 } 748 } 749 else { 750 if ( zSig0 == 0 ) zExp = 0; 751 } 752 return packFloatx80( zSign, zExp, zSig0 ); 753 754 } 755 756 /*---------------------------------------------------------------------------- 757 | Takes an abstract floating-point value having sign `zSign', exponent 758 | `zExp', and significand formed by the concatenation of `zSig0' and `zSig1', 759 | and returns the proper extended double-precision floating-point value 760 | corresponding to the abstract input. This routine is just like 761 | `roundAndPackFloatx80' except that the input significand does not have to be 762 | normalized. 763 *----------------------------------------------------------------------------*/ 764 765 static floatx80 766 normalizeRoundAndPackFloatx80( 767 int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1 768 STATUS_PARAM) 769 { 770 int8 shiftCount; 771 772 if ( zSig0 == 0 ) { 773 zSig0 = zSig1; 774 zSig1 = 0; 775 zExp -= 64; 776 } 777 shiftCount = countLeadingZeros64( zSig0 ); 778 shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 ); 779 zExp -= shiftCount; 780 return 781 roundAndPackFloatx80( roundingPrecision, zSign, zExp, zSig0, zSig1 STATUS_VAR); 782 783 } 784 785 #endif 786 787 #ifdef FLOAT128 788 789 /*---------------------------------------------------------------------------- 790 | Returns the least-significant 64 fraction bits of the quadruple-precision 791 | floating-point value `a'. 792 *----------------------------------------------------------------------------*/ 793 794 INLINE bits64 extractFloat128Frac1( float128 a ) 795 { 796 797 return a.low; 798 799 } 800 801 /*---------------------------------------------------------------------------- 802 | Returns the most-significant 48 fraction bits of the quadruple-precision 803 | floating-point value `a'. 804 *----------------------------------------------------------------------------*/ 805 806 INLINE bits64 extractFloat128Frac0( float128 a ) 807 { 808 809 return a.high & LIT64( 0x0000FFFFFFFFFFFF ); 810 811 } 812 813 /*---------------------------------------------------------------------------- 814 | Returns the exponent bits of the quadruple-precision floating-point value 815 | `a'. 816 *----------------------------------------------------------------------------*/ 817 818 INLINE int32 extractFloat128Exp( float128 a ) 819 { 820 821 return ( a.high>>48 ) & 0x7FFF; 822 823 } 824 825 /*---------------------------------------------------------------------------- 826 | Returns the sign bit of the quadruple-precision floating-point value `a'. 827 *----------------------------------------------------------------------------*/ 828 829 INLINE flag extractFloat128Sign( float128 a ) 830 { 831 832 return a.high>>63; 833 834 } 835 836 /*---------------------------------------------------------------------------- 837 | Normalizes the subnormal quadruple-precision floating-point value 838 | represented by the denormalized significand formed by the concatenation of 839 | `aSig0' and `aSig1'. The normalized exponent is stored at the location 840 | pointed to by `zExpPtr'. The most significant 49 bits of the normalized 841 | significand are stored at the location pointed to by `zSig0Ptr', and the 842 | least significant 64 bits of the normalized significand are stored at the 843 | location pointed to by `zSig1Ptr'. 844 *----------------------------------------------------------------------------*/ 845 846 static void 847 normalizeFloat128Subnormal( 848 bits64 aSig0, 849 bits64 aSig1, 850 int32 *zExpPtr, 851 bits64 *zSig0Ptr, 852 bits64 *zSig1Ptr 853 ) 854 { 855 int8 shiftCount; 856 857 if ( aSig0 == 0 ) { 858 shiftCount = countLeadingZeros64( aSig1 ) - 15; 859 if ( shiftCount < 0 ) { 860 *zSig0Ptr = aSig1>>( - shiftCount ); 861 *zSig1Ptr = aSig1<<( shiftCount & 63 ); 862 } 863 else { 864 *zSig0Ptr = aSig1<<shiftCount; 865 *zSig1Ptr = 0; 866 } 867 *zExpPtr = - shiftCount - 63; 868 } 869 else { 870 shiftCount = countLeadingZeros64( aSig0 ) - 15; 871 shortShift128Left( aSig0, aSig1, shiftCount, zSig0Ptr, zSig1Ptr ); 872 *zExpPtr = 1 - shiftCount; 873 } 874 875 } 876 877 /*---------------------------------------------------------------------------- 878 | Packs the sign `zSign', the exponent `zExp', and the significand formed 879 | by the concatenation of `zSig0' and `zSig1' into a quadruple-precision 880 | floating-point value, returning the result. After being shifted into the 881 | proper positions, the three fields `zSign', `zExp', and `zSig0' are simply 882 | added together to form the most significant 32 bits of the result. This 883 | means that any integer portion of `zSig0' will be added into the exponent. 884 | Since a properly normalized significand will have an integer portion equal 885 | to 1, the `zExp' input should be 1 less than the desired result exponent 886 | whenever `zSig0' and `zSig1' concatenated form a complete, normalized 887 | significand. 888 *----------------------------------------------------------------------------*/ 889 890 INLINE float128 891 packFloat128( flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1 ) 892 { 893 float128 z; 894 895 z.low = zSig1; 896 z.high = ( ( (bits64) zSign )<<63 ) + ( ( (bits64) zExp )<<48 ) + zSig0; 897 return z; 898 899 } 900 901 /*---------------------------------------------------------------------------- 902 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 903 | and extended significand formed by the concatenation of `zSig0', `zSig1', 904 | and `zSig2', and returns the proper quadruple-precision floating-point value 905 | corresponding to the abstract input. Ordinarily, the abstract value is 906 | simply rounded and packed into the quadruple-precision format, with the 907 | inexact exception raised if the abstract input cannot be represented 908 | exactly. However, if the abstract value is too large, the overflow and 909 | inexact exceptions are raised and an infinity or maximal finite value is 910 | returned. If the abstract value is too small, the input value is rounded to 911 | a subnormal number, and the underflow and inexact exceptions are raised if 912 | the abstract input cannot be represented exactly as a subnormal quadruple- 913 | precision floating-point number. 914 | The input significand must be normalized or smaller. If the input 915 | significand is not normalized, `zExp' must be 0; in that case, the result 916 | returned is a subnormal number, and it must not require rounding. In the 917 | usual case that the input significand is normalized, `zExp' must be 1 less 918 | than the ``true'' floating-point exponent. The handling of underflow and 919 | overflow follows the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 920 *----------------------------------------------------------------------------*/ 921 922 static float128 923 roundAndPackFloat128( 924 flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1, bits64 zSig2 STATUS_PARAM) 925 { 926 int8 roundingMode; 927 flag roundNearestEven, increment, isTiny; 928 929 roundingMode = STATUS(float_rounding_mode); 930 roundNearestEven = ( roundingMode == float_round_nearest_even ); 931 increment = ( (sbits64) zSig2 < 0 ); 932 if ( ! roundNearestEven ) { 933 if ( roundingMode == float_round_to_zero ) { 934 increment = 0; 935 } 936 else { 937 if ( zSign ) { 938 increment = ( roundingMode == float_round_down ) && zSig2; 939 } 940 else { 941 increment = ( roundingMode == float_round_up ) && zSig2; 942 } 943 } 944 } 945 if ( 0x7FFD <= (bits32) zExp ) { 946 if ( ( 0x7FFD < zExp ) 947 || ( ( zExp == 0x7FFD ) 948 && eq128( 949 LIT64( 0x0001FFFFFFFFFFFF ), 950 LIT64( 0xFFFFFFFFFFFFFFFF ), 951 zSig0, 952 zSig1 953 ) 954 && increment 955 ) 956 ) { 957 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 958 if ( ( roundingMode == float_round_to_zero ) 959 || ( zSign && ( roundingMode == float_round_up ) ) 960 || ( ! zSign && ( roundingMode == float_round_down ) ) 961 ) { 962 return 963 packFloat128( 964 zSign, 965 0x7FFE, 966 LIT64( 0x0000FFFFFFFFFFFF ), 967 LIT64( 0xFFFFFFFFFFFFFFFF ) 968 ); 969 } 970 return packFloat128( zSign, 0x7FFF, 0, 0 ); 971 } 972 if ( zExp < 0 ) { 973 if ( STATUS(flush_to_zero) ) return packFloat128( zSign, 0, 0, 0 ); 974 isTiny = 975 ( STATUS(float_detect_tininess) == float_tininess_before_rounding ) 976 || ( zExp < -1 ) 977 || ! increment 978 || lt128( 979 zSig0, 980 zSig1, 981 LIT64( 0x0001FFFFFFFFFFFF ), 982 LIT64( 0xFFFFFFFFFFFFFFFF ) 983 ); 984 shift128ExtraRightJamming( 985 zSig0, zSig1, zSig2, - zExp, &zSig0, &zSig1, &zSig2 ); 986 zExp = 0; 987 if ( isTiny && zSig2 ) float_raise( float_flag_underflow STATUS_VAR); 988 if ( roundNearestEven ) { 989 increment = ( (sbits64) zSig2 < 0 ); 990 } 991 else { 992 if ( zSign ) { 993 increment = ( roundingMode == float_round_down ) && zSig2; 994 } 995 else { 996 increment = ( roundingMode == float_round_up ) && zSig2; 997 } 998 } 999 } 1000 } 1001 if ( zSig2 ) STATUS(float_exception_flags) |= float_flag_inexact; 1002 if ( increment ) { 1003 add128( zSig0, zSig1, 0, 1, &zSig0, &zSig1 ); 1004 zSig1 &= ~ ( ( zSig2 + zSig2 == 0 ) & roundNearestEven ); 1005 } 1006 else { 1007 if ( ( zSig0 | zSig1 ) == 0 ) zExp = 0; 1008 } 1009 return packFloat128( zSign, zExp, zSig0, zSig1 ); 1010 1011 } 1012 1013 /*---------------------------------------------------------------------------- 1014 | Takes an abstract floating-point value having sign `zSign', exponent `zExp', 1015 | and significand formed by the concatenation of `zSig0' and `zSig1', and 1016 | returns the proper quadruple-precision floating-point value corresponding 1017 | to the abstract input. This routine is just like `roundAndPackFloat128' 1018 | except that the input significand has fewer bits and does not have to be 1019 | normalized. In all cases, `zExp' must be 1 less than the ``true'' floating- 1020 | point exponent. 1021 *----------------------------------------------------------------------------*/ 1022 1023 static float128 1024 normalizeRoundAndPackFloat128( 1025 flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1 STATUS_PARAM) 1026 { 1027 int8 shiftCount; 1028 bits64 zSig2; 1029 1030 if ( zSig0 == 0 ) { 1031 zSig0 = zSig1; 1032 zSig1 = 0; 1033 zExp -= 64; 1034 } 1035 shiftCount = countLeadingZeros64( zSig0 ) - 15; 1036 if ( 0 <= shiftCount ) { 1037 zSig2 = 0; 1038 shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 ); 1039 } 1040 else { 1041 shift128ExtraRightJamming( 1042 zSig0, zSig1, 0, - shiftCount, &zSig0, &zSig1, &zSig2 ); 1043 } 1044 zExp -= shiftCount; 1045 return roundAndPackFloat128( zSign, zExp, zSig0, zSig1, zSig2 STATUS_VAR); 1046 1047 } 1048 1049 #endif 1050 1051 /*---------------------------------------------------------------------------- 1052 | Returns the result of converting the 32-bit two's complement integer `a' 1053 | to the single-precision floating-point format. The conversion is performed 1054 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1055 *----------------------------------------------------------------------------*/ 1056 1057 float32 int32_to_float32( int32 a STATUS_PARAM ) 1058 { 1059 flag zSign; 1060 1061 if ( a == 0 ) return float32_zero; 1062 if ( a == (sbits32) 0x80000000 ) return packFloat32( 1, 0x9E, 0 ); 1063 zSign = ( a < 0 ); 1064 return normalizeRoundAndPackFloat32( zSign, 0x9C, zSign ? - a : a STATUS_VAR ); 1065 1066 } 1067 1068 /*---------------------------------------------------------------------------- 1069 | Returns the result of converting the 32-bit two's complement integer `a' 1070 | to the double-precision floating-point format. The conversion is performed 1071 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1072 *----------------------------------------------------------------------------*/ 1073 1074 float64 int32_to_float64( int32 a STATUS_PARAM ) 1075 { 1076 flag zSign; 1077 uint32 absA; 1078 int8 shiftCount; 1079 bits64 zSig; 1080 1081 if ( a == 0 ) return float64_zero; 1082 zSign = ( a < 0 ); 1083 absA = zSign ? - a : a; 1084 shiftCount = countLeadingZeros32( absA ) + 21; 1085 zSig = absA; 1086 return packFloat64( zSign, 0x432 - shiftCount, zSig<<shiftCount ); 1087 1088 } 1089 1090 #ifdef FLOATX80 1091 1092 /*---------------------------------------------------------------------------- 1093 | Returns the result of converting the 32-bit two's complement integer `a' 1094 | to the extended double-precision floating-point format. The conversion 1095 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 1096 | Arithmetic. 1097 *----------------------------------------------------------------------------*/ 1098 1099 floatx80 int32_to_floatx80( int32 a STATUS_PARAM ) 1100 { 1101 flag zSign; 1102 uint32 absA; 1103 int8 shiftCount; 1104 bits64 zSig; 1105 1106 if ( a == 0 ) return packFloatx80( 0, 0, 0 ); 1107 zSign = ( a < 0 ); 1108 absA = zSign ? - a : a; 1109 shiftCount = countLeadingZeros32( absA ) + 32; 1110 zSig = absA; 1111 return packFloatx80( zSign, 0x403E - shiftCount, zSig<<shiftCount ); 1112 1113 } 1114 1115 #endif 1116 1117 #ifdef FLOAT128 1118 1119 /*---------------------------------------------------------------------------- 1120 | Returns the result of converting the 32-bit two's complement integer `a' to 1121 | the quadruple-precision floating-point format. The conversion is performed 1122 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1123 *----------------------------------------------------------------------------*/ 1124 1125 float128 int32_to_float128( int32 a STATUS_PARAM ) 1126 { 1127 flag zSign; 1128 uint32 absA; 1129 int8 shiftCount; 1130 bits64 zSig0; 1131 1132 if ( a == 0 ) return packFloat128( 0, 0, 0, 0 ); 1133 zSign = ( a < 0 ); 1134 absA = zSign ? - a : a; 1135 shiftCount = countLeadingZeros32( absA ) + 17; 1136 zSig0 = absA; 1137 return packFloat128( zSign, 0x402E - shiftCount, zSig0<<shiftCount, 0 ); 1138 1139 } 1140 1141 #endif 1142 1143 /*---------------------------------------------------------------------------- 1144 | Returns the result of converting the 64-bit two's complement integer `a' 1145 | to the single-precision floating-point format. The conversion is performed 1146 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1147 *----------------------------------------------------------------------------*/ 1148 1149 float32 int64_to_float32( int64 a STATUS_PARAM ) 1150 { 1151 flag zSign; 1152 uint64 absA; 1153 int8 shiftCount; 1154 1155 if ( a == 0 ) return float32_zero; 1156 zSign = ( a < 0 ); 1157 absA = zSign ? - a : a; 1158 shiftCount = countLeadingZeros64( absA ) - 40; 1159 if ( 0 <= shiftCount ) { 1160 return packFloat32( zSign, 0x95 - shiftCount, absA<<shiftCount ); 1161 } 1162 else { 1163 shiftCount += 7; 1164 if ( shiftCount < 0 ) { 1165 shift64RightJamming( absA, - shiftCount, &absA ); 1166 } 1167 else { 1168 absA <<= shiftCount; 1169 } 1170 return roundAndPackFloat32( zSign, 0x9C - shiftCount, absA STATUS_VAR ); 1171 } 1172 1173 } 1174 1175 float32 uint64_to_float32( uint64 a STATUS_PARAM ) 1176 { 1177 int8 shiftCount; 1178 1179 if ( a == 0 ) return float32_zero; 1180 shiftCount = countLeadingZeros64( a ) - 40; 1181 if ( 0 <= shiftCount ) { 1182 return packFloat32( 1 > 0, 0x95 - shiftCount, a<<shiftCount ); 1183 } 1184 else { 1185 shiftCount += 7; 1186 if ( shiftCount < 0 ) { 1187 shift64RightJamming( a, - shiftCount, &a ); 1188 } 1189 else { 1190 a <<= shiftCount; 1191 } 1192 return roundAndPackFloat32( 1 > 0, 0x9C - shiftCount, a STATUS_VAR ); 1193 } 1194 } 1195 1196 /*---------------------------------------------------------------------------- 1197 | Returns the result of converting the 64-bit two's complement integer `a' 1198 | to the double-precision floating-point format. The conversion is performed 1199 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1200 *----------------------------------------------------------------------------*/ 1201 1202 float64 int64_to_float64( int64 a STATUS_PARAM ) 1203 { 1204 flag zSign; 1205 1206 if ( a == 0 ) return float64_zero; 1207 if ( a == (sbits64) LIT64( 0x8000000000000000 ) ) { 1208 return packFloat64( 1, 0x43E, 0 ); 1209 } 1210 zSign = ( a < 0 ); 1211 return normalizeRoundAndPackFloat64( zSign, 0x43C, zSign ? - a : a STATUS_VAR ); 1212 1213 } 1214 1215 float64 uint64_to_float64( uint64 a STATUS_PARAM ) 1216 { 1217 if ( a == 0 ) return float64_zero; 1218 return normalizeRoundAndPackFloat64( 0, 0x43C, a STATUS_VAR ); 1219 1220 } 1221 1222 #ifdef FLOATX80 1223 1224 /*---------------------------------------------------------------------------- 1225 | Returns the result of converting the 64-bit two's complement integer `a' 1226 | to the extended double-precision floating-point format. The conversion 1227 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 1228 | Arithmetic. 1229 *----------------------------------------------------------------------------*/ 1230 1231 floatx80 int64_to_floatx80( int64 a STATUS_PARAM ) 1232 { 1233 flag zSign; 1234 uint64 absA; 1235 int8 shiftCount; 1236 1237 if ( a == 0 ) return packFloatx80( 0, 0, 0 ); 1238 zSign = ( a < 0 ); 1239 absA = zSign ? - a : a; 1240 shiftCount = countLeadingZeros64( absA ); 1241 return packFloatx80( zSign, 0x403E - shiftCount, absA<<shiftCount ); 1242 1243 } 1244 1245 #endif 1246 1247 #ifdef FLOAT128 1248 1249 /*---------------------------------------------------------------------------- 1250 | Returns the result of converting the 64-bit two's complement integer `a' to 1251 | the quadruple-precision floating-point format. The conversion is performed 1252 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1253 *----------------------------------------------------------------------------*/ 1254 1255 float128 int64_to_float128( int64 a STATUS_PARAM ) 1256 { 1257 flag zSign; 1258 uint64 absA; 1259 int8 shiftCount; 1260 int32 zExp; 1261 bits64 zSig0, zSig1; 1262 1263 if ( a == 0 ) return packFloat128( 0, 0, 0, 0 ); 1264 zSign = ( a < 0 ); 1265 absA = zSign ? - a : a; 1266 shiftCount = countLeadingZeros64( absA ) + 49; 1267 zExp = 0x406E - shiftCount; 1268 if ( 64 <= shiftCount ) { 1269 zSig1 = 0; 1270 zSig0 = absA; 1271 shiftCount -= 64; 1272 } 1273 else { 1274 zSig1 = absA; 1275 zSig0 = 0; 1276 } 1277 shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 ); 1278 return packFloat128( zSign, zExp, zSig0, zSig1 ); 1279 1280 } 1281 1282 #endif 1283 1284 /*---------------------------------------------------------------------------- 1285 | Returns the result of converting the single-precision floating-point value 1286 | `a' to the 32-bit two's complement integer format. The conversion is 1287 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1288 | Arithmetic---which means in particular that the conversion is rounded 1289 | according to the current rounding mode. If `a' is a NaN, the largest 1290 | positive integer is returned. Otherwise, if the conversion overflows, the 1291 | largest integer with the same sign as `a' is returned. 1292 *----------------------------------------------------------------------------*/ 1293 1294 int32 float32_to_int32( float32 a STATUS_PARAM ) 1295 { 1296 flag aSign; 1297 int16 aExp, shiftCount; 1298 bits32 aSig; 1299 bits64 aSig64; 1300 1301 aSig = extractFloat32Frac( a ); 1302 aExp = extractFloat32Exp( a ); 1303 aSign = extractFloat32Sign( a ); 1304 if ( ( aExp == 0xFF ) && aSig ) aSign = 0; 1305 if ( aExp ) aSig |= 0x00800000; 1306 shiftCount = 0xAF - aExp; 1307 aSig64 = aSig; 1308 aSig64 <<= 32; 1309 if ( 0 < shiftCount ) shift64RightJamming( aSig64, shiftCount, &aSig64 ); 1310 return roundAndPackInt32( aSign, aSig64 STATUS_VAR ); 1311 1312 } 1313 1314 /*---------------------------------------------------------------------------- 1315 | Returns the result of converting the single-precision floating-point value 1316 | `a' to the 32-bit two's complement integer format. The conversion is 1317 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1318 | Arithmetic, except that the conversion is always rounded toward zero. 1319 | If `a' is a NaN, the largest positive integer is returned. Otherwise, if 1320 | the conversion overflows, the largest integer with the same sign as `a' is 1321 | returned. 1322 *----------------------------------------------------------------------------*/ 1323 1324 int32 float32_to_int32_round_to_zero( float32 a STATUS_PARAM ) 1325 { 1326 flag aSign; 1327 int16 aExp, shiftCount; 1328 bits32 aSig; 1329 int32 z; 1330 1331 aSig = extractFloat32Frac( a ); 1332 aExp = extractFloat32Exp( a ); 1333 aSign = extractFloat32Sign( a ); 1334 shiftCount = aExp - 0x9E; 1335 if ( 0 <= shiftCount ) { 1336 if ( float32_val(a) != 0xCF000000 ) { 1337 float_raise( float_flag_invalid STATUS_VAR); 1338 if ( ! aSign || ( ( aExp == 0xFF ) && aSig ) ) return 0x7FFFFFFF; 1339 } 1340 return (sbits32) 0x80000000; 1341 } 1342 else if ( aExp <= 0x7E ) { 1343 if ( aExp | aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 1344 return 0; 1345 } 1346 aSig = ( aSig | 0x00800000 )<<8; 1347 z = aSig>>( - shiftCount ); 1348 if ( (bits32) ( aSig<<( shiftCount & 31 ) ) ) { 1349 STATUS(float_exception_flags) |= float_flag_inexact; 1350 } 1351 if ( aSign ) z = - z; 1352 return z; 1353 1354 } 1355 1356 /*---------------------------------------------------------------------------- 1357 | Returns the result of converting the single-precision floating-point value 1358 | `a' to the 64-bit two's complement integer format. The conversion is 1359 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1360 | Arithmetic---which means in particular that the conversion is rounded 1361 | according to the current rounding mode. If `a' is a NaN, the largest 1362 | positive integer is returned. Otherwise, if the conversion overflows, the 1363 | largest integer with the same sign as `a' is returned. 1364 *----------------------------------------------------------------------------*/ 1365 1366 int64 float32_to_int64( float32 a STATUS_PARAM ) 1367 { 1368 flag aSign; 1369 int16 aExp, shiftCount; 1370 bits32 aSig; 1371 bits64 aSig64, aSigExtra; 1372 1373 aSig = extractFloat32Frac( a ); 1374 aExp = extractFloat32Exp( a ); 1375 aSign = extractFloat32Sign( a ); 1376 shiftCount = 0xBE - aExp; 1377 if ( shiftCount < 0 ) { 1378 float_raise( float_flag_invalid STATUS_VAR); 1379 if ( ! aSign || ( ( aExp == 0xFF ) && aSig ) ) { 1380 return LIT64( 0x7FFFFFFFFFFFFFFF ); 1381 } 1382 return (sbits64) LIT64( 0x8000000000000000 ); 1383 } 1384 if ( aExp ) aSig |= 0x00800000; 1385 aSig64 = aSig; 1386 aSig64 <<= 40; 1387 shift64ExtraRightJamming( aSig64, 0, shiftCount, &aSig64, &aSigExtra ); 1388 return roundAndPackInt64( aSign, aSig64, aSigExtra STATUS_VAR ); 1389 1390 } 1391 1392 /*---------------------------------------------------------------------------- 1393 | Returns the result of converting the single-precision floating-point value 1394 | `a' to the 64-bit two's complement integer format. The conversion is 1395 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1396 | Arithmetic, except that the conversion is always rounded toward zero. If 1397 | `a' is a NaN, the largest positive integer is returned. Otherwise, if the 1398 | conversion overflows, the largest integer with the same sign as `a' is 1399 | returned. 1400 *----------------------------------------------------------------------------*/ 1401 1402 int64 float32_to_int64_round_to_zero( float32 a STATUS_PARAM ) 1403 { 1404 flag aSign; 1405 int16 aExp, shiftCount; 1406 bits32 aSig; 1407 bits64 aSig64; 1408 int64 z; 1409 1410 aSig = extractFloat32Frac( a ); 1411 aExp = extractFloat32Exp( a ); 1412 aSign = extractFloat32Sign( a ); 1413 shiftCount = aExp - 0xBE; 1414 if ( 0 <= shiftCount ) { 1415 if ( float32_val(a) != 0xDF000000 ) { 1416 float_raise( float_flag_invalid STATUS_VAR); 1417 if ( ! aSign || ( ( aExp == 0xFF ) && aSig ) ) { 1418 return LIT64( 0x7FFFFFFFFFFFFFFF ); 1419 } 1420 } 1421 return (sbits64) LIT64( 0x8000000000000000 ); 1422 } 1423 else if ( aExp <= 0x7E ) { 1424 if ( aExp | aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 1425 return 0; 1426 } 1427 aSig64 = aSig | 0x00800000; 1428 aSig64 <<= 40; 1429 z = aSig64>>( - shiftCount ); 1430 if ( (bits64) ( aSig64<<( shiftCount & 63 ) ) ) { 1431 STATUS(float_exception_flags) |= float_flag_inexact; 1432 } 1433 if ( aSign ) z = - z; 1434 return z; 1435 1436 } 1437 1438 /*---------------------------------------------------------------------------- 1439 | Returns the result of converting the single-precision floating-point value 1440 | `a' to the double-precision floating-point format. The conversion is 1441 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1442 | Arithmetic. 1443 *----------------------------------------------------------------------------*/ 1444 1445 float64 float32_to_float64( float32 a STATUS_PARAM ) 1446 { 1447 flag aSign; 1448 int16 aExp; 1449 bits32 aSig; 1450 1451 aSig = extractFloat32Frac( a ); 1452 aExp = extractFloat32Exp( a ); 1453 aSign = extractFloat32Sign( a ); 1454 if ( aExp == 0xFF ) { 1455 if ( aSig ) return commonNaNToFloat64( float32ToCommonNaN( a STATUS_VAR )); 1456 return packFloat64( aSign, 0x7FF, 0 ); 1457 } 1458 if ( aExp == 0 ) { 1459 if ( aSig == 0 ) return packFloat64( aSign, 0, 0 ); 1460 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1461 --aExp; 1462 } 1463 return packFloat64( aSign, aExp + 0x380, ( (bits64) aSig )<<29 ); 1464 1465 } 1466 1467 #ifdef FLOATX80 1468 1469 /*---------------------------------------------------------------------------- 1470 | Returns the result of converting the single-precision floating-point value 1471 | `a' to the extended double-precision floating-point format. The conversion 1472 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 1473 | Arithmetic. 1474 *----------------------------------------------------------------------------*/ 1475 1476 floatx80 float32_to_floatx80( float32 a STATUS_PARAM ) 1477 { 1478 flag aSign; 1479 int16 aExp; 1480 bits32 aSig; 1481 1482 aSig = extractFloat32Frac( a ); 1483 aExp = extractFloat32Exp( a ); 1484 aSign = extractFloat32Sign( a ); 1485 if ( aExp == 0xFF ) { 1486 if ( aSig ) return commonNaNToFloatx80( float32ToCommonNaN( a STATUS_VAR ) ); 1487 return packFloatx80( aSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 1488 } 1489 if ( aExp == 0 ) { 1490 if ( aSig == 0 ) return packFloatx80( aSign, 0, 0 ); 1491 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1492 } 1493 aSig |= 0x00800000; 1494 return packFloatx80( aSign, aExp + 0x3F80, ( (bits64) aSig )<<40 ); 1495 1496 } 1497 1498 #endif 1499 1500 #ifdef FLOAT128 1501 1502 /*---------------------------------------------------------------------------- 1503 | Returns the result of converting the single-precision floating-point value 1504 | `a' to the double-precision floating-point format. The conversion is 1505 | performed according to the IEC/IEEE Standard for Binary Floating-Point 1506 | Arithmetic. 1507 *----------------------------------------------------------------------------*/ 1508 1509 float128 float32_to_float128( float32 a STATUS_PARAM ) 1510 { 1511 flag aSign; 1512 int16 aExp; 1513 bits32 aSig; 1514 1515 aSig = extractFloat32Frac( a ); 1516 aExp = extractFloat32Exp( a ); 1517 aSign = extractFloat32Sign( a ); 1518 if ( aExp == 0xFF ) { 1519 if ( aSig ) return commonNaNToFloat128( float32ToCommonNaN( a STATUS_VAR ) ); 1520 return packFloat128( aSign, 0x7FFF, 0, 0 ); 1521 } 1522 if ( aExp == 0 ) { 1523 if ( aSig == 0 ) return packFloat128( aSign, 0, 0, 0 ); 1524 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1525 --aExp; 1526 } 1527 return packFloat128( aSign, aExp + 0x3F80, ( (bits64) aSig )<<25, 0 ); 1528 1529 } 1530 1531 #endif 1532 1533 /*---------------------------------------------------------------------------- 1534 | Rounds the single-precision floating-point value `a' to an integer, and 1535 | returns the result as a single-precision floating-point value. The 1536 | operation is performed according to the IEC/IEEE Standard for Binary 1537 | Floating-Point Arithmetic. 1538 *----------------------------------------------------------------------------*/ 1539 1540 float32 float32_round_to_int( float32 a STATUS_PARAM) 1541 { 1542 flag aSign; 1543 int16 aExp; 1544 bits32 lastBitMask, roundBitsMask; 1545 int8 roundingMode; 1546 bits32 z; 1547 1548 aExp = extractFloat32Exp( a ); 1549 if ( 0x96 <= aExp ) { 1550 if ( ( aExp == 0xFF ) && extractFloat32Frac( a ) ) { 1551 return propagateFloat32NaN( a, a STATUS_VAR ); 1552 } 1553 return a; 1554 } 1555 if ( aExp <= 0x7E ) { 1556 if ( (bits32) ( float32_val(a)<<1 ) == 0 ) return a; 1557 STATUS(float_exception_flags) |= float_flag_inexact; 1558 aSign = extractFloat32Sign( a ); 1559 switch ( STATUS(float_rounding_mode) ) { 1560 case float_round_nearest_even: 1561 if ( ( aExp == 0x7E ) && extractFloat32Frac( a ) ) { 1562 return packFloat32( aSign, 0x7F, 0 ); 1563 } 1564 break; 1565 case float_round_down: 1566 return make_float32(aSign ? 0xBF800000 : 0); 1567 case float_round_up: 1568 return make_float32(aSign ? 0x80000000 : 0x3F800000); 1569 } 1570 return packFloat32( aSign, 0, 0 ); 1571 } 1572 lastBitMask = 1; 1573 lastBitMask <<= 0x96 - aExp; 1574 roundBitsMask = lastBitMask - 1; 1575 z = float32_val(a); 1576 roundingMode = STATUS(float_rounding_mode); 1577 if ( roundingMode == float_round_nearest_even ) { 1578 z += lastBitMask>>1; 1579 if ( ( z & roundBitsMask ) == 0 ) z &= ~ lastBitMask; 1580 } 1581 else if ( roundingMode != float_round_to_zero ) { 1582 if ( extractFloat32Sign( make_float32(z) ) ^ ( roundingMode == float_round_up ) ) { 1583 z += roundBitsMask; 1584 } 1585 } 1586 z &= ~ roundBitsMask; 1587 if ( z != float32_val(a) ) STATUS(float_exception_flags) |= float_flag_inexact; 1588 return make_float32(z); 1589 1590 } 1591 1592 /*---------------------------------------------------------------------------- 1593 | Returns the result of adding the absolute values of the single-precision 1594 | floating-point values `a' and `b'. If `zSign' is 1, the sum is negated 1595 | before being returned. `zSign' is ignored if the result is a NaN. 1596 | The addition is performed according to the IEC/IEEE Standard for Binary 1597 | Floating-Point Arithmetic. 1598 *----------------------------------------------------------------------------*/ 1599 1600 static float32 addFloat32Sigs( float32 a, float32 b, flag zSign STATUS_PARAM) 1601 { 1602 int16 aExp, bExp, zExp; 1603 bits32 aSig, bSig, zSig; 1604 int16 expDiff; 1605 1606 aSig = extractFloat32Frac( a ); 1607 aExp = extractFloat32Exp( a ); 1608 bSig = extractFloat32Frac( b ); 1609 bExp = extractFloat32Exp( b ); 1610 expDiff = aExp - bExp; 1611 aSig <<= 6; 1612 bSig <<= 6; 1613 if ( 0 < expDiff ) { 1614 if ( aExp == 0xFF ) { 1615 if ( aSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1616 return a; 1617 } 1618 if ( bExp == 0 ) { 1619 --expDiff; 1620 } 1621 else { 1622 bSig |= 0x20000000; 1623 } 1624 shift32RightJamming( bSig, expDiff, &bSig ); 1625 zExp = aExp; 1626 } 1627 else if ( expDiff < 0 ) { 1628 if ( bExp == 0xFF ) { 1629 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1630 return packFloat32( zSign, 0xFF, 0 ); 1631 } 1632 if ( aExp == 0 ) { 1633 ++expDiff; 1634 } 1635 else { 1636 aSig |= 0x20000000; 1637 } 1638 shift32RightJamming( aSig, - expDiff, &aSig ); 1639 zExp = bExp; 1640 } 1641 else { 1642 if ( aExp == 0xFF ) { 1643 if ( aSig | bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1644 return a; 1645 } 1646 if ( aExp == 0 ) { 1647 if ( STATUS(flush_to_zero) ) return packFloat32( zSign, 0, 0 ); 1648 return packFloat32( zSign, 0, ( aSig + bSig )>>6 ); 1649 } 1650 zSig = 0x40000000 + aSig + bSig; 1651 zExp = aExp; 1652 goto roundAndPack; 1653 } 1654 aSig |= 0x20000000; 1655 zSig = ( aSig + bSig )<<1; 1656 --zExp; 1657 if ( (sbits32) zSig < 0 ) { 1658 zSig = aSig + bSig; 1659 ++zExp; 1660 } 1661 roundAndPack: 1662 return roundAndPackFloat32( zSign, zExp, zSig STATUS_VAR ); 1663 1664 } 1665 1666 /*---------------------------------------------------------------------------- 1667 | Returns the result of subtracting the absolute values of the single- 1668 | precision floating-point values `a' and `b'. If `zSign' is 1, the 1669 | difference is negated before being returned. `zSign' is ignored if the 1670 | result is a NaN. The subtraction is performed according to the IEC/IEEE 1671 | Standard for Binary Floating-Point Arithmetic. 1672 *----------------------------------------------------------------------------*/ 1673 1674 static float32 subFloat32Sigs( float32 a, float32 b, flag zSign STATUS_PARAM) 1675 { 1676 int16 aExp, bExp, zExp; 1677 bits32 aSig, bSig, zSig; 1678 int16 expDiff; 1679 1680 aSig = extractFloat32Frac( a ); 1681 aExp = extractFloat32Exp( a ); 1682 bSig = extractFloat32Frac( b ); 1683 bExp = extractFloat32Exp( b ); 1684 expDiff = aExp - bExp; 1685 aSig <<= 7; 1686 bSig <<= 7; 1687 if ( 0 < expDiff ) goto aExpBigger; 1688 if ( expDiff < 0 ) goto bExpBigger; 1689 if ( aExp == 0xFF ) { 1690 if ( aSig | bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1691 float_raise( float_flag_invalid STATUS_VAR); 1692 return float32_default_nan; 1693 } 1694 if ( aExp == 0 ) { 1695 aExp = 1; 1696 bExp = 1; 1697 } 1698 if ( bSig < aSig ) goto aBigger; 1699 if ( aSig < bSig ) goto bBigger; 1700 return packFloat32( STATUS(float_rounding_mode) == float_round_down, 0, 0 ); 1701 bExpBigger: 1702 if ( bExp == 0xFF ) { 1703 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1704 return packFloat32( zSign ^ 1, 0xFF, 0 ); 1705 } 1706 if ( aExp == 0 ) { 1707 ++expDiff; 1708 } 1709 else { 1710 aSig |= 0x40000000; 1711 } 1712 shift32RightJamming( aSig, - expDiff, &aSig ); 1713 bSig |= 0x40000000; 1714 bBigger: 1715 zSig = bSig - aSig; 1716 zExp = bExp; 1717 zSign ^= 1; 1718 goto normalizeRoundAndPack; 1719 aExpBigger: 1720 if ( aExp == 0xFF ) { 1721 if ( aSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1722 return a; 1723 } 1724 if ( bExp == 0 ) { 1725 --expDiff; 1726 } 1727 else { 1728 bSig |= 0x40000000; 1729 } 1730 shift32RightJamming( bSig, expDiff, &bSig ); 1731 aSig |= 0x40000000; 1732 aBigger: 1733 zSig = aSig - bSig; 1734 zExp = aExp; 1735 normalizeRoundAndPack: 1736 --zExp; 1737 return normalizeRoundAndPackFloat32( zSign, zExp, zSig STATUS_VAR ); 1738 1739 } 1740 1741 /*---------------------------------------------------------------------------- 1742 | Returns the result of adding the single-precision floating-point values `a' 1743 | and `b'. The operation is performed according to the IEC/IEEE Standard for 1744 | Binary Floating-Point Arithmetic. 1745 *----------------------------------------------------------------------------*/ 1746 1747 float32 float32_add( float32 a, float32 b STATUS_PARAM ) 1748 { 1749 flag aSign, bSign; 1750 1751 aSign = extractFloat32Sign( a ); 1752 bSign = extractFloat32Sign( b ); 1753 if ( aSign == bSign ) { 1754 return addFloat32Sigs( a, b, aSign STATUS_VAR); 1755 } 1756 else { 1757 return subFloat32Sigs( a, b, aSign STATUS_VAR ); 1758 } 1759 1760 } 1761 1762 /*---------------------------------------------------------------------------- 1763 | Returns the result of subtracting the single-precision floating-point values 1764 | `a' and `b'. The operation is performed according to the IEC/IEEE Standard 1765 | for Binary Floating-Point Arithmetic. 1766 *----------------------------------------------------------------------------*/ 1767 1768 float32 float32_sub( float32 a, float32 b STATUS_PARAM ) 1769 { 1770 flag aSign, bSign; 1771 1772 aSign = extractFloat32Sign( a ); 1773 bSign = extractFloat32Sign( b ); 1774 if ( aSign == bSign ) { 1775 return subFloat32Sigs( a, b, aSign STATUS_VAR ); 1776 } 1777 else { 1778 return addFloat32Sigs( a, b, aSign STATUS_VAR ); 1779 } 1780 1781 } 1782 1783 /*---------------------------------------------------------------------------- 1784 | Returns the result of multiplying the single-precision floating-point values 1785 | `a' and `b'. The operation is performed according to the IEC/IEEE Standard 1786 | for Binary Floating-Point Arithmetic. 1787 *----------------------------------------------------------------------------*/ 1788 1789 float32 float32_mul( float32 a, float32 b STATUS_PARAM ) 1790 { 1791 flag aSign, bSign, zSign; 1792 int16 aExp, bExp, zExp; 1793 bits32 aSig, bSig; 1794 bits64 zSig64; 1795 bits32 zSig; 1796 1797 aSig = extractFloat32Frac( a ); 1798 aExp = extractFloat32Exp( a ); 1799 aSign = extractFloat32Sign( a ); 1800 bSig = extractFloat32Frac( b ); 1801 bExp = extractFloat32Exp( b ); 1802 bSign = extractFloat32Sign( b ); 1803 zSign = aSign ^ bSign; 1804 if ( aExp == 0xFF ) { 1805 if ( aSig || ( ( bExp == 0xFF ) && bSig ) ) { 1806 return propagateFloat32NaN( a, b STATUS_VAR ); 1807 } 1808 if ( ( bExp | bSig ) == 0 ) { 1809 float_raise( float_flag_invalid STATUS_VAR); 1810 return float32_default_nan; 1811 } 1812 return packFloat32( zSign, 0xFF, 0 ); 1813 } 1814 if ( bExp == 0xFF ) { 1815 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1816 if ( ( aExp | aSig ) == 0 ) { 1817 float_raise( float_flag_invalid STATUS_VAR); 1818 return float32_default_nan; 1819 } 1820 return packFloat32( zSign, 0xFF, 0 ); 1821 } 1822 if ( aExp == 0 ) { 1823 if ( aSig == 0 ) return packFloat32( zSign, 0, 0 ); 1824 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1825 } 1826 if ( bExp == 0 ) { 1827 if ( bSig == 0 ) return packFloat32( zSign, 0, 0 ); 1828 normalizeFloat32Subnormal( bSig, &bExp, &bSig ); 1829 } 1830 zExp = aExp + bExp - 0x7F; 1831 aSig = ( aSig | 0x00800000 )<<7; 1832 bSig = ( bSig | 0x00800000 )<<8; 1833 shift64RightJamming( ( (bits64) aSig ) * bSig, 32, &zSig64 ); 1834 zSig = zSig64; 1835 if ( 0 <= (sbits32) ( zSig<<1 ) ) { 1836 zSig <<= 1; 1837 --zExp; 1838 } 1839 return roundAndPackFloat32( zSign, zExp, zSig STATUS_VAR ); 1840 1841 } 1842 1843 /*---------------------------------------------------------------------------- 1844 | Returns the result of dividing the single-precision floating-point value `a' 1845 | by the corresponding value `b'. The operation is performed according to the 1846 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1847 *----------------------------------------------------------------------------*/ 1848 1849 float32 float32_div( float32 a, float32 b STATUS_PARAM ) 1850 { 1851 flag aSign, bSign, zSign; 1852 int16 aExp, bExp, zExp; 1853 bits32 aSig, bSig, zSig; 1854 1855 aSig = extractFloat32Frac( a ); 1856 aExp = extractFloat32Exp( a ); 1857 aSign = extractFloat32Sign( a ); 1858 bSig = extractFloat32Frac( b ); 1859 bExp = extractFloat32Exp( b ); 1860 bSign = extractFloat32Sign( b ); 1861 zSign = aSign ^ bSign; 1862 if ( aExp == 0xFF ) { 1863 if ( aSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1864 if ( bExp == 0xFF ) { 1865 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1866 float_raise( float_flag_invalid STATUS_VAR); 1867 return float32_default_nan; 1868 } 1869 return packFloat32( zSign, 0xFF, 0 ); 1870 } 1871 if ( bExp == 0xFF ) { 1872 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1873 return packFloat32( zSign, 0, 0 ); 1874 } 1875 if ( bExp == 0 ) { 1876 if ( bSig == 0 ) { 1877 if ( ( aExp | aSig ) == 0 ) { 1878 float_raise( float_flag_invalid STATUS_VAR); 1879 return float32_default_nan; 1880 } 1881 float_raise( float_flag_divbyzero STATUS_VAR); 1882 return packFloat32( zSign, 0xFF, 0 ); 1883 } 1884 normalizeFloat32Subnormal( bSig, &bExp, &bSig ); 1885 } 1886 if ( aExp == 0 ) { 1887 if ( aSig == 0 ) return packFloat32( zSign, 0, 0 ); 1888 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1889 } 1890 zExp = aExp - bExp + 0x7D; 1891 aSig = ( aSig | 0x00800000 )<<7; 1892 bSig = ( bSig | 0x00800000 )<<8; 1893 if ( bSig <= ( aSig + aSig ) ) { 1894 aSig >>= 1; 1895 ++zExp; 1896 } 1897 zSig = ( ( (bits64) aSig )<<32 ) / bSig; 1898 if ( ( zSig & 0x3F ) == 0 ) { 1899 zSig |= ( (bits64) bSig * zSig != ( (bits64) aSig )<<32 ); 1900 } 1901 return roundAndPackFloat32( zSign, zExp, zSig STATUS_VAR ); 1902 1903 } 1904 1905 /*---------------------------------------------------------------------------- 1906 | Returns the remainder of the single-precision floating-point value `a' 1907 | with respect to the corresponding value `b'. The operation is performed 1908 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 1909 *----------------------------------------------------------------------------*/ 1910 1911 float32 float32_rem( float32 a, float32 b STATUS_PARAM ) 1912 { 1913 flag aSign, bSign, zSign; 1914 int16 aExp, bExp, expDiff; 1915 bits32 aSig, bSig; 1916 bits32 q; 1917 bits64 aSig64, bSig64, q64; 1918 bits32 alternateASig; 1919 sbits32 sigMean; 1920 1921 aSig = extractFloat32Frac( a ); 1922 aExp = extractFloat32Exp( a ); 1923 aSign = extractFloat32Sign( a ); 1924 bSig = extractFloat32Frac( b ); 1925 bExp = extractFloat32Exp( b ); 1926 bSign = extractFloat32Sign( b ); 1927 if ( aExp == 0xFF ) { 1928 if ( aSig || ( ( bExp == 0xFF ) && bSig ) ) { 1929 return propagateFloat32NaN( a, b STATUS_VAR ); 1930 } 1931 float_raise( float_flag_invalid STATUS_VAR); 1932 return float32_default_nan; 1933 } 1934 if ( bExp == 0xFF ) { 1935 if ( bSig ) return propagateFloat32NaN( a, b STATUS_VAR ); 1936 return a; 1937 } 1938 if ( bExp == 0 ) { 1939 if ( bSig == 0 ) { 1940 float_raise( float_flag_invalid STATUS_VAR); 1941 return float32_default_nan; 1942 } 1943 normalizeFloat32Subnormal( bSig, &bExp, &bSig ); 1944 } 1945 if ( aExp == 0 ) { 1946 if ( aSig == 0 ) return a; 1947 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 1948 } 1949 expDiff = aExp - bExp; 1950 aSig |= 0x00800000; 1951 bSig |= 0x00800000; 1952 if ( expDiff < 32 ) { 1953 aSig <<= 8; 1954 bSig <<= 8; 1955 if ( expDiff < 0 ) { 1956 if ( expDiff < -1 ) return a; 1957 aSig >>= 1; 1958 } 1959 q = ( bSig <= aSig ); 1960 if ( q ) aSig -= bSig; 1961 if ( 0 < expDiff ) { 1962 q = ( ( (bits64) aSig )<<32 ) / bSig; 1963 q >>= 32 - expDiff; 1964 bSig >>= 2; 1965 aSig = ( ( aSig>>1 )<<( expDiff - 1 ) ) - bSig * q; 1966 } 1967 else { 1968 aSig >>= 2; 1969 bSig >>= 2; 1970 } 1971 } 1972 else { 1973 if ( bSig <= aSig ) aSig -= bSig; 1974 aSig64 = ( (bits64) aSig )<<40; 1975 bSig64 = ( (bits64) bSig )<<40; 1976 expDiff -= 64; 1977 while ( 0 < expDiff ) { 1978 q64 = estimateDiv128To64( aSig64, 0, bSig64 ); 1979 q64 = ( 2 < q64 ) ? q64 - 2 : 0; 1980 aSig64 = - ( ( bSig * q64 )<<38 ); 1981 expDiff -= 62; 1982 } 1983 expDiff += 64; 1984 q64 = estimateDiv128To64( aSig64, 0, bSig64 ); 1985 q64 = ( 2 < q64 ) ? q64 - 2 : 0; 1986 q = q64>>( 64 - expDiff ); 1987 bSig <<= 6; 1988 aSig = ( ( aSig64>>33 )<<( expDiff - 1 ) ) - bSig * q; 1989 } 1990 do { 1991 alternateASig = aSig; 1992 ++q; 1993 aSig -= bSig; 1994 } while ( 0 <= (sbits32) aSig ); 1995 sigMean = aSig + alternateASig; 1996 if ( ( sigMean < 0 ) || ( ( sigMean == 0 ) && ( q & 1 ) ) ) { 1997 aSig = alternateASig; 1998 } 1999 zSign = ( (sbits32) aSig < 0 ); 2000 if ( zSign ) aSig = - aSig; 2001 return normalizeRoundAndPackFloat32( aSign ^ zSign, bExp, aSig STATUS_VAR ); 2002 2003 } 2004 2005 /*---------------------------------------------------------------------------- 2006 | Returns the square root of the single-precision floating-point value `a'. 2007 | The operation is performed according to the IEC/IEEE Standard for Binary 2008 | Floating-Point Arithmetic. 2009 *----------------------------------------------------------------------------*/ 2010 2011 float32 float32_sqrt( float32 a STATUS_PARAM ) 2012 { 2013 flag aSign; 2014 int16 aExp, zExp; 2015 bits32 aSig, zSig; 2016 bits64 rem, term; 2017 2018 aSig = extractFloat32Frac( a ); 2019 aExp = extractFloat32Exp( a ); 2020 aSign = extractFloat32Sign( a ); 2021 if ( aExp == 0xFF ) { 2022 if ( aSig ) return propagateFloat32NaN( a, float32_zero STATUS_VAR ); 2023 if ( ! aSign ) return a; 2024 float_raise( float_flag_invalid STATUS_VAR); 2025 return float32_default_nan; 2026 } 2027 if ( aSign ) { 2028 if ( ( aExp | aSig ) == 0 ) return a; 2029 float_raise( float_flag_invalid STATUS_VAR); 2030 return float32_default_nan; 2031 } 2032 if ( aExp == 0 ) { 2033 if ( aSig == 0 ) return float32_zero; 2034 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 2035 } 2036 zExp = ( ( aExp - 0x7F )>>1 ) + 0x7E; 2037 aSig = ( aSig | 0x00800000 )<<8; 2038 zSig = estimateSqrt32( aExp, aSig ) + 2; 2039 if ( ( zSig & 0x7F ) <= 5 ) { 2040 if ( zSig < 2 ) { 2041 zSig = 0x7FFFFFFF; 2042 goto roundAndPack; 2043 } 2044 aSig >>= aExp & 1; 2045 term = ( (bits64) zSig ) * zSig; 2046 rem = ( ( (bits64) aSig )<<32 ) - term; 2047 while ( (sbits64) rem < 0 ) { 2048 --zSig; 2049 rem += ( ( (bits64) zSig )<<1 ) | 1; 2050 } 2051 zSig |= ( rem != 0 ); 2052 } 2053 shift32RightJamming( zSig, 1, &zSig ); 2054 roundAndPack: 2055 return roundAndPackFloat32( 0, zExp, zSig STATUS_VAR ); 2056 2057 } 2058 2059 /*---------------------------------------------------------------------------- 2060 | Returns the binary log of the single-precision floating-point value `a'. 2061 | The operation is performed according to the IEC/IEEE Standard for Binary 2062 | Floating-Point Arithmetic. 2063 *----------------------------------------------------------------------------*/ 2064 float32 float32_log2( float32 a STATUS_PARAM ) 2065 { 2066 flag aSign, zSign; 2067 int16 aExp; 2068 bits32 aSig, zSig, i; 2069 2070 aSig = extractFloat32Frac( a ); 2071 aExp = extractFloat32Exp( a ); 2072 aSign = extractFloat32Sign( a ); 2073 2074 if ( aExp == 0 ) { 2075 if ( aSig == 0 ) return packFloat32( 1, 0xFF, 0 ); 2076 normalizeFloat32Subnormal( aSig, &aExp, &aSig ); 2077 } 2078 if ( aSign ) { 2079 float_raise( float_flag_invalid STATUS_VAR); 2080 return float32_default_nan; 2081 } 2082 if ( aExp == 0xFF ) { 2083 if ( aSig ) return propagateFloat32NaN( a, float32_zero STATUS_VAR ); 2084 return a; 2085 } 2086 2087 aExp -= 0x7F; 2088 aSig |= 0x00800000; 2089 zSign = aExp < 0; 2090 zSig = aExp << 23; 2091 2092 for (i = 1 << 22; i > 0; i >>= 1) { 2093 aSig = ( (bits64)aSig * aSig ) >> 23; 2094 if ( aSig & 0x01000000 ) { 2095 aSig >>= 1; 2096 zSig |= i; 2097 } 2098 } 2099 2100 if ( zSign ) 2101 zSig = -zSig; 2102 2103 return normalizeRoundAndPackFloat32( zSign, 0x85, zSig STATUS_VAR ); 2104 } 2105 2106 /*---------------------------------------------------------------------------- 2107 | Returns 1 if the single-precision floating-point value `a' is equal to 2108 | the corresponding value `b', and 0 otherwise. The comparison is performed 2109 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 2110 *----------------------------------------------------------------------------*/ 2111 2112 int float32_eq( float32 a, float32 b STATUS_PARAM ) 2113 { 2114 2115 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2116 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2117 ) { 2118 if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) { 2119 float_raise( float_flag_invalid STATUS_VAR); 2120 } 2121 return 0; 2122 } 2123 return ( float32_val(a) == float32_val(b) ) || 2124 ( (bits32) ( ( float32_val(a) | float32_val(b) )<<1 ) == 0 ); 2125 2126 } 2127 2128 /*---------------------------------------------------------------------------- 2129 | Returns 1 if the single-precision floating-point value `a' is less than 2130 | or equal to the corresponding value `b', and 0 otherwise. The comparison 2131 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 2132 | Arithmetic. 2133 *----------------------------------------------------------------------------*/ 2134 2135 int float32_le( float32 a, float32 b STATUS_PARAM ) 2136 { 2137 flag aSign, bSign; 2138 bits32 av, bv; 2139 2140 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2141 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2142 ) { 2143 float_raise( float_flag_invalid STATUS_VAR); 2144 return 0; 2145 } 2146 aSign = extractFloat32Sign( a ); 2147 bSign = extractFloat32Sign( b ); 2148 av = float32_val(a); 2149 bv = float32_val(b); 2150 if ( aSign != bSign ) return aSign || ( (bits32) ( ( av | bv )<<1 ) == 0 ); 2151 return ( av == bv ) || ( aSign ^ ( av < bv ) ); 2152 2153 } 2154 2155 /*---------------------------------------------------------------------------- 2156 | Returns 1 if the single-precision floating-point value `a' is less than 2157 | the corresponding value `b', and 0 otherwise. The comparison is performed 2158 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 2159 *----------------------------------------------------------------------------*/ 2160 2161 int float32_lt( float32 a, float32 b STATUS_PARAM ) 2162 { 2163 flag aSign, bSign; 2164 bits32 av, bv; 2165 2166 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2167 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2168 ) { 2169 float_raise( float_flag_invalid STATUS_VAR); 2170 return 0; 2171 } 2172 aSign = extractFloat32Sign( a ); 2173 bSign = extractFloat32Sign( b ); 2174 av = float32_val(a); 2175 bv = float32_val(b); 2176 if ( aSign != bSign ) return aSign && ( (bits32) ( ( av | bv )<<1 ) != 0 ); 2177 return ( av != bv ) && ( aSign ^ ( av < bv ) ); 2178 2179 } 2180 2181 /*---------------------------------------------------------------------------- 2182 | Returns 1 if the single-precision floating-point value `a' is equal to 2183 | the corresponding value `b', and 0 otherwise. The invalid exception is 2184 | raised if either operand is a NaN. Otherwise, the comparison is performed 2185 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 2186 *----------------------------------------------------------------------------*/ 2187 2188 int float32_eq_signaling( float32 a, float32 b STATUS_PARAM ) 2189 { 2190 bits32 av, bv; 2191 2192 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2193 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2194 ) { 2195 float_raise( float_flag_invalid STATUS_VAR); 2196 return 0; 2197 } 2198 av = float32_val(a); 2199 bv = float32_val(b); 2200 return ( av == bv ) || ( (bits32) ( ( av | bv )<<1 ) == 0 ); 2201 2202 } 2203 2204 /*---------------------------------------------------------------------------- 2205 | Returns 1 if the single-precision floating-point value `a' is less than or 2206 | equal to the corresponding value `b', and 0 otherwise. Quiet NaNs do not 2207 | cause an exception. Otherwise, the comparison is performed according to the 2208 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 2209 *----------------------------------------------------------------------------*/ 2210 2211 int float32_le_quiet( float32 a, float32 b STATUS_PARAM ) 2212 { 2213 flag aSign, bSign; 2214 bits32 av, bv; 2215 2216 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2217 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2218 ) { 2219 if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) { 2220 float_raise( float_flag_invalid STATUS_VAR); 2221 } 2222 return 0; 2223 } 2224 aSign = extractFloat32Sign( a ); 2225 bSign = extractFloat32Sign( b ); 2226 av = float32_val(a); 2227 bv = float32_val(b); 2228 if ( aSign != bSign ) return aSign || ( (bits32) ( ( av | bv )<<1 ) == 0 ); 2229 return ( av == bv ) || ( aSign ^ ( av < bv ) ); 2230 2231 } 2232 2233 /*---------------------------------------------------------------------------- 2234 | Returns 1 if the single-precision floating-point value `a' is less than 2235 | the corresponding value `b', and 0 otherwise. Quiet NaNs do not cause an 2236 | exception. Otherwise, the comparison is performed according to the IEC/IEEE 2237 | Standard for Binary Floating-Point Arithmetic. 2238 *----------------------------------------------------------------------------*/ 2239 2240 int float32_lt_quiet( float32 a, float32 b STATUS_PARAM ) 2241 { 2242 flag aSign, bSign; 2243 bits32 av, bv; 2244 2245 if ( ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) ) 2246 || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) ) 2247 ) { 2248 if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) { 2249 float_raise( float_flag_invalid STATUS_VAR); 2250 } 2251 return 0; 2252 } 2253 aSign = extractFloat32Sign( a ); 2254 bSign = extractFloat32Sign( b ); 2255 av = float32_val(a); 2256 bv = float32_val(b); 2257 if ( aSign != bSign ) return aSign && ( (bits32) ( ( av | bv )<<1 ) != 0 ); 2258 return ( av != bv ) && ( aSign ^ ( av < bv ) ); 2259 2260 } 2261 2262 /*---------------------------------------------------------------------------- 2263 | Returns the result of converting the double-precision floating-point value 2264 | `a' to the 32-bit two's complement integer format. The conversion is 2265 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2266 | Arithmetic---which means in particular that the conversion is rounded 2267 | according to the current rounding mode. If `a' is a NaN, the largest 2268 | positive integer is returned. Otherwise, if the conversion overflows, the 2269 | largest integer with the same sign as `a' is returned. 2270 *----------------------------------------------------------------------------*/ 2271 2272 int32 float64_to_int32( float64 a STATUS_PARAM ) 2273 { 2274 flag aSign; 2275 int16 aExp, shiftCount; 2276 bits64 aSig; 2277 2278 aSig = extractFloat64Frac( a ); 2279 aExp = extractFloat64Exp( a ); 2280 aSign = extractFloat64Sign( a ); 2281 if ( ( aExp == 0x7FF ) && aSig ) aSign = 0; 2282 if ( aExp ) aSig |= LIT64( 0x0010000000000000 ); 2283 shiftCount = 0x42C - aExp; 2284 if ( 0 < shiftCount ) shift64RightJamming( aSig, shiftCount, &aSig ); 2285 return roundAndPackInt32( aSign, aSig STATUS_VAR ); 2286 2287 } 2288 2289 /*---------------------------------------------------------------------------- 2290 | Returns the result of converting the double-precision floating-point value 2291 | `a' to the 32-bit two's complement integer format. The conversion is 2292 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2293 | Arithmetic, except that the conversion is always rounded toward zero. 2294 | If `a' is a NaN, the largest positive integer is returned. Otherwise, if 2295 | the conversion overflows, the largest integer with the same sign as `a' is 2296 | returned. 2297 *----------------------------------------------------------------------------*/ 2298 2299 int32 float64_to_int32_round_to_zero( float64 a STATUS_PARAM ) 2300 { 2301 flag aSign; 2302 int16 aExp, shiftCount; 2303 bits64 aSig, savedASig; 2304 int32 z; 2305 2306 aSig = extractFloat64Frac( a ); 2307 aExp = extractFloat64Exp( a ); 2308 aSign = extractFloat64Sign( a ); 2309 if ( 0x41E < aExp ) { 2310 if ( ( aExp == 0x7FF ) && aSig ) aSign = 0; 2311 goto invalid; 2312 } 2313 else if ( aExp < 0x3FF ) { 2314 if ( aExp || aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 2315 return 0; 2316 } 2317 aSig |= LIT64( 0x0010000000000000 ); 2318 shiftCount = 0x433 - aExp; 2319 savedASig = aSig; 2320 aSig >>= shiftCount; 2321 z = aSig; 2322 if ( aSign ) z = - z; 2323 if ( ( z < 0 ) ^ aSign ) { 2324 invalid: 2325 float_raise( float_flag_invalid STATUS_VAR); 2326 return aSign ? (sbits32) 0x80000000 : 0x7FFFFFFF; 2327 } 2328 if ( ( aSig<<shiftCount ) != savedASig ) { 2329 STATUS(float_exception_flags) |= float_flag_inexact; 2330 } 2331 return z; 2332 2333 } 2334 2335 /*---------------------------------------------------------------------------- 2336 | Returns the result of converting the double-precision floating-point value 2337 | `a' to the 64-bit two's complement integer format. The conversion is 2338 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2339 | Arithmetic---which means in particular that the conversion is rounded 2340 | according to the current rounding mode. If `a' is a NaN, the largest 2341 | positive integer is returned. Otherwise, if the conversion overflows, the 2342 | largest integer with the same sign as `a' is returned. 2343 *----------------------------------------------------------------------------*/ 2344 2345 int64 float64_to_int64( float64 a STATUS_PARAM ) 2346 { 2347 flag aSign; 2348 int16 aExp, shiftCount; 2349 bits64 aSig, aSigExtra; 2350 2351 aSig = extractFloat64Frac( a ); 2352 aExp = extractFloat64Exp( a ); 2353 aSign = extractFloat64Sign( a ); 2354 if ( aExp ) aSig |= LIT64( 0x0010000000000000 ); 2355 shiftCount = 0x433 - aExp; 2356 if ( shiftCount <= 0 ) { 2357 if ( 0x43E < aExp ) { 2358 float_raise( float_flag_invalid STATUS_VAR); 2359 if ( ! aSign 2360 || ( ( aExp == 0x7FF ) 2361 && ( aSig != LIT64( 0x0010000000000000 ) ) ) 2362 ) { 2363 return LIT64( 0x7FFFFFFFFFFFFFFF ); 2364 } 2365 return (sbits64) LIT64( 0x8000000000000000 ); 2366 } 2367 aSigExtra = 0; 2368 aSig <<= - shiftCount; 2369 } 2370 else { 2371 shift64ExtraRightJamming( aSig, 0, shiftCount, &aSig, &aSigExtra ); 2372 } 2373 return roundAndPackInt64( aSign, aSig, aSigExtra STATUS_VAR ); 2374 2375 } 2376 2377 /*---------------------------------------------------------------------------- 2378 | Returns the result of converting the double-precision floating-point value 2379 | `a' to the 64-bit two's complement integer format. The conversion is 2380 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2381 | Arithmetic, except that the conversion is always rounded toward zero. 2382 | If `a' is a NaN, the largest positive integer is returned. Otherwise, if 2383 | the conversion overflows, the largest integer with the same sign as `a' is 2384 | returned. 2385 *----------------------------------------------------------------------------*/ 2386 2387 int64 float64_to_int64_round_to_zero( float64 a STATUS_PARAM ) 2388 { 2389 flag aSign; 2390 int16 aExp, shiftCount; 2391 bits64 aSig; 2392 int64 z; 2393 2394 aSig = extractFloat64Frac( a ); 2395 aExp = extractFloat64Exp( a ); 2396 aSign = extractFloat64Sign( a ); 2397 if ( aExp ) aSig |= LIT64( 0x0010000000000000 ); 2398 shiftCount = aExp - 0x433; 2399 if ( 0 <= shiftCount ) { 2400 if ( 0x43E <= aExp ) { 2401 if ( float64_val(a) != LIT64( 0xC3E0000000000000 ) ) { 2402 float_raise( float_flag_invalid STATUS_VAR); 2403 if ( ! aSign 2404 || ( ( aExp == 0x7FF ) 2405 && ( aSig != LIT64( 0x0010000000000000 ) ) ) 2406 ) { 2407 return LIT64( 0x7FFFFFFFFFFFFFFF ); 2408 } 2409 } 2410 return (sbits64) LIT64( 0x8000000000000000 ); 2411 } 2412 z = aSig<<shiftCount; 2413 } 2414 else { 2415 if ( aExp < 0x3FE ) { 2416 if ( aExp | aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 2417 return 0; 2418 } 2419 z = aSig>>( - shiftCount ); 2420 if ( (bits64) ( aSig<<( shiftCount & 63 ) ) ) { 2421 STATUS(float_exception_flags) |= float_flag_inexact; 2422 } 2423 } 2424 if ( aSign ) z = - z; 2425 return z; 2426 2427 } 2428 2429 /*---------------------------------------------------------------------------- 2430 | Returns the result of converting the double-precision floating-point value 2431 | `a' to the single-precision floating-point format. The conversion is 2432 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2433 | Arithmetic. 2434 *----------------------------------------------------------------------------*/ 2435 2436 float32 float64_to_float32( float64 a STATUS_PARAM ) 2437 { 2438 flag aSign; 2439 int16 aExp; 2440 bits64 aSig; 2441 bits32 zSig; 2442 2443 aSig = extractFloat64Frac( a ); 2444 aExp = extractFloat64Exp( a ); 2445 aSign = extractFloat64Sign( a ); 2446 if ( aExp == 0x7FF ) { 2447 if ( aSig ) return commonNaNToFloat32( float64ToCommonNaN( a STATUS_VAR ) ); 2448 return packFloat32( aSign, 0xFF, 0 ); 2449 } 2450 shift64RightJamming( aSig, 22, &aSig ); 2451 zSig = aSig; 2452 if ( aExp || zSig ) { 2453 zSig |= 0x40000000; 2454 aExp -= 0x381; 2455 } 2456 return roundAndPackFloat32( aSign, aExp, zSig STATUS_VAR ); 2457 2458 } 2459 2460 2461 /*---------------------------------------------------------------------------- 2462 | Packs the sign `zSign', exponent `zExp', and significand `zSig' into a 2463 | half-precision floating-point value, returning the result. After being 2464 | shifted into the proper positions, the three fields are simply added 2465 | together to form the result. This means that any integer portion of `zSig' 2466 | will be added into the exponent. Since a properly normalized significand 2467 | will have an integer portion equal to 1, the `zExp' input should be 1 less 2468 | than the desired result exponent whenever `zSig' is a complete, normalized 2469 | significand. 2470 *----------------------------------------------------------------------------*/ 2471 static bits16 packFloat16(flag zSign, int16 zExp, bits16 zSig) 2472 { 2473 return (((bits32)zSign) << 15) + (((bits32)zExp) << 10) + zSig; 2474 } 2475 2476 /* Half precision floats come in two formats: standard IEEE and "ARM" format. 2477 The latter gains extra exponent range by omitting the NaN/Inf encodings. */ 2478 2479 float32 float16_to_float32( bits16 a, flag ieee STATUS_PARAM ) 2480 { 2481 flag aSign; 2482 int16 aExp; 2483 bits32 aSig; 2484 2485 aSign = a >> 15; 2486 aExp = (a >> 10) & 0x1f; 2487 aSig = a & 0x3ff; 2488 2489 if (aExp == 0x1f && ieee) { 2490 if (aSig) { 2491 /* Make sure correct exceptions are raised. */ 2492 float32ToCommonNaN(a STATUS_VAR); 2493 aSig |= 0x200; 2494 } 2495 return packFloat32(aSign, 0xff, aSig << 13); 2496 } 2497 if (aExp == 0) { 2498 int8 shiftCount; 2499 2500 if (aSig == 0) { 2501 return packFloat32(aSign, 0, 0); 2502 } 2503 2504 shiftCount = countLeadingZeros32( aSig ) - 21; 2505 aSig = aSig << shiftCount; 2506 aExp = -shiftCount; 2507 } 2508 return packFloat32( aSign, aExp + 0x70, aSig << 13); 2509 } 2510 2511 bits16 float32_to_float16( float32 a, flag ieee STATUS_PARAM) 2512 { 2513 flag aSign; 2514 int16 aExp; 2515 bits32 aSig; 2516 bits32 mask; 2517 bits32 increment; 2518 int8 roundingMode; 2519 2520 aSig = extractFloat32Frac( a ); 2521 aExp = extractFloat32Exp( a ); 2522 aSign = extractFloat32Sign( a ); 2523 if ( aExp == 0xFF ) { 2524 if (aSig) { 2525 /* Make sure correct exceptions are raised. */ 2526 float32ToCommonNaN(a STATUS_VAR); 2527 aSig |= 0x00400000; 2528 } 2529 return packFloat16(aSign, 0x1f, aSig >> 13); 2530 } 2531 if (aExp == 0 && aSign == 0) { 2532 return packFloat16(aSign, 0, 0); 2533 } 2534 /* Decimal point between bits 22 and 23. */ 2535 aSig |= 0x00800000; 2536 aExp -= 0x7f; 2537 if (aExp < -14) { 2538 mask = 0x007fffff; 2539 if (aExp < -24) { 2540 aExp = -25; 2541 } else { 2542 mask >>= 24 + aExp; 2543 } 2544 } else { 2545 mask = 0x00001fff; 2546 } 2547 if (aSig & mask) { 2548 float_raise( float_flag_underflow STATUS_VAR ); 2549 roundingMode = STATUS(float_rounding_mode); 2550 switch (roundingMode) { 2551 case float_round_nearest_even: 2552 increment = (mask + 1) >> 1; 2553 if ((aSig & mask) == increment) { 2554 increment = aSig & (increment << 1); 2555 } 2556 break; 2557 case float_round_up: 2558 increment = aSign ? 0 : mask; 2559 break; 2560 case float_round_down: 2561 increment = aSign ? mask : 0; 2562 break; 2563 default: /* round_to_zero */ 2564 increment = 0; 2565 break; 2566 } 2567 aSig += increment; 2568 if (aSig >= 0x01000000) { 2569 aSig >>= 1; 2570 aExp++; 2571 } 2572 } else if (aExp < -14 2573 && STATUS(float_detect_tininess) == float_tininess_before_rounding) { 2574 float_raise( float_flag_underflow STATUS_VAR); 2575 } 2576 2577 if (ieee) { 2578 if (aExp > 15) { 2579 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 2580 return packFloat16(aSign, 0x1f, 0); 2581 } 2582 } else { 2583 if (aExp > 16) { 2584 float_raise( float_flag_overflow | float_flag_inexact STATUS_VAR); 2585 return packFloat16(aSign, 0x1f, 0x3ff); 2586 } 2587 } 2588 if (aExp < -24) { 2589 return packFloat16(aSign, 0, 0); 2590 } 2591 if (aExp < -14) { 2592 aSig >>= -14 - aExp; 2593 aExp = -14; 2594 } 2595 return packFloat16(aSign, aExp + 14, aSig >> 13); 2596 } 2597 2598 #ifdef FLOATX80 2599 2600 /*---------------------------------------------------------------------------- 2601 | Returns the result of converting the double-precision floating-point value 2602 | `a' to the extended double-precision floating-point format. The conversion 2603 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 2604 | Arithmetic. 2605 *----------------------------------------------------------------------------*/ 2606 2607 floatx80 float64_to_floatx80( float64 a STATUS_PARAM ) 2608 { 2609 flag aSign; 2610 int16 aExp; 2611 bits64 aSig; 2612 2613 aSig = extractFloat64Frac( a ); 2614 aExp = extractFloat64Exp( a ); 2615 aSign = extractFloat64Sign( a ); 2616 if ( aExp == 0x7FF ) { 2617 if ( aSig ) return commonNaNToFloatx80( float64ToCommonNaN( a STATUS_VAR ) ); 2618 return packFloatx80( aSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 2619 } 2620 if ( aExp == 0 ) { 2621 if ( aSig == 0 ) return packFloatx80( aSign, 0, 0 ); 2622 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 2623 } 2624 return 2625 packFloatx80( 2626 aSign, aExp + 0x3C00, ( aSig | LIT64( 0x0010000000000000 ) )<<11 ); 2627 2628 } 2629 2630 #endif 2631 2632 #ifdef FLOAT128 2633 2634 /*---------------------------------------------------------------------------- 2635 | Returns the result of converting the double-precision floating-point value 2636 | `a' to the quadruple-precision floating-point format. The conversion is 2637 | performed according to the IEC/IEEE Standard for Binary Floating-Point 2638 | Arithmetic. 2639 *----------------------------------------------------------------------------*/ 2640 2641 float128 float64_to_float128( float64 a STATUS_PARAM ) 2642 { 2643 flag aSign; 2644 int16 aExp; 2645 bits64 aSig, zSig0, zSig1; 2646 2647 aSig = extractFloat64Frac( a ); 2648 aExp = extractFloat64Exp( a ); 2649 aSign = extractFloat64Sign( a ); 2650 if ( aExp == 0x7FF ) { 2651 if ( aSig ) return commonNaNToFloat128( float64ToCommonNaN( a STATUS_VAR ) ); 2652 return packFloat128( aSign, 0x7FFF, 0, 0 ); 2653 } 2654 if ( aExp == 0 ) { 2655 if ( aSig == 0 ) return packFloat128( aSign, 0, 0, 0 ); 2656 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 2657 --aExp; 2658 } 2659 shift128Right( aSig, 0, 4, &zSig0, &zSig1 ); 2660 return packFloat128( aSign, aExp + 0x3C00, zSig0, zSig1 ); 2661 2662 } 2663 2664 #endif 2665 2666 /*---------------------------------------------------------------------------- 2667 | Rounds the double-precision floating-point value `a' to an integer, and 2668 | returns the result as a double-precision floating-point value. The 2669 | operation is performed according to the IEC/IEEE Standard for Binary 2670 | Floating-Point Arithmetic. 2671 *----------------------------------------------------------------------------*/ 2672 2673 float64 float64_round_to_int( float64 a STATUS_PARAM ) 2674 { 2675 flag aSign; 2676 int16 aExp; 2677 bits64 lastBitMask, roundBitsMask; 2678 int8 roundingMode; 2679 bits64 z; 2680 2681 aExp = extractFloat64Exp( a ); 2682 if ( 0x433 <= aExp ) { 2683 if ( ( aExp == 0x7FF ) && extractFloat64Frac( a ) ) { 2684 return propagateFloat64NaN( a, a STATUS_VAR ); 2685 } 2686 return a; 2687 } 2688 if ( aExp < 0x3FF ) { 2689 if ( (bits64) ( float64_val(a)<<1 ) == 0 ) return a; 2690 STATUS(float_exception_flags) |= float_flag_inexact; 2691 aSign = extractFloat64Sign( a ); 2692 switch ( STATUS(float_rounding_mode) ) { 2693 case float_round_nearest_even: 2694 if ( ( aExp == 0x3FE ) && extractFloat64Frac( a ) ) { 2695 return packFloat64( aSign, 0x3FF, 0 ); 2696 } 2697 break; 2698 case float_round_down: 2699 return make_float64(aSign ? LIT64( 0xBFF0000000000000 ) : 0); 2700 case float_round_up: 2701 return make_float64( 2702 aSign ? LIT64( 0x8000000000000000 ) : LIT64( 0x3FF0000000000000 )); 2703 } 2704 return packFloat64( aSign, 0, 0 ); 2705 } 2706 lastBitMask = 1; 2707 lastBitMask <<= 0x433 - aExp; 2708 roundBitsMask = lastBitMask - 1; 2709 z = float64_val(a); 2710 roundingMode = STATUS(float_rounding_mode); 2711 if ( roundingMode == float_round_nearest_even ) { 2712 z += lastBitMask>>1; 2713 if ( ( z & roundBitsMask ) == 0 ) z &= ~ lastBitMask; 2714 } 2715 else if ( roundingMode != float_round_to_zero ) { 2716 if ( extractFloat64Sign( make_float64(z) ) ^ ( roundingMode == float_round_up ) ) { 2717 z += roundBitsMask; 2718 } 2719 } 2720 z &= ~ roundBitsMask; 2721 if ( z != float64_val(a) ) 2722 STATUS(float_exception_flags) |= float_flag_inexact; 2723 return make_float64(z); 2724 2725 } 2726 2727 float64 float64_trunc_to_int( float64 a STATUS_PARAM) 2728 { 2729 int oldmode; 2730 float64 res; 2731 oldmode = STATUS(float_rounding_mode); 2732 STATUS(float_rounding_mode) = float_round_to_zero; 2733 res = float64_round_to_int(a STATUS_VAR); 2734 STATUS(float_rounding_mode) = oldmode; 2735 return res; 2736 } 2737 2738 /*---------------------------------------------------------------------------- 2739 | Returns the result of adding the absolute values of the double-precision 2740 | floating-point values `a' and `b'. If `zSign' is 1, the sum is negated 2741 | before being returned. `zSign' is ignored if the result is a NaN. 2742 | The addition is performed according to the IEC/IEEE Standard for Binary 2743 | Floating-Point Arithmetic. 2744 *----------------------------------------------------------------------------*/ 2745 2746 static float64 addFloat64Sigs( float64 a, float64 b, flag zSign STATUS_PARAM ) 2747 { 2748 int16 aExp, bExp, zExp; 2749 bits64 aSig, bSig, zSig; 2750 int16 expDiff; 2751 2752 aSig = extractFloat64Frac( a ); 2753 aExp = extractFloat64Exp( a ); 2754 bSig = extractFloat64Frac( b ); 2755 bExp = extractFloat64Exp( b ); 2756 expDiff = aExp - bExp; 2757 aSig <<= 9; 2758 bSig <<= 9; 2759 if ( 0 < expDiff ) { 2760 if ( aExp == 0x7FF ) { 2761 if ( aSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2762 return a; 2763 } 2764 if ( bExp == 0 ) { 2765 --expDiff; 2766 } 2767 else { 2768 bSig |= LIT64( 0x2000000000000000 ); 2769 } 2770 shift64RightJamming( bSig, expDiff, &bSig ); 2771 zExp = aExp; 2772 } 2773 else if ( expDiff < 0 ) { 2774 if ( bExp == 0x7FF ) { 2775 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2776 return packFloat64( zSign, 0x7FF, 0 ); 2777 } 2778 if ( aExp == 0 ) { 2779 ++expDiff; 2780 } 2781 else { 2782 aSig |= LIT64( 0x2000000000000000 ); 2783 } 2784 shift64RightJamming( aSig, - expDiff, &aSig ); 2785 zExp = bExp; 2786 } 2787 else { 2788 if ( aExp == 0x7FF ) { 2789 if ( aSig | bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2790 return a; 2791 } 2792 if ( aExp == 0 ) { 2793 if ( STATUS(flush_to_zero) ) return packFloat64( zSign, 0, 0 ); 2794 return packFloat64( zSign, 0, ( aSig + bSig )>>9 ); 2795 } 2796 zSig = LIT64( 0x4000000000000000 ) + aSig + bSig; 2797 zExp = aExp; 2798 goto roundAndPack; 2799 } 2800 aSig |= LIT64( 0x2000000000000000 ); 2801 zSig = ( aSig + bSig )<<1; 2802 --zExp; 2803 if ( (sbits64) zSig < 0 ) { 2804 zSig = aSig + bSig; 2805 ++zExp; 2806 } 2807 roundAndPack: 2808 return roundAndPackFloat64( zSign, zExp, zSig STATUS_VAR ); 2809 2810 } 2811 2812 /*---------------------------------------------------------------------------- 2813 | Returns the result of subtracting the absolute values of the double- 2814 | precision floating-point values `a' and `b'. If `zSign' is 1, the 2815 | difference is negated before being returned. `zSign' is ignored if the 2816 | result is a NaN. The subtraction is performed according to the IEC/IEEE 2817 | Standard for Binary Floating-Point Arithmetic. 2818 *----------------------------------------------------------------------------*/ 2819 2820 static float64 subFloat64Sigs( float64 a, float64 b, flag zSign STATUS_PARAM ) 2821 { 2822 int16 aExp, bExp, zExp; 2823 bits64 aSig, bSig, zSig; 2824 int16 expDiff; 2825 2826 aSig = extractFloat64Frac( a ); 2827 aExp = extractFloat64Exp( a ); 2828 bSig = extractFloat64Frac( b ); 2829 bExp = extractFloat64Exp( b ); 2830 expDiff = aExp - bExp; 2831 aSig <<= 10; 2832 bSig <<= 10; 2833 if ( 0 < expDiff ) goto aExpBigger; 2834 if ( expDiff < 0 ) goto bExpBigger; 2835 if ( aExp == 0x7FF ) { 2836 if ( aSig | bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2837 float_raise( float_flag_invalid STATUS_VAR); 2838 return float64_default_nan; 2839 } 2840 if ( aExp == 0 ) { 2841 aExp = 1; 2842 bExp = 1; 2843 } 2844 if ( bSig < aSig ) goto aBigger; 2845 if ( aSig < bSig ) goto bBigger; 2846 return packFloat64( STATUS(float_rounding_mode) == float_round_down, 0, 0 ); 2847 bExpBigger: 2848 if ( bExp == 0x7FF ) { 2849 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2850 return packFloat64( zSign ^ 1, 0x7FF, 0 ); 2851 } 2852 if ( aExp == 0 ) { 2853 ++expDiff; 2854 } 2855 else { 2856 aSig |= LIT64( 0x4000000000000000 ); 2857 } 2858 shift64RightJamming( aSig, - expDiff, &aSig ); 2859 bSig |= LIT64( 0x4000000000000000 ); 2860 bBigger: 2861 zSig = bSig - aSig; 2862 zExp = bExp; 2863 zSign ^= 1; 2864 goto normalizeRoundAndPack; 2865 aExpBigger: 2866 if ( aExp == 0x7FF ) { 2867 if ( aSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2868 return a; 2869 } 2870 if ( bExp == 0 ) { 2871 --expDiff; 2872 } 2873 else { 2874 bSig |= LIT64( 0x4000000000000000 ); 2875 } 2876 shift64RightJamming( bSig, expDiff, &bSig ); 2877 aSig |= LIT64( 0x4000000000000000 ); 2878 aBigger: 2879 zSig = aSig - bSig; 2880 zExp = aExp; 2881 normalizeRoundAndPack: 2882 --zExp; 2883 return normalizeRoundAndPackFloat64( zSign, zExp, zSig STATUS_VAR ); 2884 2885 } 2886 2887 /*---------------------------------------------------------------------------- 2888 | Returns the result of adding the double-precision floating-point values `a' 2889 | and `b'. The operation is performed according to the IEC/IEEE Standard for 2890 | Binary Floating-Point Arithmetic. 2891 *----------------------------------------------------------------------------*/ 2892 2893 float64 float64_add( float64 a, float64 b STATUS_PARAM ) 2894 { 2895 flag aSign, bSign; 2896 2897 aSign = extractFloat64Sign( a ); 2898 bSign = extractFloat64Sign( b ); 2899 if ( aSign == bSign ) { 2900 return addFloat64Sigs( a, b, aSign STATUS_VAR ); 2901 } 2902 else { 2903 return subFloat64Sigs( a, b, aSign STATUS_VAR ); 2904 } 2905 2906 } 2907 2908 /*---------------------------------------------------------------------------- 2909 | Returns the result of subtracting the double-precision floating-point values 2910 | `a' and `b'. The operation is performed according to the IEC/IEEE Standard 2911 | for Binary Floating-Point Arithmetic. 2912 *----------------------------------------------------------------------------*/ 2913 2914 float64 float64_sub( float64 a, float64 b STATUS_PARAM ) 2915 { 2916 flag aSign, bSign; 2917 2918 aSign = extractFloat64Sign( a ); 2919 bSign = extractFloat64Sign( b ); 2920 if ( aSign == bSign ) { 2921 return subFloat64Sigs( a, b, aSign STATUS_VAR ); 2922 } 2923 else { 2924 return addFloat64Sigs( a, b, aSign STATUS_VAR ); 2925 } 2926 2927 } 2928 2929 /*---------------------------------------------------------------------------- 2930 | Returns the result of multiplying the double-precision floating-point values 2931 | `a' and `b'. The operation is performed according to the IEC/IEEE Standard 2932 | for Binary Floating-Point Arithmetic. 2933 *----------------------------------------------------------------------------*/ 2934 2935 float64 float64_mul( float64 a, float64 b STATUS_PARAM ) 2936 { 2937 flag aSign, bSign, zSign; 2938 int16 aExp, bExp, zExp; 2939 bits64 aSig, bSig, zSig0, zSig1; 2940 2941 aSig = extractFloat64Frac( a ); 2942 aExp = extractFloat64Exp( a ); 2943 aSign = extractFloat64Sign( a ); 2944 bSig = extractFloat64Frac( b ); 2945 bExp = extractFloat64Exp( b ); 2946 bSign = extractFloat64Sign( b ); 2947 zSign = aSign ^ bSign; 2948 if ( aExp == 0x7FF ) { 2949 if ( aSig || ( ( bExp == 0x7FF ) && bSig ) ) { 2950 return propagateFloat64NaN( a, b STATUS_VAR ); 2951 } 2952 if ( ( bExp | bSig ) == 0 ) { 2953 float_raise( float_flag_invalid STATUS_VAR); 2954 return float64_default_nan; 2955 } 2956 return packFloat64( zSign, 0x7FF, 0 ); 2957 } 2958 if ( bExp == 0x7FF ) { 2959 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 2960 if ( ( aExp | aSig ) == 0 ) { 2961 float_raise( float_flag_invalid STATUS_VAR); 2962 return float64_default_nan; 2963 } 2964 return packFloat64( zSign, 0x7FF, 0 ); 2965 } 2966 if ( aExp == 0 ) { 2967 if ( aSig == 0 ) return packFloat64( zSign, 0, 0 ); 2968 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 2969 } 2970 if ( bExp == 0 ) { 2971 if ( bSig == 0 ) return packFloat64( zSign, 0, 0 ); 2972 normalizeFloat64Subnormal( bSig, &bExp, &bSig ); 2973 } 2974 zExp = aExp + bExp - 0x3FF; 2975 aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<10; 2976 bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11; 2977 mul64To128( aSig, bSig, &zSig0, &zSig1 ); 2978 zSig0 |= ( zSig1 != 0 ); 2979 if ( 0 <= (sbits64) ( zSig0<<1 ) ) { 2980 zSig0 <<= 1; 2981 --zExp; 2982 } 2983 return roundAndPackFloat64( zSign, zExp, zSig0 STATUS_VAR ); 2984 2985 } 2986 2987 /*---------------------------------------------------------------------------- 2988 | Returns the result of dividing the double-precision floating-point value `a' 2989 | by the corresponding value `b'. The operation is performed according to 2990 | the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 2991 *----------------------------------------------------------------------------*/ 2992 2993 float64 float64_div( float64 a, float64 b STATUS_PARAM ) 2994 { 2995 flag aSign, bSign, zSign; 2996 int16 aExp, bExp, zExp; 2997 bits64 aSig, bSig, zSig; 2998 bits64 rem0, rem1; 2999 bits64 term0, term1; 3000 3001 aSig = extractFloat64Frac( a ); 3002 aExp = extractFloat64Exp( a ); 3003 aSign = extractFloat64Sign( a ); 3004 bSig = extractFloat64Frac( b ); 3005 bExp = extractFloat64Exp( b ); 3006 bSign = extractFloat64Sign( b ); 3007 zSign = aSign ^ bSign; 3008 if ( aExp == 0x7FF ) { 3009 if ( aSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 3010 if ( bExp == 0x7FF ) { 3011 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 3012 float_raise( float_flag_invalid STATUS_VAR); 3013 return float64_default_nan; 3014 } 3015 return packFloat64( zSign, 0x7FF, 0 ); 3016 } 3017 if ( bExp == 0x7FF ) { 3018 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 3019 return packFloat64( zSign, 0, 0 ); 3020 } 3021 if ( bExp == 0 ) { 3022 if ( bSig == 0 ) { 3023 if ( ( aExp | aSig ) == 0 ) { 3024 float_raise( float_flag_invalid STATUS_VAR); 3025 return float64_default_nan; 3026 } 3027 float_raise( float_flag_divbyzero STATUS_VAR); 3028 return packFloat64( zSign, 0x7FF, 0 ); 3029 } 3030 normalizeFloat64Subnormal( bSig, &bExp, &bSig ); 3031 } 3032 if ( aExp == 0 ) { 3033 if ( aSig == 0 ) return packFloat64( zSign, 0, 0 ); 3034 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 3035 } 3036 zExp = aExp - bExp + 0x3FD; 3037 aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<10; 3038 bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11; 3039 if ( bSig <= ( aSig + aSig ) ) { 3040 aSig >>= 1; 3041 ++zExp; 3042 } 3043 zSig = estimateDiv128To64( aSig, 0, bSig ); 3044 if ( ( zSig & 0x1FF ) <= 2 ) { 3045 mul64To128( bSig, zSig, &term0, &term1 ); 3046 sub128( aSig, 0, term0, term1, &rem0, &rem1 ); 3047 while ( (sbits64) rem0 < 0 ) { 3048 --zSig; 3049 add128( rem0, rem1, 0, bSig, &rem0, &rem1 ); 3050 } 3051 zSig |= ( rem1 != 0 ); 3052 } 3053 return roundAndPackFloat64( zSign, zExp, zSig STATUS_VAR ); 3054 3055 } 3056 3057 /*---------------------------------------------------------------------------- 3058 | Returns the remainder of the double-precision floating-point value `a' 3059 | with respect to the corresponding value `b'. The operation is performed 3060 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3061 *----------------------------------------------------------------------------*/ 3062 3063 float64 float64_rem( float64 a, float64 b STATUS_PARAM ) 3064 { 3065 flag aSign, bSign, zSign; 3066 int16 aExp, bExp, expDiff; 3067 bits64 aSig, bSig; 3068 bits64 q, alternateASig; 3069 sbits64 sigMean; 3070 3071 aSig = extractFloat64Frac( a ); 3072 aExp = extractFloat64Exp( a ); 3073 aSign = extractFloat64Sign( a ); 3074 bSig = extractFloat64Frac( b ); 3075 bExp = extractFloat64Exp( b ); 3076 bSign = extractFloat64Sign( b ); 3077 if ( aExp == 0x7FF ) { 3078 if ( aSig || ( ( bExp == 0x7FF ) && bSig ) ) { 3079 return propagateFloat64NaN( a, b STATUS_VAR ); 3080 } 3081 float_raise( float_flag_invalid STATUS_VAR); 3082 return float64_default_nan; 3083 } 3084 if ( bExp == 0x7FF ) { 3085 if ( bSig ) return propagateFloat64NaN( a, b STATUS_VAR ); 3086 return a; 3087 } 3088 if ( bExp == 0 ) { 3089 if ( bSig == 0 ) { 3090 float_raise( float_flag_invalid STATUS_VAR); 3091 return float64_default_nan; 3092 } 3093 normalizeFloat64Subnormal( bSig, &bExp, &bSig ); 3094 } 3095 if ( aExp == 0 ) { 3096 if ( aSig == 0 ) return a; 3097 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 3098 } 3099 expDiff = aExp - bExp; 3100 aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<11; 3101 bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11; 3102 if ( expDiff < 0 ) { 3103 if ( expDiff < -1 ) return a; 3104 aSig >>= 1; 3105 } 3106 q = ( bSig <= aSig ); 3107 if ( q ) aSig -= bSig; 3108 expDiff -= 64; 3109 while ( 0 < expDiff ) { 3110 q = estimateDiv128To64( aSig, 0, bSig ); 3111 q = ( 2 < q ) ? q - 2 : 0; 3112 aSig = - ( ( bSig>>2 ) * q ); 3113 expDiff -= 62; 3114 } 3115 expDiff += 64; 3116 if ( 0 < expDiff ) { 3117 q = estimateDiv128To64( aSig, 0, bSig ); 3118 q = ( 2 < q ) ? q - 2 : 0; 3119 q >>= 64 - expDiff; 3120 bSig >>= 2; 3121 aSig = ( ( aSig>>1 )<<( expDiff - 1 ) ) - bSig * q; 3122 } 3123 else { 3124 aSig >>= 2; 3125 bSig >>= 2; 3126 } 3127 do { 3128 alternateASig = aSig; 3129 ++q; 3130 aSig -= bSig; 3131 } while ( 0 <= (sbits64) aSig ); 3132 sigMean = aSig + alternateASig; 3133 if ( ( sigMean < 0 ) || ( ( sigMean == 0 ) && ( q & 1 ) ) ) { 3134 aSig = alternateASig; 3135 } 3136 zSign = ( (sbits64) aSig < 0 ); 3137 if ( zSign ) aSig = - aSig; 3138 return normalizeRoundAndPackFloat64( aSign ^ zSign, bExp, aSig STATUS_VAR ); 3139 3140 } 3141 3142 /*---------------------------------------------------------------------------- 3143 | Returns the square root of the double-precision floating-point value `a'. 3144 | The operation is performed according to the IEC/IEEE Standard for Binary 3145 | Floating-Point Arithmetic. 3146 *----------------------------------------------------------------------------*/ 3147 3148 float64 float64_sqrt( float64 a STATUS_PARAM ) 3149 { 3150 flag aSign; 3151 int16 aExp, zExp; 3152 bits64 aSig, zSig, doubleZSig; 3153 bits64 rem0, rem1, term0, term1; 3154 3155 aSig = extractFloat64Frac( a ); 3156 aExp = extractFloat64Exp( a ); 3157 aSign = extractFloat64Sign( a ); 3158 if ( aExp == 0x7FF ) { 3159 if ( aSig ) return propagateFloat64NaN( a, a STATUS_VAR ); 3160 if ( ! aSign ) return a; 3161 float_raise( float_flag_invalid STATUS_VAR); 3162 return float64_default_nan; 3163 } 3164 if ( aSign ) { 3165 if ( ( aExp | aSig ) == 0 ) return a; 3166 float_raise( float_flag_invalid STATUS_VAR); 3167 return float64_default_nan; 3168 } 3169 if ( aExp == 0 ) { 3170 if ( aSig == 0 ) return float64_zero; 3171 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 3172 } 3173 zExp = ( ( aExp - 0x3FF )>>1 ) + 0x3FE; 3174 aSig |= LIT64( 0x0010000000000000 ); 3175 zSig = estimateSqrt32( aExp, aSig>>21 ); 3176 aSig <<= 9 - ( aExp & 1 ); 3177 zSig = estimateDiv128To64( aSig, 0, zSig<<32 ) + ( zSig<<30 ); 3178 if ( ( zSig & 0x1FF ) <= 5 ) { 3179 doubleZSig = zSig<<1; 3180 mul64To128( zSig, zSig, &term0, &term1 ); 3181 sub128( aSig, 0, term0, term1, &rem0, &rem1 ); 3182 while ( (sbits64) rem0 < 0 ) { 3183 --zSig; 3184 doubleZSig -= 2; 3185 add128( rem0, rem1, zSig>>63, doubleZSig | 1, &rem0, &rem1 ); 3186 } 3187 zSig |= ( ( rem0 | rem1 ) != 0 ); 3188 } 3189 return roundAndPackFloat64( 0, zExp, zSig STATUS_VAR ); 3190 3191 } 3192 3193 /*---------------------------------------------------------------------------- 3194 | Returns the binary log of the double-precision floating-point value `a'. 3195 | The operation is performed according to the IEC/IEEE Standard for Binary 3196 | Floating-Point Arithmetic. 3197 *----------------------------------------------------------------------------*/ 3198 float64 float64_log2( float64 a STATUS_PARAM ) 3199 { 3200 flag aSign, zSign; 3201 int16 aExp; 3202 bits64 aSig, aSig0, aSig1, zSig, i; 3203 3204 aSig = extractFloat64Frac( a ); 3205 aExp = extractFloat64Exp( a ); 3206 aSign = extractFloat64Sign( a ); 3207 3208 if ( aExp == 0 ) { 3209 if ( aSig == 0 ) return packFloat64( 1, 0x7FF, 0 ); 3210 normalizeFloat64Subnormal( aSig, &aExp, &aSig ); 3211 } 3212 if ( aSign ) { 3213 float_raise( float_flag_invalid STATUS_VAR); 3214 return float64_default_nan; 3215 } 3216 if ( aExp == 0x7FF ) { 3217 if ( aSig ) return propagateFloat64NaN( a, float64_zero STATUS_VAR ); 3218 return a; 3219 } 3220 3221 aExp -= 0x3FF; 3222 aSig |= LIT64( 0x0010000000000000 ); 3223 zSign = aExp < 0; 3224 zSig = (bits64)aExp << 52; 3225 for (i = 1LL << 51; i > 0; i >>= 1) { 3226 mul64To128( aSig, aSig, &aSig0, &aSig1 ); 3227 aSig = ( aSig0 << 12 ) | ( aSig1 >> 52 ); 3228 if ( aSig & LIT64( 0x0020000000000000 ) ) { 3229 aSig >>= 1; 3230 zSig |= i; 3231 } 3232 } 3233 3234 if ( zSign ) 3235 zSig = -zSig; 3236 return normalizeRoundAndPackFloat64( zSign, 0x408, zSig STATUS_VAR ); 3237 } 3238 3239 /*---------------------------------------------------------------------------- 3240 | Returns 1 if the double-precision floating-point value `a' is equal to the 3241 | corresponding value `b', and 0 otherwise. The comparison is performed 3242 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3243 *----------------------------------------------------------------------------*/ 3244 3245 int float64_eq( float64 a, float64 b STATUS_PARAM ) 3246 { 3247 bits64 av, bv; 3248 3249 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3250 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3251 ) { 3252 if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) { 3253 float_raise( float_flag_invalid STATUS_VAR); 3254 } 3255 return 0; 3256 } 3257 av = float64_val(a); 3258 bv = float64_val(b); 3259 return ( av == bv ) || ( (bits64) ( ( av | bv )<<1 ) == 0 ); 3260 3261 } 3262 3263 /*---------------------------------------------------------------------------- 3264 | Returns 1 if the double-precision floating-point value `a' is less than or 3265 | equal to the corresponding value `b', and 0 otherwise. The comparison is 3266 | performed according to the IEC/IEEE Standard for Binary Floating-Point 3267 | Arithmetic. 3268 *----------------------------------------------------------------------------*/ 3269 3270 int float64_le( float64 a, float64 b STATUS_PARAM ) 3271 { 3272 flag aSign, bSign; 3273 bits64 av, bv; 3274 3275 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3276 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3277 ) { 3278 float_raise( float_flag_invalid STATUS_VAR); 3279 return 0; 3280 } 3281 aSign = extractFloat64Sign( a ); 3282 bSign = extractFloat64Sign( b ); 3283 av = float64_val(a); 3284 bv = float64_val(b); 3285 if ( aSign != bSign ) return aSign || ( (bits64) ( ( av | bv )<<1 ) == 0 ); 3286 return ( av == bv ) || ( aSign ^ ( av < bv ) ); 3287 3288 } 3289 3290 /*---------------------------------------------------------------------------- 3291 | Returns 1 if the double-precision floating-point value `a' is less than 3292 | the corresponding value `b', and 0 otherwise. The comparison is performed 3293 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3294 *----------------------------------------------------------------------------*/ 3295 3296 int float64_lt( float64 a, float64 b STATUS_PARAM ) 3297 { 3298 flag aSign, bSign; 3299 bits64 av, bv; 3300 3301 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3302 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3303 ) { 3304 float_raise( float_flag_invalid STATUS_VAR); 3305 return 0; 3306 } 3307 aSign = extractFloat64Sign( a ); 3308 bSign = extractFloat64Sign( b ); 3309 av = float64_val(a); 3310 bv = float64_val(b); 3311 if ( aSign != bSign ) return aSign && ( (bits64) ( ( av | bv )<<1 ) != 0 ); 3312 return ( av != bv ) && ( aSign ^ ( av < bv ) ); 3313 3314 } 3315 3316 /*---------------------------------------------------------------------------- 3317 | Returns 1 if the double-precision floating-point value `a' is equal to the 3318 | corresponding value `b', and 0 otherwise. The invalid exception is raised 3319 | if either operand is a NaN. Otherwise, the comparison is performed 3320 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3321 *----------------------------------------------------------------------------*/ 3322 3323 int float64_eq_signaling( float64 a, float64 b STATUS_PARAM ) 3324 { 3325 bits64 av, bv; 3326 3327 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3328 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3329 ) { 3330 float_raise( float_flag_invalid STATUS_VAR); 3331 return 0; 3332 } 3333 av = float64_val(a); 3334 bv = float64_val(b); 3335 return ( av == bv ) || ( (bits64) ( ( av | bv )<<1 ) == 0 ); 3336 3337 } 3338 3339 /*---------------------------------------------------------------------------- 3340 | Returns 1 if the double-precision floating-point value `a' is less than or 3341 | equal to the corresponding value `b', and 0 otherwise. Quiet NaNs do not 3342 | cause an exception. Otherwise, the comparison is performed according to the 3343 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3344 *----------------------------------------------------------------------------*/ 3345 3346 int float64_le_quiet( float64 a, float64 b STATUS_PARAM ) 3347 { 3348 flag aSign, bSign; 3349 bits64 av, bv; 3350 3351 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3352 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3353 ) { 3354 if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) { 3355 float_raise( float_flag_invalid STATUS_VAR); 3356 } 3357 return 0; 3358 } 3359 aSign = extractFloat64Sign( a ); 3360 bSign = extractFloat64Sign( b ); 3361 av = float64_val(a); 3362 bv = float64_val(b); 3363 if ( aSign != bSign ) return aSign || ( (bits64) ( ( av | bv )<<1 ) == 0 ); 3364 return ( av == bv ) || ( aSign ^ ( av < bv ) ); 3365 3366 } 3367 3368 /*---------------------------------------------------------------------------- 3369 | Returns 1 if the double-precision floating-point value `a' is less than 3370 | the corresponding value `b', and 0 otherwise. Quiet NaNs do not cause an 3371 | exception. Otherwise, the comparison is performed according to the IEC/IEEE 3372 | Standard for Binary Floating-Point Arithmetic. 3373 *----------------------------------------------------------------------------*/ 3374 3375 int float64_lt_quiet( float64 a, float64 b STATUS_PARAM ) 3376 { 3377 flag aSign, bSign; 3378 bits64 av, bv; 3379 3380 if ( ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) ) 3381 || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) ) 3382 ) { 3383 if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) { 3384 float_raise( float_flag_invalid STATUS_VAR); 3385 } 3386 return 0; 3387 } 3388 aSign = extractFloat64Sign( a ); 3389 bSign = extractFloat64Sign( b ); 3390 av = float64_val(a); 3391 bv = float64_val(b); 3392 if ( aSign != bSign ) return aSign && ( (bits64) ( ( av | bv )<<1 ) != 0 ); 3393 return ( av != bv ) && ( aSign ^ ( av < bv ) ); 3394 3395 } 3396 3397 #ifdef FLOATX80 3398 3399 /*---------------------------------------------------------------------------- 3400 | Returns the result of converting the extended double-precision floating- 3401 | point value `a' to the 32-bit two's complement integer format. The 3402 | conversion is performed according to the IEC/IEEE Standard for Binary 3403 | Floating-Point Arithmetic---which means in particular that the conversion 3404 | is rounded according to the current rounding mode. If `a' is a NaN, the 3405 | largest positive integer is returned. Otherwise, if the conversion 3406 | overflows, the largest integer with the same sign as `a' is returned. 3407 *----------------------------------------------------------------------------*/ 3408 3409 int32 floatx80_to_int32( floatx80 a STATUS_PARAM ) 3410 { 3411 flag aSign; 3412 int32 aExp, shiftCount; 3413 bits64 aSig; 3414 3415 aSig = extractFloatx80Frac( a ); 3416 aExp = extractFloatx80Exp( a ); 3417 aSign = extractFloatx80Sign( a ); 3418 if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) aSign = 0; 3419 shiftCount = 0x4037 - aExp; 3420 if ( shiftCount <= 0 ) shiftCount = 1; 3421 shift64RightJamming( aSig, shiftCount, &aSig ); 3422 return roundAndPackInt32( aSign, aSig STATUS_VAR ); 3423 3424 } 3425 3426 /*---------------------------------------------------------------------------- 3427 | Returns the result of converting the extended double-precision floating- 3428 | point value `a' to the 32-bit two's complement integer format. The 3429 | conversion is performed according to the IEC/IEEE Standard for Binary 3430 | Floating-Point Arithmetic, except that the conversion is always rounded 3431 | toward zero. If `a' is a NaN, the largest positive integer is returned. 3432 | Otherwise, if the conversion overflows, the largest integer with the same 3433 | sign as `a' is returned. 3434 *----------------------------------------------------------------------------*/ 3435 3436 int32 floatx80_to_int32_round_to_zero( floatx80 a STATUS_PARAM ) 3437 { 3438 flag aSign; 3439 int32 aExp, shiftCount; 3440 bits64 aSig, savedASig; 3441 int32 z; 3442 3443 aSig = extractFloatx80Frac( a ); 3444 aExp = extractFloatx80Exp( a ); 3445 aSign = extractFloatx80Sign( a ); 3446 if ( 0x401E < aExp ) { 3447 if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) aSign = 0; 3448 goto invalid; 3449 } 3450 else if ( aExp < 0x3FFF ) { 3451 if ( aExp || aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 3452 return 0; 3453 } 3454 shiftCount = 0x403E - aExp; 3455 savedASig = aSig; 3456 aSig >>= shiftCount; 3457 z = aSig; 3458 if ( aSign ) z = - z; 3459 if ( ( z < 0 ) ^ aSign ) { 3460 invalid: 3461 float_raise( float_flag_invalid STATUS_VAR); 3462 return aSign ? (sbits32) 0x80000000 : 0x7FFFFFFF; 3463 } 3464 if ( ( aSig<<shiftCount ) != savedASig ) { 3465 STATUS(float_exception_flags) |= float_flag_inexact; 3466 } 3467 return z; 3468 3469 } 3470 3471 /*---------------------------------------------------------------------------- 3472 | Returns the result of converting the extended double-precision floating- 3473 | point value `a' to the 64-bit two's complement integer format. The 3474 | conversion is performed according to the IEC/IEEE Standard for Binary 3475 | Floating-Point Arithmetic---which means in particular that the conversion 3476 | is rounded according to the current rounding mode. If `a' is a NaN, 3477 | the largest positive integer is returned. Otherwise, if the conversion 3478 | overflows, the largest integer with the same sign as `a' is returned. 3479 *----------------------------------------------------------------------------*/ 3480 3481 int64 floatx80_to_int64( floatx80 a STATUS_PARAM ) 3482 { 3483 flag aSign; 3484 int32 aExp, shiftCount; 3485 bits64 aSig, aSigExtra; 3486 3487 aSig = extractFloatx80Frac( a ); 3488 aExp = extractFloatx80Exp( a ); 3489 aSign = extractFloatx80Sign( a ); 3490 shiftCount = 0x403E - aExp; 3491 if ( shiftCount <= 0 ) { 3492 if ( shiftCount ) { 3493 float_raise( float_flag_invalid STATUS_VAR); 3494 if ( ! aSign 3495 || ( ( aExp == 0x7FFF ) 3496 && ( aSig != LIT64( 0x8000000000000000 ) ) ) 3497 ) { 3498 return LIT64( 0x7FFFFFFFFFFFFFFF ); 3499 } 3500 return (sbits64) LIT64( 0x8000000000000000 ); 3501 } 3502 aSigExtra = 0; 3503 } 3504 else { 3505 shift64ExtraRightJamming( aSig, 0, shiftCount, &aSig, &aSigExtra ); 3506 } 3507 return roundAndPackInt64( aSign, aSig, aSigExtra STATUS_VAR ); 3508 3509 } 3510 3511 /*---------------------------------------------------------------------------- 3512 | Returns the result of converting the extended double-precision floating- 3513 | point value `a' to the 64-bit two's complement integer format. The 3514 | conversion is performed according to the IEC/IEEE Standard for Binary 3515 | Floating-Point Arithmetic, except that the conversion is always rounded 3516 | toward zero. If `a' is a NaN, the largest positive integer is returned. 3517 | Otherwise, if the conversion overflows, the largest integer with the same 3518 | sign as `a' is returned. 3519 *----------------------------------------------------------------------------*/ 3520 3521 int64 floatx80_to_int64_round_to_zero( floatx80 a STATUS_PARAM ) 3522 { 3523 flag aSign; 3524 int32 aExp, shiftCount; 3525 bits64 aSig; 3526 int64 z; 3527 3528 aSig = extractFloatx80Frac( a ); 3529 aExp = extractFloatx80Exp( a ); 3530 aSign = extractFloatx80Sign( a ); 3531 shiftCount = aExp - 0x403E; 3532 if ( 0 <= shiftCount ) { 3533 aSig &= LIT64( 0x7FFFFFFFFFFFFFFF ); 3534 if ( ( a.high != 0xC03E ) || aSig ) { 3535 float_raise( float_flag_invalid STATUS_VAR); 3536 if ( ! aSign || ( ( aExp == 0x7FFF ) && aSig ) ) { 3537 return LIT64( 0x7FFFFFFFFFFFFFFF ); 3538 } 3539 } 3540 return (sbits64) LIT64( 0x8000000000000000 ); 3541 } 3542 else if ( aExp < 0x3FFF ) { 3543 if ( aExp | aSig ) STATUS(float_exception_flags) |= float_flag_inexact; 3544 return 0; 3545 } 3546 z = aSig>>( - shiftCount ); 3547 if ( (bits64) ( aSig<<( shiftCount & 63 ) ) ) { 3548 STATUS(float_exception_flags) |= float_flag_inexact; 3549 } 3550 if ( aSign ) z = - z; 3551 return z; 3552 3553 } 3554 3555 /*---------------------------------------------------------------------------- 3556 | Returns the result of converting the extended double-precision floating- 3557 | point value `a' to the single-precision floating-point format. The 3558 | conversion is performed according to the IEC/IEEE Standard for Binary 3559 | Floating-Point Arithmetic. 3560 *----------------------------------------------------------------------------*/ 3561 3562 float32 floatx80_to_float32( floatx80 a STATUS_PARAM ) 3563 { 3564 flag aSign; 3565 int32 aExp; 3566 bits64 aSig; 3567 3568 aSig = extractFloatx80Frac( a ); 3569 aExp = extractFloatx80Exp( a ); 3570 aSign = extractFloatx80Sign( a ); 3571 if ( aExp == 0x7FFF ) { 3572 if ( (bits64) ( aSig<<1 ) ) { 3573 return commonNaNToFloat32( floatx80ToCommonNaN( a STATUS_VAR ) ); 3574 } 3575 return packFloat32( aSign, 0xFF, 0 ); 3576 } 3577 shift64RightJamming( aSig, 33, &aSig ); 3578 if ( aExp || aSig ) aExp -= 0x3F81; 3579 return roundAndPackFloat32( aSign, aExp, aSig STATUS_VAR ); 3580 3581 } 3582 3583 /*---------------------------------------------------------------------------- 3584 | Returns the result of converting the extended double-precision floating- 3585 | point value `a' to the double-precision floating-point format. The 3586 | conversion is performed according to the IEC/IEEE Standard for Binary 3587 | Floating-Point Arithmetic. 3588 *----------------------------------------------------------------------------*/ 3589 3590 float64 floatx80_to_float64( floatx80 a STATUS_PARAM ) 3591 { 3592 flag aSign; 3593 int32 aExp; 3594 bits64 aSig, zSig; 3595 3596 aSig = extractFloatx80Frac( a ); 3597 aExp = extractFloatx80Exp( a ); 3598 aSign = extractFloatx80Sign( a ); 3599 if ( aExp == 0x7FFF ) { 3600 if ( (bits64) ( aSig<<1 ) ) { 3601 return commonNaNToFloat64( floatx80ToCommonNaN( a STATUS_VAR ) ); 3602 } 3603 return packFloat64( aSign, 0x7FF, 0 ); 3604 } 3605 shift64RightJamming( aSig, 1, &zSig ); 3606 if ( aExp || aSig ) aExp -= 0x3C01; 3607 return roundAndPackFloat64( aSign, aExp, zSig STATUS_VAR ); 3608 3609 } 3610 3611 #ifdef FLOAT128 3612 3613 /*---------------------------------------------------------------------------- 3614 | Returns the result of converting the extended double-precision floating- 3615 | point value `a' to the quadruple-precision floating-point format. The 3616 | conversion is performed according to the IEC/IEEE Standard for Binary 3617 | Floating-Point Arithmetic. 3618 *----------------------------------------------------------------------------*/ 3619 3620 float128 floatx80_to_float128( floatx80 a STATUS_PARAM ) 3621 { 3622 flag aSign; 3623 int16 aExp; 3624 bits64 aSig, zSig0, zSig1; 3625 3626 aSig = extractFloatx80Frac( a ); 3627 aExp = extractFloatx80Exp( a ); 3628 aSign = extractFloatx80Sign( a ); 3629 if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) { 3630 return commonNaNToFloat128( floatx80ToCommonNaN( a STATUS_VAR ) ); 3631 } 3632 shift128Right( aSig<<1, 0, 16, &zSig0, &zSig1 ); 3633 return packFloat128( aSign, aExp, zSig0, zSig1 ); 3634 3635 } 3636 3637 #endif 3638 3639 /*---------------------------------------------------------------------------- 3640 | Rounds the extended double-precision floating-point value `a' to an integer, 3641 | and returns the result as an extended quadruple-precision floating-point 3642 | value. The operation is performed according to the IEC/IEEE Standard for 3643 | Binary Floating-Point Arithmetic. 3644 *----------------------------------------------------------------------------*/ 3645 3646 floatx80 floatx80_round_to_int( floatx80 a STATUS_PARAM ) 3647 { 3648 flag aSign; 3649 int32 aExp; 3650 bits64 lastBitMask, roundBitsMask; 3651 int8 roundingMode; 3652 floatx80 z; 3653 3654 aExp = extractFloatx80Exp( a ); 3655 if ( 0x403E <= aExp ) { 3656 if ( ( aExp == 0x7FFF ) && (bits64) ( extractFloatx80Frac( a )<<1 ) ) { 3657 return propagateFloatx80NaN( a, a STATUS_VAR ); 3658 } 3659 return a; 3660 } 3661 if ( aExp < 0x3FFF ) { 3662 if ( ( aExp == 0 ) 3663 && ( (bits64) ( extractFloatx80Frac( a )<<1 ) == 0 ) ) { 3664 return a; 3665 } 3666 STATUS(float_exception_flags) |= float_flag_inexact; 3667 aSign = extractFloatx80Sign( a ); 3668 switch ( STATUS(float_rounding_mode) ) { 3669 case float_round_nearest_even: 3670 if ( ( aExp == 0x3FFE ) && (bits64) ( extractFloatx80Frac( a )<<1 ) 3671 ) { 3672 return 3673 packFloatx80( aSign, 0x3FFF, LIT64( 0x8000000000000000 ) ); 3674 } 3675 break; 3676 case float_round_down: 3677 return 3678 aSign ? 3679 packFloatx80( 1, 0x3FFF, LIT64( 0x8000000000000000 ) ) 3680 : packFloatx80( 0, 0, 0 ); 3681 case float_round_up: 3682 return 3683 aSign ? packFloatx80( 1, 0, 0 ) 3684 : packFloatx80( 0, 0x3FFF, LIT64( 0x8000000000000000 ) ); 3685 } 3686 return packFloatx80( aSign, 0, 0 ); 3687 } 3688 lastBitMask = 1; 3689 lastBitMask <<= 0x403E - aExp; 3690 roundBitsMask = lastBitMask - 1; 3691 z = a; 3692 roundingMode = STATUS(float_rounding_mode); 3693 if ( roundingMode == float_round_nearest_even ) { 3694 z.low += lastBitMask>>1; 3695 if ( ( z.low & roundBitsMask ) == 0 ) z.low &= ~ lastBitMask; 3696 } 3697 else if ( roundingMode != float_round_to_zero ) { 3698 if ( extractFloatx80Sign( z ) ^ ( roundingMode == float_round_up ) ) { 3699 z.low += roundBitsMask; 3700 } 3701 } 3702 z.low &= ~ roundBitsMask; 3703 if ( z.low == 0 ) { 3704 ++z.high; 3705 z.low = LIT64( 0x8000000000000000 ); 3706 } 3707 if ( z.low != a.low ) STATUS(float_exception_flags) |= float_flag_inexact; 3708 return z; 3709 3710 } 3711 3712 /*---------------------------------------------------------------------------- 3713 | Returns the result of adding the absolute values of the extended double- 3714 | precision floating-point values `a' and `b'. If `zSign' is 1, the sum is 3715 | negated before being returned. `zSign' is ignored if the result is a NaN. 3716 | The addition is performed according to the IEC/IEEE Standard for Binary 3717 | Floating-Point Arithmetic. 3718 *----------------------------------------------------------------------------*/ 3719 3720 static floatx80 addFloatx80Sigs( floatx80 a, floatx80 b, flag zSign STATUS_PARAM) 3721 { 3722 int32 aExp, bExp, zExp; 3723 bits64 aSig, bSig, zSig0, zSig1; 3724 int32 expDiff; 3725 3726 aSig = extractFloatx80Frac( a ); 3727 aExp = extractFloatx80Exp( a ); 3728 bSig = extractFloatx80Frac( b ); 3729 bExp = extractFloatx80Exp( b ); 3730 expDiff = aExp - bExp; 3731 if ( 0 < expDiff ) { 3732 if ( aExp == 0x7FFF ) { 3733 if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3734 return a; 3735 } 3736 if ( bExp == 0 ) --expDiff; 3737 shift64ExtraRightJamming( bSig, 0, expDiff, &bSig, &zSig1 ); 3738 zExp = aExp; 3739 } 3740 else if ( expDiff < 0 ) { 3741 if ( bExp == 0x7FFF ) { 3742 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3743 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3744 } 3745 if ( aExp == 0 ) ++expDiff; 3746 shift64ExtraRightJamming( aSig, 0, - expDiff, &aSig, &zSig1 ); 3747 zExp = bExp; 3748 } 3749 else { 3750 if ( aExp == 0x7FFF ) { 3751 if ( (bits64) ( ( aSig | bSig )<<1 ) ) { 3752 return propagateFloatx80NaN( a, b STATUS_VAR ); 3753 } 3754 return a; 3755 } 3756 zSig1 = 0; 3757 zSig0 = aSig + bSig; 3758 if ( aExp == 0 ) { 3759 normalizeFloatx80Subnormal( zSig0, &zExp, &zSig0 ); 3760 goto roundAndPack; 3761 } 3762 zExp = aExp; 3763 goto shiftRight1; 3764 } 3765 zSig0 = aSig + bSig; 3766 if ( (sbits64) zSig0 < 0 ) goto roundAndPack; 3767 shiftRight1: 3768 shift64ExtraRightJamming( zSig0, zSig1, 1, &zSig0, &zSig1 ); 3769 zSig0 |= LIT64( 0x8000000000000000 ); 3770 ++zExp; 3771 roundAndPack: 3772 return 3773 roundAndPackFloatx80( 3774 STATUS(floatx80_rounding_precision), zSign, zExp, zSig0, zSig1 STATUS_VAR ); 3775 3776 } 3777 3778 /*---------------------------------------------------------------------------- 3779 | Returns the result of subtracting the absolute values of the extended 3780 | double-precision floating-point values `a' and `b'. If `zSign' is 1, the 3781 | difference is negated before being returned. `zSign' is ignored if the 3782 | result is a NaN. The subtraction is performed according to the IEC/IEEE 3783 | Standard for Binary Floating-Point Arithmetic. 3784 *----------------------------------------------------------------------------*/ 3785 3786 static floatx80 subFloatx80Sigs( floatx80 a, floatx80 b, flag zSign STATUS_PARAM ) 3787 { 3788 int32 aExp, bExp, zExp; 3789 bits64 aSig, bSig, zSig0, zSig1; 3790 int32 expDiff; 3791 floatx80 z; 3792 3793 aSig = extractFloatx80Frac( a ); 3794 aExp = extractFloatx80Exp( a ); 3795 bSig = extractFloatx80Frac( b ); 3796 bExp = extractFloatx80Exp( b ); 3797 expDiff = aExp - bExp; 3798 if ( 0 < expDiff ) goto aExpBigger; 3799 if ( expDiff < 0 ) goto bExpBigger; 3800 if ( aExp == 0x7FFF ) { 3801 if ( (bits64) ( ( aSig | bSig )<<1 ) ) { 3802 return propagateFloatx80NaN( a, b STATUS_VAR ); 3803 } 3804 float_raise( float_flag_invalid STATUS_VAR); 3805 z.low = floatx80_default_nan_low; 3806 z.high = floatx80_default_nan_high; 3807 return z; 3808 } 3809 if ( aExp == 0 ) { 3810 aExp = 1; 3811 bExp = 1; 3812 } 3813 zSig1 = 0; 3814 if ( bSig < aSig ) goto aBigger; 3815 if ( aSig < bSig ) goto bBigger; 3816 return packFloatx80( STATUS(float_rounding_mode) == float_round_down, 0, 0 ); 3817 bExpBigger: 3818 if ( bExp == 0x7FFF ) { 3819 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3820 return packFloatx80( zSign ^ 1, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3821 } 3822 if ( aExp == 0 ) ++expDiff; 3823 shift128RightJamming( aSig, 0, - expDiff, &aSig, &zSig1 ); 3824 bBigger: 3825 sub128( bSig, 0, aSig, zSig1, &zSig0, &zSig1 ); 3826 zExp = bExp; 3827 zSign ^= 1; 3828 goto normalizeRoundAndPack; 3829 aExpBigger: 3830 if ( aExp == 0x7FFF ) { 3831 if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3832 return a; 3833 } 3834 if ( bExp == 0 ) --expDiff; 3835 shift128RightJamming( bSig, 0, expDiff, &bSig, &zSig1 ); 3836 aBigger: 3837 sub128( aSig, 0, bSig, zSig1, &zSig0, &zSig1 ); 3838 zExp = aExp; 3839 normalizeRoundAndPack: 3840 return 3841 normalizeRoundAndPackFloatx80( 3842 STATUS(floatx80_rounding_precision), zSign, zExp, zSig0, zSig1 STATUS_VAR ); 3843 3844 } 3845 3846 /*---------------------------------------------------------------------------- 3847 | Returns the result of adding the extended double-precision floating-point 3848 | values `a' and `b'. The operation is performed according to the IEC/IEEE 3849 | Standard for Binary Floating-Point Arithmetic. 3850 *----------------------------------------------------------------------------*/ 3851 3852 floatx80 floatx80_add( floatx80 a, floatx80 b STATUS_PARAM ) 3853 { 3854 flag aSign, bSign; 3855 3856 aSign = extractFloatx80Sign( a ); 3857 bSign = extractFloatx80Sign( b ); 3858 if ( aSign == bSign ) { 3859 return addFloatx80Sigs( a, b, aSign STATUS_VAR ); 3860 } 3861 else { 3862 return subFloatx80Sigs( a, b, aSign STATUS_VAR ); 3863 } 3864 3865 } 3866 3867 /*---------------------------------------------------------------------------- 3868 | Returns the result of subtracting the extended double-precision floating- 3869 | point values `a' and `b'. The operation is performed according to the 3870 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3871 *----------------------------------------------------------------------------*/ 3872 3873 floatx80 floatx80_sub( floatx80 a, floatx80 b STATUS_PARAM ) 3874 { 3875 flag aSign, bSign; 3876 3877 aSign = extractFloatx80Sign( a ); 3878 bSign = extractFloatx80Sign( b ); 3879 if ( aSign == bSign ) { 3880 return subFloatx80Sigs( a, b, aSign STATUS_VAR ); 3881 } 3882 else { 3883 return addFloatx80Sigs( a, b, aSign STATUS_VAR ); 3884 } 3885 3886 } 3887 3888 /*---------------------------------------------------------------------------- 3889 | Returns the result of multiplying the extended double-precision floating- 3890 | point values `a' and `b'. The operation is performed according to the 3891 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3892 *----------------------------------------------------------------------------*/ 3893 3894 floatx80 floatx80_mul( floatx80 a, floatx80 b STATUS_PARAM ) 3895 { 3896 flag aSign, bSign, zSign; 3897 int32 aExp, bExp, zExp; 3898 bits64 aSig, bSig, zSig0, zSig1; 3899 floatx80 z; 3900 3901 aSig = extractFloatx80Frac( a ); 3902 aExp = extractFloatx80Exp( a ); 3903 aSign = extractFloatx80Sign( a ); 3904 bSig = extractFloatx80Frac( b ); 3905 bExp = extractFloatx80Exp( b ); 3906 bSign = extractFloatx80Sign( b ); 3907 zSign = aSign ^ bSign; 3908 if ( aExp == 0x7FFF ) { 3909 if ( (bits64) ( aSig<<1 ) 3910 || ( ( bExp == 0x7FFF ) && (bits64) ( bSig<<1 ) ) ) { 3911 return propagateFloatx80NaN( a, b STATUS_VAR ); 3912 } 3913 if ( ( bExp | bSig ) == 0 ) goto invalid; 3914 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3915 } 3916 if ( bExp == 0x7FFF ) { 3917 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3918 if ( ( aExp | aSig ) == 0 ) { 3919 invalid: 3920 float_raise( float_flag_invalid STATUS_VAR); 3921 z.low = floatx80_default_nan_low; 3922 z.high = floatx80_default_nan_high; 3923 return z; 3924 } 3925 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3926 } 3927 if ( aExp == 0 ) { 3928 if ( aSig == 0 ) return packFloatx80( zSign, 0, 0 ); 3929 normalizeFloatx80Subnormal( aSig, &aExp, &aSig ); 3930 } 3931 if ( bExp == 0 ) { 3932 if ( bSig == 0 ) return packFloatx80( zSign, 0, 0 ); 3933 normalizeFloatx80Subnormal( bSig, &bExp, &bSig ); 3934 } 3935 zExp = aExp + bExp - 0x3FFE; 3936 mul64To128( aSig, bSig, &zSig0, &zSig1 ); 3937 if ( 0 < (sbits64) zSig0 ) { 3938 shortShift128Left( zSig0, zSig1, 1, &zSig0, &zSig1 ); 3939 --zExp; 3940 } 3941 return 3942 roundAndPackFloatx80( 3943 STATUS(floatx80_rounding_precision), zSign, zExp, zSig0, zSig1 STATUS_VAR ); 3944 3945 } 3946 3947 /*---------------------------------------------------------------------------- 3948 | Returns the result of dividing the extended double-precision floating-point 3949 | value `a' by the corresponding value `b'. The operation is performed 3950 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 3951 *----------------------------------------------------------------------------*/ 3952 3953 floatx80 floatx80_div( floatx80 a, floatx80 b STATUS_PARAM ) 3954 { 3955 flag aSign, bSign, zSign; 3956 int32 aExp, bExp, zExp; 3957 bits64 aSig, bSig, zSig0, zSig1; 3958 bits64 rem0, rem1, rem2, term0, term1, term2; 3959 floatx80 z; 3960 3961 aSig = extractFloatx80Frac( a ); 3962 aExp = extractFloatx80Exp( a ); 3963 aSign = extractFloatx80Sign( a ); 3964 bSig = extractFloatx80Frac( b ); 3965 bExp = extractFloatx80Exp( b ); 3966 bSign = extractFloatx80Sign( b ); 3967 zSign = aSign ^ bSign; 3968 if ( aExp == 0x7FFF ) { 3969 if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3970 if ( bExp == 0x7FFF ) { 3971 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3972 goto invalid; 3973 } 3974 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3975 } 3976 if ( bExp == 0x7FFF ) { 3977 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 3978 return packFloatx80( zSign, 0, 0 ); 3979 } 3980 if ( bExp == 0 ) { 3981 if ( bSig == 0 ) { 3982 if ( ( aExp | aSig ) == 0 ) { 3983 invalid: 3984 float_raise( float_flag_invalid STATUS_VAR); 3985 z.low = floatx80_default_nan_low; 3986 z.high = floatx80_default_nan_high; 3987 return z; 3988 } 3989 float_raise( float_flag_divbyzero STATUS_VAR); 3990 return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 3991 } 3992 normalizeFloatx80Subnormal( bSig, &bExp, &bSig ); 3993 } 3994 if ( aExp == 0 ) { 3995 if ( aSig == 0 ) return packFloatx80( zSign, 0, 0 ); 3996 normalizeFloatx80Subnormal( aSig, &aExp, &aSig ); 3997 } 3998 zExp = aExp - bExp + 0x3FFE; 3999 rem1 = 0; 4000 if ( bSig <= aSig ) { 4001 shift128Right( aSig, 0, 1, &aSig, &rem1 ); 4002 ++zExp; 4003 } 4004 zSig0 = estimateDiv128To64( aSig, rem1, bSig ); 4005 mul64To128( bSig, zSig0, &term0, &term1 ); 4006 sub128( aSig, rem1, term0, term1, &rem0, &rem1 ); 4007 while ( (sbits64) rem0 < 0 ) { 4008 --zSig0; 4009 add128( rem0, rem1, 0, bSig, &rem0, &rem1 ); 4010 } 4011 zSig1 = estimateDiv128To64( rem1, 0, bSig ); 4012 if ( (bits64) ( zSig1<<1 ) <= 8 ) { 4013 mul64To128( bSig, zSig1, &term1, &term2 ); 4014 sub128( rem1, 0, term1, term2, &rem1, &rem2 ); 4015 while ( (sbits64) rem1 < 0 ) { 4016 --zSig1; 4017 add128( rem1, rem2, 0, bSig, &rem1, &rem2 ); 4018 } 4019 zSig1 |= ( ( rem1 | rem2 ) != 0 ); 4020 } 4021 return 4022 roundAndPackFloatx80( 4023 STATUS(floatx80_rounding_precision), zSign, zExp, zSig0, zSig1 STATUS_VAR ); 4024 4025 } 4026 4027 /*---------------------------------------------------------------------------- 4028 | Returns the remainder of the extended double-precision floating-point value 4029 | `a' with respect to the corresponding value `b'. The operation is performed 4030 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 4031 *----------------------------------------------------------------------------*/ 4032 4033 floatx80 floatx80_rem( floatx80 a, floatx80 b STATUS_PARAM ) 4034 { 4035 flag aSign, bSign, zSign; 4036 int32 aExp, bExp, expDiff; 4037 bits64 aSig0, aSig1, bSig; 4038 bits64 q, term0, term1, alternateASig0, alternateASig1; 4039 floatx80 z; 4040 4041 aSig0 = extractFloatx80Frac( a ); 4042 aExp = extractFloatx80Exp( a ); 4043 aSign = extractFloatx80Sign( a ); 4044 bSig = extractFloatx80Frac( b ); 4045 bExp = extractFloatx80Exp( b ); 4046 bSign = extractFloatx80Sign( b ); 4047 if ( aExp == 0x7FFF ) { 4048 if ( (bits64) ( aSig0<<1 ) 4049 || ( ( bExp == 0x7FFF ) && (bits64) ( bSig<<1 ) ) ) { 4050 return propagateFloatx80NaN( a, b STATUS_VAR ); 4051 } 4052 goto invalid; 4053 } 4054 if ( bExp == 0x7FFF ) { 4055 if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b STATUS_VAR ); 4056 return a; 4057 } 4058 if ( bExp == 0 ) { 4059 if ( bSig == 0 ) { 4060 invalid: 4061 float_raise( float_flag_invalid STATUS_VAR); 4062 z.low = floatx80_default_nan_low; 4063 z.high = floatx80_default_nan_high; 4064 return z; 4065 } 4066 normalizeFloatx80Subnormal( bSig, &bExp, &bSig ); 4067 } 4068 if ( aExp == 0 ) { 4069 if ( (bits64) ( aSig0<<1 ) == 0 ) return a; 4070 normalizeFloatx80Subnormal( aSig0, &aExp, &aSig0 ); 4071 } 4072 bSig |= LIT64( 0x8000000000000000 ); 4073 zSign = aSign; 4074 expDiff = aExp - bExp; 4075 aSig1 = 0; 4076 if ( expDiff < 0 ) { 4077 if ( expDiff < -1 ) return a; 4078 shift128Right( aSig0, 0, 1, &aSig0, &aSig1 ); 4079 expDiff = 0; 4080 } 4081 q = ( bSig <= aSig0 ); 4082 if ( q ) aSig0 -= bSig; 4083 expDiff -= 64; 4084 while ( 0 < expDiff ) { 4085 q = estimateDiv128To64( aSig0, aSig1, bSig ); 4086 q = ( 2 < q ) ? q - 2 : 0; 4087 mul64To128( bSig, q, &term0, &term1 ); 4088 sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 ); 4089 shortShift128Left( aSig0, aSig1, 62, &aSig0, &aSig1 ); 4090 expDiff -= 62; 4091 } 4092 expDiff += 64; 4093 if ( 0 < expDiff ) { 4094 q = estimateDiv128To64( aSig0, aSig1, bSig ); 4095 q = ( 2 < q ) ? q - 2 : 0; 4096 q >>= 64 - expDiff; 4097 mul64To128( bSig, q<<( 64 - expDiff ), &term0, &term1 ); 4098 sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 ); 4099 shortShift128Left( 0, bSig, 64 - expDiff, &term0, &term1 ); 4100 while ( le128( term0, term1, aSig0, aSig1 ) ) { 4101 ++q; 4102 sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 ); 4103 } 4104 } 4105 else { 4106 term1 = 0; 4107 term0 = bSig; 4108 } 4109 sub128( term0, term1, aSig0, aSig1, &alternateASig0, &alternateASig1 ); 4110 if ( lt128( alternateASig0, alternateASig1, aSig0, aSig1 ) 4111 || ( eq128( alternateASig0, alternateASig1, aSig0, aSig1 ) 4112 && ( q & 1 ) ) 4113 ) { 4114 aSig0 = alternateASig0; 4115 aSig1 = alternateASig1; 4116 zSign = ! zSign; 4117 } 4118 return 4119 normalizeRoundAndPackFloatx80( 4120 80, zSign, bExp + expDiff, aSig0, aSig1 STATUS_VAR ); 4121 4122 } 4123 4124 /*---------------------------------------------------------------------------- 4125 | Returns the square root of the extended double-precision floating-point 4126 | value `a'. The operation is performed according to the IEC/IEEE Standard 4127 | for Binary Floating-Point Arithmetic. 4128 *----------------------------------------------------------------------------*/ 4129 4130 floatx80 floatx80_sqrt( floatx80 a STATUS_PARAM ) 4131 { 4132 flag aSign; 4133 int32 aExp, zExp; 4134 bits64 aSig0, aSig1, zSig0, zSig1, doubleZSig0; 4135 bits64 rem0, rem1, rem2, rem3, term0, term1, term2, term3; 4136 floatx80 z; 4137 4138 aSig0 = extractFloatx80Frac( a ); 4139 aExp = extractFloatx80Exp( a ); 4140 aSign = extractFloatx80Sign( a ); 4141 if ( aExp == 0x7FFF ) { 4142 if ( (bits64) ( aSig0<<1 ) ) return propagateFloatx80NaN( a, a STATUS_VAR ); 4143 if ( ! aSign ) return a; 4144 goto invalid; 4145 } 4146 if ( aSign ) { 4147 if ( ( aExp | aSig0 ) == 0 ) return a; 4148 invalid: 4149 float_raise( float_flag_invalid STATUS_VAR); 4150 z.low = floatx80_default_nan_low; 4151 z.high = floatx80_default_nan_high; 4152 return z; 4153 } 4154 if ( aExp == 0 ) { 4155 if ( aSig0 == 0 ) return packFloatx80( 0, 0, 0 ); 4156 normalizeFloatx80Subnormal( aSig0, &aExp, &aSig0 ); 4157 } 4158 zExp = ( ( aExp - 0x3FFF )>>1 ) + 0x3FFF; 4159 zSig0 = estimateSqrt32( aExp, aSig0>>32 ); 4160 shift128Right( aSig0, 0, 2 + ( aExp & 1 ), &aSig0, &aSig1 ); 4161 zSig0 = estimateDiv128To64( aSig0, aSig1, zSig0<<32 ) + ( zSig0<<30 ); 4162 doubleZSig0 = zSig0<<1; 4163 mul64To128( zSig0, zSig0, &term0, &term1 ); 4164 sub128( aSig0, aSig1, term0, term1, &rem0, &rem1 ); 4165 while ( (sbits64) rem0 < 0 ) { 4166 --zSig0; 4167 doubleZSig0 -= 2; 4168 add128( rem0, rem1, zSig0>>63, doubleZSig0 | 1, &rem0, &rem1 ); 4169 } 4170 zSig1 = estimateDiv128To64( rem1, 0, doubleZSig0 ); 4171 if ( ( zSig1 & LIT64( 0x3FFFFFFFFFFFFFFF ) ) <= 5 ) { 4172 if ( zSig1 == 0 ) zSig1 = 1; 4173 mul64To128( doubleZSig0, zSig1, &term1, &term2 ); 4174 sub128( rem1, 0, term1, term2, &rem1, &rem2 ); 4175 mul64To128( zSig1, zSig1, &term2, &term3 ); 4176 sub192( rem1, rem2, 0, 0, term2, term3, &rem1, &rem2, &rem3 ); 4177 while ( (sbits64) rem1 < 0 ) { 4178 --zSig1; 4179 shortShift128Left( 0, zSig1, 1, &term2, &term3 ); 4180 term3 |= 1; 4181 term2 |= doubleZSig0; 4182 add192( rem1, rem2, rem3, 0, term2, term3, &rem1, &rem2, &rem3 ); 4183 } 4184 zSig1 |= ( ( rem1 | rem2 | rem3 ) != 0 ); 4185 } 4186 shortShift128Left( 0, zSig1, 1, &zSig0, &zSig1 ); 4187 zSig0 |= doubleZSig0; 4188 return 4189 roundAndPackFloatx80( 4190 STATUS(floatx80_rounding_precision), 0, zExp, zSig0, zSig1 STATUS_VAR ); 4191 4192 } 4193 4194 /*---------------------------------------------------------------------------- 4195 | Returns 1 if the extended double-precision floating-point value `a' is 4196 | equal to the corresponding value `b', and 0 otherwise. The comparison is 4197 | performed according to the IEC/IEEE Standard for Binary Floating-Point 4198 | Arithmetic. 4199 *----------------------------------------------------------------------------*/ 4200 4201 int floatx80_eq( floatx80 a, floatx80 b STATUS_PARAM ) 4202 { 4203 4204 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4205 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4206 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4207 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4208 ) { 4209 if ( floatx80_is_signaling_nan( a ) 4210 || floatx80_is_signaling_nan( b ) ) { 4211 float_raise( float_flag_invalid STATUS_VAR); 4212 } 4213 return 0; 4214 } 4215 return 4216 ( a.low == b.low ) 4217 && ( ( a.high == b.high ) 4218 || ( ( a.low == 0 ) 4219 && ( (bits16) ( ( a.high | b.high )<<1 ) == 0 ) ) 4220 ); 4221 4222 } 4223 4224 /*---------------------------------------------------------------------------- 4225 | Returns 1 if the extended double-precision floating-point value `a' is 4226 | less than or equal to the corresponding value `b', and 0 otherwise. The 4227 | comparison is performed according to the IEC/IEEE Standard for Binary 4228 | Floating-Point Arithmetic. 4229 *----------------------------------------------------------------------------*/ 4230 4231 int floatx80_le( floatx80 a, floatx80 b STATUS_PARAM ) 4232 { 4233 flag aSign, bSign; 4234 4235 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4236 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4237 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4238 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4239 ) { 4240 float_raise( float_flag_invalid STATUS_VAR); 4241 return 0; 4242 } 4243 aSign = extractFloatx80Sign( a ); 4244 bSign = extractFloatx80Sign( b ); 4245 if ( aSign != bSign ) { 4246 return 4247 aSign 4248 || ( ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 4249 == 0 ); 4250 } 4251 return 4252 aSign ? le128( b.high, b.low, a.high, a.low ) 4253 : le128( a.high, a.low, b.high, b.low ); 4254 4255 } 4256 4257 /*---------------------------------------------------------------------------- 4258 | Returns 1 if the extended double-precision floating-point value `a' is 4259 | less than the corresponding value `b', and 0 otherwise. The comparison 4260 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4261 | Arithmetic. 4262 *----------------------------------------------------------------------------*/ 4263 4264 int floatx80_lt( floatx80 a, floatx80 b STATUS_PARAM ) 4265 { 4266 flag aSign, bSign; 4267 4268 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4269 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4270 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4271 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4272 ) { 4273 float_raise( float_flag_invalid STATUS_VAR); 4274 return 0; 4275 } 4276 aSign = extractFloatx80Sign( a ); 4277 bSign = extractFloatx80Sign( b ); 4278 if ( aSign != bSign ) { 4279 return 4280 aSign 4281 && ( ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 4282 != 0 ); 4283 } 4284 return 4285 aSign ? lt128( b.high, b.low, a.high, a.low ) 4286 : lt128( a.high, a.low, b.high, b.low ); 4287 4288 } 4289 4290 /*---------------------------------------------------------------------------- 4291 | Returns 1 if the extended double-precision floating-point value `a' is equal 4292 | to the corresponding value `b', and 0 otherwise. The invalid exception is 4293 | raised if either operand is a NaN. Otherwise, the comparison is performed 4294 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 4295 *----------------------------------------------------------------------------*/ 4296 4297 int floatx80_eq_signaling( floatx80 a, floatx80 b STATUS_PARAM ) 4298 { 4299 4300 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4301 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4302 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4303 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4304 ) { 4305 float_raise( float_flag_invalid STATUS_VAR); 4306 return 0; 4307 } 4308 return 4309 ( a.low == b.low ) 4310 && ( ( a.high == b.high ) 4311 || ( ( a.low == 0 ) 4312 && ( (bits16) ( ( a.high | b.high )<<1 ) == 0 ) ) 4313 ); 4314 4315 } 4316 4317 /*---------------------------------------------------------------------------- 4318 | Returns 1 if the extended double-precision floating-point value `a' is less 4319 | than or equal to the corresponding value `b', and 0 otherwise. Quiet NaNs 4320 | do not cause an exception. Otherwise, the comparison is performed according 4321 | to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 4322 *----------------------------------------------------------------------------*/ 4323 4324 int floatx80_le_quiet( floatx80 a, floatx80 b STATUS_PARAM ) 4325 { 4326 flag aSign, bSign; 4327 4328 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4329 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4330 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4331 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4332 ) { 4333 if ( floatx80_is_signaling_nan( a ) 4334 || floatx80_is_signaling_nan( b ) ) { 4335 float_raise( float_flag_invalid STATUS_VAR); 4336 } 4337 return 0; 4338 } 4339 aSign = extractFloatx80Sign( a ); 4340 bSign = extractFloatx80Sign( b ); 4341 if ( aSign != bSign ) { 4342 return 4343 aSign 4344 || ( ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 4345 == 0 ); 4346 } 4347 return 4348 aSign ? le128( b.high, b.low, a.high, a.low ) 4349 : le128( a.high, a.low, b.high, b.low ); 4350 4351 } 4352 4353 /*---------------------------------------------------------------------------- 4354 | Returns 1 if the extended double-precision floating-point value `a' is less 4355 | than the corresponding value `b', and 0 otherwise. Quiet NaNs do not cause 4356 | an exception. Otherwise, the comparison is performed according to the 4357 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 4358 *----------------------------------------------------------------------------*/ 4359 4360 int floatx80_lt_quiet( floatx80 a, floatx80 b STATUS_PARAM ) 4361 { 4362 flag aSign, bSign; 4363 4364 if ( ( ( extractFloatx80Exp( a ) == 0x7FFF ) 4365 && (bits64) ( extractFloatx80Frac( a )<<1 ) ) 4366 || ( ( extractFloatx80Exp( b ) == 0x7FFF ) 4367 && (bits64) ( extractFloatx80Frac( b )<<1 ) ) 4368 ) { 4369 if ( floatx80_is_signaling_nan( a ) 4370 || floatx80_is_signaling_nan( b ) ) { 4371 float_raise( float_flag_invalid STATUS_VAR); 4372 } 4373 return 0; 4374 } 4375 aSign = extractFloatx80Sign( a ); 4376 bSign = extractFloatx80Sign( b ); 4377 if ( aSign != bSign ) { 4378 return 4379 aSign 4380 && ( ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 4381 != 0 ); 4382 } 4383 return 4384 aSign ? lt128( b.high, b.low, a.high, a.low ) 4385 : lt128( a.high, a.low, b.high, b.low ); 4386 4387 } 4388 4389 #endif 4390 4391 #ifdef FLOAT128 4392 4393 /*---------------------------------------------------------------------------- 4394 | Returns the result of converting the quadruple-precision floating-point 4395 | value `a' to the 32-bit two's complement integer format. The conversion 4396 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4397 | Arithmetic---which means in particular that the conversion is rounded 4398 | according to the current rounding mode. If `a' is a NaN, the largest 4399 | positive integer is returned. Otherwise, if the conversion overflows, the 4400 | largest integer with the same sign as `a' is returned. 4401 *----------------------------------------------------------------------------*/ 4402 4403 int32 float128_to_int32( float128 a STATUS_PARAM ) 4404 { 4405 flag aSign; 4406 int32 aExp, shiftCount; 4407 bits64 aSig0, aSig1; 4408 4409 aSig1 = extractFloat128Frac1( a ); 4410 aSig0 = extractFloat128Frac0( a ); 4411 aExp = extractFloat128Exp( a ); 4412 aSign = extractFloat128Sign( a ); 4413 if ( ( aExp == 0x7FFF ) && ( aSig0 | aSig1 ) ) aSign = 0; 4414 if ( aExp ) aSig0 |= LIT64( 0x0001000000000000 ); 4415 aSig0 |= ( aSig1 != 0 ); 4416 shiftCount = 0x4028 - aExp; 4417 if ( 0 < shiftCount ) shift64RightJamming( aSig0, shiftCount, &aSig0 ); 4418 return roundAndPackInt32( aSign, aSig0 STATUS_VAR ); 4419 4420 } 4421 4422 /*---------------------------------------------------------------------------- 4423 | Returns the result of converting the quadruple-precision floating-point 4424 | value `a' to the 32-bit two's complement integer format. The conversion 4425 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4426 | Arithmetic, except that the conversion is always rounded toward zero. If 4427 | `a' is a NaN, the largest positive integer is returned. Otherwise, if the 4428 | conversion overflows, the largest integer with the same sign as `a' is 4429 | returned. 4430 *----------------------------------------------------------------------------*/ 4431 4432 int32 float128_to_int32_round_to_zero( float128 a STATUS_PARAM ) 4433 { 4434 flag aSign; 4435 int32 aExp, shiftCount; 4436 bits64 aSig0, aSig1, savedASig; 4437 int32 z; 4438 4439 aSig1 = extractFloat128Frac1( a ); 4440 aSig0 = extractFloat128Frac0( a ); 4441 aExp = extractFloat128Exp( a ); 4442 aSign = extractFloat128Sign( a ); 4443 aSig0 |= ( aSig1 != 0 ); 4444 if ( 0x401E < aExp ) { 4445 if ( ( aExp == 0x7FFF ) && aSig0 ) aSign = 0; 4446 goto invalid; 4447 } 4448 else if ( aExp < 0x3FFF ) { 4449 if ( aExp || aSig0 ) STATUS(float_exception_flags) |= float_flag_inexact; 4450 return 0; 4451 } 4452 aSig0 |= LIT64( 0x0001000000000000 ); 4453 shiftCount = 0x402F - aExp; 4454 savedASig = aSig0; 4455 aSig0 >>= shiftCount; 4456 z = aSig0; 4457 if ( aSign ) z = - z; 4458 if ( ( z < 0 ) ^ aSign ) { 4459 invalid: 4460 float_raise( float_flag_invalid STATUS_VAR); 4461 return aSign ? (sbits32) 0x80000000 : 0x7FFFFFFF; 4462 } 4463 if ( ( aSig0<<shiftCount ) != savedASig ) { 4464 STATUS(float_exception_flags) |= float_flag_inexact; 4465 } 4466 return z; 4467 4468 } 4469 4470 /*---------------------------------------------------------------------------- 4471 | Returns the result of converting the quadruple-precision floating-point 4472 | value `a' to the 64-bit two's complement integer format. The conversion 4473 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4474 | Arithmetic---which means in particular that the conversion is rounded 4475 | according to the current rounding mode. If `a' is a NaN, the largest 4476 | positive integer is returned. Otherwise, if the conversion overflows, the 4477 | largest integer with the same sign as `a' is returned. 4478 *----------------------------------------------------------------------------*/ 4479 4480 int64 float128_to_int64( float128 a STATUS_PARAM ) 4481 { 4482 flag aSign; 4483 int32 aExp, shiftCount; 4484 bits64 aSig0, aSig1; 4485 4486 aSig1 = extractFloat128Frac1( a ); 4487 aSig0 = extractFloat128Frac0( a ); 4488 aExp = extractFloat128Exp( a ); 4489 aSign = extractFloat128Sign( a ); 4490 if ( aExp ) aSig0 |= LIT64( 0x0001000000000000 ); 4491 shiftCount = 0x402F - aExp; 4492 if ( shiftCount <= 0 ) { 4493 if ( 0x403E < aExp ) { 4494 float_raise( float_flag_invalid STATUS_VAR); 4495 if ( ! aSign 4496 || ( ( aExp == 0x7FFF ) 4497 && ( aSig1 || ( aSig0 != LIT64( 0x0001000000000000 ) ) ) 4498 ) 4499 ) { 4500 return LIT64( 0x7FFFFFFFFFFFFFFF ); 4501 } 4502 return (sbits64) LIT64( 0x8000000000000000 ); 4503 } 4504 shortShift128Left( aSig0, aSig1, - shiftCount, &aSig0, &aSig1 ); 4505 } 4506 else { 4507 shift64ExtraRightJamming( aSig0, aSig1, shiftCount, &aSig0, &aSig1 ); 4508 } 4509 return roundAndPackInt64( aSign, aSig0, aSig1 STATUS_VAR ); 4510 4511 } 4512 4513 /*---------------------------------------------------------------------------- 4514 | Returns the result of converting the quadruple-precision floating-point 4515 | value `a' to the 64-bit two's complement integer format. The conversion 4516 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4517 | Arithmetic, except that the conversion is always rounded toward zero. 4518 | If `a' is a NaN, the largest positive integer is returned. Otherwise, if 4519 | the conversion overflows, the largest integer with the same sign as `a' is 4520 | returned. 4521 *----------------------------------------------------------------------------*/ 4522 4523 int64 float128_to_int64_round_to_zero( float128 a STATUS_PARAM ) 4524 { 4525 flag aSign; 4526 int32 aExp, shiftCount; 4527 bits64 aSig0, aSig1; 4528 int64 z; 4529 4530 aSig1 = extractFloat128Frac1( a ); 4531 aSig0 = extractFloat128Frac0( a ); 4532 aExp = extractFloat128Exp( a ); 4533 aSign = extractFloat128Sign( a ); 4534 if ( aExp ) aSig0 |= LIT64( 0x0001000000000000 ); 4535 shiftCount = aExp - 0x402F; 4536 if ( 0 < shiftCount ) { 4537 if ( 0x403E <= aExp ) { 4538 aSig0 &= LIT64( 0x0000FFFFFFFFFFFF ); 4539 if ( ( a.high == LIT64( 0xC03E000000000000 ) ) 4540 && ( aSig1 < LIT64( 0x0002000000000000 ) ) ) { 4541 if ( aSig1 ) STATUS(float_exception_flags) |= float_flag_inexact; 4542 } 4543 else { 4544 float_raise( float_flag_invalid STATUS_VAR); 4545 if ( ! aSign || ( ( aExp == 0x7FFF ) && ( aSig0 | aSig1 ) ) ) { 4546 return LIT64( 0x7FFFFFFFFFFFFFFF ); 4547 } 4548 } 4549 return (sbits64) LIT64( 0x8000000000000000 ); 4550 } 4551 z = ( aSig0<<shiftCount ) | ( aSig1>>( ( - shiftCount ) & 63 ) ); 4552 if ( (bits64) ( aSig1<<shiftCount ) ) { 4553 STATUS(float_exception_flags) |= float_flag_inexact; 4554 } 4555 } 4556 else { 4557 if ( aExp < 0x3FFF ) { 4558 if ( aExp | aSig0 | aSig1 ) { 4559 STATUS(float_exception_flags) |= float_flag_inexact; 4560 } 4561 return 0; 4562 } 4563 z = aSig0>>( - shiftCount ); 4564 if ( aSig1 4565 || ( shiftCount && (bits64) ( aSig0<<( shiftCount & 63 ) ) ) ) { 4566 STATUS(float_exception_flags) |= float_flag_inexact; 4567 } 4568 } 4569 if ( aSign ) z = - z; 4570 return z; 4571 4572 } 4573 4574 /*---------------------------------------------------------------------------- 4575 | Returns the result of converting the quadruple-precision floating-point 4576 | value `a' to the single-precision floating-point format. The conversion 4577 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4578 | Arithmetic. 4579 *----------------------------------------------------------------------------*/ 4580 4581 float32 float128_to_float32( float128 a STATUS_PARAM ) 4582 { 4583 flag aSign; 4584 int32 aExp; 4585 bits64 aSig0, aSig1; 4586 bits32 zSig; 4587 4588 aSig1 = extractFloat128Frac1( a ); 4589 aSig0 = extractFloat128Frac0( a ); 4590 aExp = extractFloat128Exp( a ); 4591 aSign = extractFloat128Sign( a ); 4592 if ( aExp == 0x7FFF ) { 4593 if ( aSig0 | aSig1 ) { 4594 return commonNaNToFloat32( float128ToCommonNaN( a STATUS_VAR ) ); 4595 } 4596 return packFloat32( aSign, 0xFF, 0 ); 4597 } 4598 aSig0 |= ( aSig1 != 0 ); 4599 shift64RightJamming( aSig0, 18, &aSig0 ); 4600 zSig = aSig0; 4601 if ( aExp || zSig ) { 4602 zSig |= 0x40000000; 4603 aExp -= 0x3F81; 4604 } 4605 return roundAndPackFloat32( aSign, aExp, zSig STATUS_VAR ); 4606 4607 } 4608 4609 /*---------------------------------------------------------------------------- 4610 | Returns the result of converting the quadruple-precision floating-point 4611 | value `a' to the double-precision floating-point format. The conversion 4612 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 4613 | Arithmetic. 4614 *----------------------------------------------------------------------------*/ 4615 4616 float64 float128_to_float64( float128 a STATUS_PARAM ) 4617 { 4618 flag aSign; 4619 int32 aExp; 4620 bits64 aSig0, aSig1; 4621 4622 aSig1 = extractFloat128Frac1( a ); 4623 aSig0 = extractFloat128Frac0( a ); 4624 aExp = extractFloat128Exp( a ); 4625 aSign = extractFloat128Sign( a ); 4626 if ( aExp == 0x7FFF ) { 4627 if ( aSig0 | aSig1 ) { 4628 return commonNaNToFloat64( float128ToCommonNaN( a STATUS_VAR ) ); 4629 } 4630 return packFloat64( aSign, 0x7FF, 0 ); 4631 } 4632 shortShift128Left( aSig0, aSig1, 14, &aSig0, &aSig1 ); 4633 aSig0 |= ( aSig1 != 0 ); 4634 if ( aExp || aSig0 ) { 4635 aSig0 |= LIT64( 0x4000000000000000 ); 4636 aExp -= 0x3C01; 4637 } 4638 return roundAndPackFloat64( aSign, aExp, aSig0 STATUS_VAR ); 4639 4640 } 4641 4642 #ifdef FLOATX80 4643 4644 /*---------------------------------------------------------------------------- 4645 | Returns the result of converting the quadruple-precision floating-point 4646 | value `a' to the extended double-precision floating-point format. The 4647 | conversion is performed according to the IEC/IEEE Standard for Binary 4648 | Floating-Point Arithmetic. 4649 *----------------------------------------------------------------------------*/ 4650 4651 floatx80 float128_to_floatx80( float128 a STATUS_PARAM ) 4652 { 4653 flag aSign; 4654 int32 aExp; 4655 bits64 aSig0, aSig1; 4656 4657 aSig1 = extractFloat128Frac1( a ); 4658 aSig0 = extractFloat128Frac0( a ); 4659 aExp = extractFloat128Exp( a ); 4660 aSign = extractFloat128Sign( a ); 4661 if ( aExp == 0x7FFF ) { 4662 if ( aSig0 | aSig1 ) { 4663 return commonNaNToFloatx80( float128ToCommonNaN( a STATUS_VAR ) ); 4664 } 4665 return packFloatx80( aSign, 0x7FFF, LIT64( 0x8000000000000000 ) ); 4666 } 4667 if ( aExp == 0 ) { 4668 if ( ( aSig0 | aSig1 ) == 0 ) return packFloatx80( aSign, 0, 0 ); 4669 normalizeFloat128Subnormal( aSig0, aSig1, &aExp, &aSig0, &aSig1 ); 4670 } 4671 else { 4672 aSig0 |= LIT64( 0x0001000000000000 ); 4673 } 4674 shortShift128Left( aSig0, aSig1, 15, &aSig0, &aSig1 ); 4675 return roundAndPackFloatx80( 80, aSign, aExp, aSig0, aSig1 STATUS_VAR ); 4676 4677 } 4678 4679 #endif 4680 4681 /*---------------------------------------------------------------------------- 4682 | Rounds the quadruple-precision floating-point value `a' to an integer, and 4683 | returns the result as a quadruple-precision floating-point value. The 4684 | operation is performed according to the IEC/IEEE Standard for Binary 4685 | Floating-Point Arithmetic. 4686 *----------------------------------------------------------------------------*/ 4687 4688 float128 float128_round_to_int( float128 a STATUS_PARAM ) 4689 { 4690 flag aSign; 4691 int32 aExp; 4692 bits64 lastBitMask, roundBitsMask; 4693 int8 roundingMode; 4694 float128 z; 4695 4696 aExp = extractFloat128Exp( a ); 4697 if ( 0x402F <= aExp ) { 4698 if ( 0x406F <= aExp ) { 4699 if ( ( aExp == 0x7FFF ) 4700 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) 4701 ) { 4702 return propagateFloat128NaN( a, a STATUS_VAR ); 4703 } 4704 return a; 4705 } 4706 lastBitMask = 1; 4707 lastBitMask = ( lastBitMask<<( 0x406E - aExp ) )<<1; 4708 roundBitsMask = lastBitMask - 1; 4709 z = a; 4710 roundingMode = STATUS(float_rounding_mode); 4711 if ( roundingMode == float_round_nearest_even ) { 4712 if ( lastBitMask ) { 4713 add128( z.high, z.low, 0, lastBitMask>>1, &z.high, &z.low ); 4714 if ( ( z.low & roundBitsMask ) == 0 ) z.low &= ~ lastBitMask; 4715 } 4716 else { 4717 if ( (sbits64) z.low < 0 ) { 4718 ++z.high; 4719 if ( (bits64) ( z.low<<1 ) == 0 ) z.high &= ~1; 4720 } 4721 } 4722 } 4723 else if ( roundingMode != float_round_to_zero ) { 4724 if ( extractFloat128Sign( z ) 4725 ^ ( roundingMode == float_round_up ) ) { 4726 add128( z.high, z.low, 0, roundBitsMask, &z.high, &z.low ); 4727 } 4728 } 4729 z.low &= ~ roundBitsMask; 4730 } 4731 else { 4732 if ( aExp < 0x3FFF ) { 4733 if ( ( ( (bits64) ( a.high<<1 ) ) | a.low ) == 0 ) return a; 4734 STATUS(float_exception_flags) |= float_flag_inexact; 4735 aSign = extractFloat128Sign( a ); 4736 switch ( STATUS(float_rounding_mode) ) { 4737 case float_round_nearest_even: 4738 if ( ( aExp == 0x3FFE ) 4739 && ( extractFloat128Frac0( a ) 4740 | extractFloat128Frac1( a ) ) 4741 ) { 4742 return packFloat128( aSign, 0x3FFF, 0, 0 ); 4743 } 4744 break; 4745 case float_round_down: 4746 return 4747 aSign ? packFloat128( 1, 0x3FFF, 0, 0 ) 4748 : packFloat128( 0, 0, 0, 0 ); 4749 case float_round_up: 4750 return 4751 aSign ? packFloat128( 1, 0, 0, 0 ) 4752 : packFloat128( 0, 0x3FFF, 0, 0 ); 4753 } 4754 return packFloat128( aSign, 0, 0, 0 ); 4755 } 4756 lastBitMask = 1; 4757 lastBitMask <<= 0x402F - aExp; 4758 roundBitsMask = lastBitMask - 1; 4759 z.low = 0; 4760 z.high = a.high; 4761 roundingMode = STATUS(float_rounding_mode); 4762 if ( roundingMode == float_round_nearest_even ) { 4763 z.high += lastBitMask>>1; 4764 if ( ( ( z.high & roundBitsMask ) | a.low ) == 0 ) { 4765 z.high &= ~ lastBitMask; 4766 } 4767 } 4768 else if ( roundingMode != float_round_to_zero ) { 4769 if ( extractFloat128Sign( z ) 4770 ^ ( roundingMode == float_round_up ) ) { 4771 z.high |= ( a.low != 0 ); 4772 z.high += roundBitsMask; 4773 } 4774 } 4775 z.high &= ~ roundBitsMask; 4776 } 4777 if ( ( z.low != a.low ) || ( z.high != a.high ) ) { 4778 STATUS(float_exception_flags) |= float_flag_inexact; 4779 } 4780 return z; 4781 4782 } 4783 4784 /*---------------------------------------------------------------------------- 4785 | Returns the result of adding the absolute values of the quadruple-precision 4786 | floating-point values `a' and `b'. If `zSign' is 1, the sum is negated 4787 | before being returned. `zSign' is ignored if the result is a NaN. 4788 | The addition is performed according to the IEC/IEEE Standard for Binary 4789 | Floating-Point Arithmetic. 4790 *----------------------------------------------------------------------------*/ 4791 4792 static float128 addFloat128Sigs( float128 a, float128 b, flag zSign STATUS_PARAM) 4793 { 4794 int32 aExp, bExp, zExp; 4795 bits64 aSig0, aSig1, bSig0, bSig1, zSig0, zSig1, zSig2; 4796 int32 expDiff; 4797 4798 aSig1 = extractFloat128Frac1( a ); 4799 aSig0 = extractFloat128Frac0( a ); 4800 aExp = extractFloat128Exp( a ); 4801 bSig1 = extractFloat128Frac1( b ); 4802 bSig0 = extractFloat128Frac0( b ); 4803 bExp = extractFloat128Exp( b ); 4804 expDiff = aExp - bExp; 4805 if ( 0 < expDiff ) { 4806 if ( aExp == 0x7FFF ) { 4807 if ( aSig0 | aSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 4808 return a; 4809 } 4810 if ( bExp == 0 ) { 4811 --expDiff; 4812 } 4813 else { 4814 bSig0 |= LIT64( 0x0001000000000000 ); 4815 } 4816 shift128ExtraRightJamming( 4817 bSig0, bSig1, 0, expDiff, &bSig0, &bSig1, &zSig2 ); 4818 zExp = aExp; 4819 } 4820 else if ( expDiff < 0 ) { 4821 if ( bExp == 0x7FFF ) { 4822 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 4823 return packFloat128( zSign, 0x7FFF, 0, 0 ); 4824 } 4825 if ( aExp == 0 ) { 4826 ++expDiff; 4827 } 4828 else { 4829 aSig0 |= LIT64( 0x0001000000000000 ); 4830 } 4831 shift128ExtraRightJamming( 4832 aSig0, aSig1, 0, - expDiff, &aSig0, &aSig1, &zSig2 ); 4833 zExp = bExp; 4834 } 4835 else { 4836 if ( aExp == 0x7FFF ) { 4837 if ( aSig0 | aSig1 | bSig0 | bSig1 ) { 4838 return propagateFloat128NaN( a, b STATUS_VAR ); 4839 } 4840 return a; 4841 } 4842 add128( aSig0, aSig1, bSig0, bSig1, &zSig0, &zSig1 ); 4843 if ( aExp == 0 ) { 4844 if ( STATUS(flush_to_zero) ) return packFloat128( zSign, 0, 0, 0 ); 4845 return packFloat128( zSign, 0, zSig0, zSig1 ); 4846 } 4847 zSig2 = 0; 4848 zSig0 |= LIT64( 0x0002000000000000 ); 4849 zExp = aExp; 4850 goto shiftRight1; 4851 } 4852 aSig0 |= LIT64( 0x0001000000000000 ); 4853 add128( aSig0, aSig1, bSig0, bSig1, &zSig0, &zSig1 ); 4854 --zExp; 4855 if ( zSig0 < LIT64( 0x0002000000000000 ) ) goto roundAndPack; 4856 ++zExp; 4857 shiftRight1: 4858 shift128ExtraRightJamming( 4859 zSig0, zSig1, zSig2, 1, &zSig0, &zSig1, &zSig2 ); 4860 roundAndPack: 4861 return roundAndPackFloat128( zSign, zExp, zSig0, zSig1, zSig2 STATUS_VAR ); 4862 4863 } 4864 4865 /*---------------------------------------------------------------------------- 4866 | Returns the result of subtracting the absolute values of the quadruple- 4867 | precision floating-point values `a' and `b'. If `zSign' is 1, the 4868 | difference is negated before being returned. `zSign' is ignored if the 4869 | result is a NaN. The subtraction is performed according to the IEC/IEEE 4870 | Standard for Binary Floating-Point Arithmetic. 4871 *----------------------------------------------------------------------------*/ 4872 4873 static float128 subFloat128Sigs( float128 a, float128 b, flag zSign STATUS_PARAM) 4874 { 4875 int32 aExp, bExp, zExp; 4876 bits64 aSig0, aSig1, bSig0, bSig1, zSig0, zSig1; 4877 int32 expDiff; 4878 float128 z; 4879 4880 aSig1 = extractFloat128Frac1( a ); 4881 aSig0 = extractFloat128Frac0( a ); 4882 aExp = extractFloat128Exp( a ); 4883 bSig1 = extractFloat128Frac1( b ); 4884 bSig0 = extractFloat128Frac0( b ); 4885 bExp = extractFloat128Exp( b ); 4886 expDiff = aExp - bExp; 4887 shortShift128Left( aSig0, aSig1, 14, &aSig0, &aSig1 ); 4888 shortShift128Left( bSig0, bSig1, 14, &bSig0, &bSig1 ); 4889 if ( 0 < expDiff ) goto aExpBigger; 4890 if ( expDiff < 0 ) goto bExpBigger; 4891 if ( aExp == 0x7FFF ) { 4892 if ( aSig0 | aSig1 | bSig0 | bSig1 ) { 4893 return propagateFloat128NaN( a, b STATUS_VAR ); 4894 } 4895 float_raise( float_flag_invalid STATUS_VAR); 4896 z.low = float128_default_nan_low; 4897 z.high = float128_default_nan_high; 4898 return z; 4899 } 4900 if ( aExp == 0 ) { 4901 aExp = 1; 4902 bExp = 1; 4903 } 4904 if ( bSig0 < aSig0 ) goto aBigger; 4905 if ( aSig0 < bSig0 ) goto bBigger; 4906 if ( bSig1 < aSig1 ) goto aBigger; 4907 if ( aSig1 < bSig1 ) goto bBigger; 4908 return packFloat128( STATUS(float_rounding_mode) == float_round_down, 0, 0, 0 ); 4909 bExpBigger: 4910 if ( bExp == 0x7FFF ) { 4911 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 4912 return packFloat128( zSign ^ 1, 0x7FFF, 0, 0 ); 4913 } 4914 if ( aExp == 0 ) { 4915 ++expDiff; 4916 } 4917 else { 4918 aSig0 |= LIT64( 0x4000000000000000 ); 4919 } 4920 shift128RightJamming( aSig0, aSig1, - expDiff, &aSig0, &aSig1 ); 4921 bSig0 |= LIT64( 0x4000000000000000 ); 4922 bBigger: 4923 sub128( bSig0, bSig1, aSig0, aSig1, &zSig0, &zSig1 ); 4924 zExp = bExp; 4925 zSign ^= 1; 4926 goto normalizeRoundAndPack; 4927 aExpBigger: 4928 if ( aExp == 0x7FFF ) { 4929 if ( aSig0 | aSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 4930 return a; 4931 } 4932 if ( bExp == 0 ) { 4933 --expDiff; 4934 } 4935 else { 4936 bSig0 |= LIT64( 0x4000000000000000 ); 4937 } 4938 shift128RightJamming( bSig0, bSig1, expDiff, &bSig0, &bSig1 ); 4939 aSig0 |= LIT64( 0x4000000000000000 ); 4940 aBigger: 4941 sub128( aSig0, aSig1, bSig0, bSig1, &zSig0, &zSig1 ); 4942 zExp = aExp; 4943 normalizeRoundAndPack: 4944 --zExp; 4945 return normalizeRoundAndPackFloat128( zSign, zExp - 14, zSig0, zSig1 STATUS_VAR ); 4946 4947 } 4948 4949 /*---------------------------------------------------------------------------- 4950 | Returns the result of adding the quadruple-precision floating-point values 4951 | `a' and `b'. The operation is performed according to the IEC/IEEE Standard 4952 | for Binary Floating-Point Arithmetic. 4953 *----------------------------------------------------------------------------*/ 4954 4955 float128 float128_add( float128 a, float128 b STATUS_PARAM ) 4956 { 4957 flag aSign, bSign; 4958 4959 aSign = extractFloat128Sign( a ); 4960 bSign = extractFloat128Sign( b ); 4961 if ( aSign == bSign ) { 4962 return addFloat128Sigs( a, b, aSign STATUS_VAR ); 4963 } 4964 else { 4965 return subFloat128Sigs( a, b, aSign STATUS_VAR ); 4966 } 4967 4968 } 4969 4970 /*---------------------------------------------------------------------------- 4971 | Returns the result of subtracting the quadruple-precision floating-point 4972 | values `a' and `b'. The operation is performed according to the IEC/IEEE 4973 | Standard for Binary Floating-Point Arithmetic. 4974 *----------------------------------------------------------------------------*/ 4975 4976 float128 float128_sub( float128 a, float128 b STATUS_PARAM ) 4977 { 4978 flag aSign, bSign; 4979 4980 aSign = extractFloat128Sign( a ); 4981 bSign = extractFloat128Sign( b ); 4982 if ( aSign == bSign ) { 4983 return subFloat128Sigs( a, b, aSign STATUS_VAR ); 4984 } 4985 else { 4986 return addFloat128Sigs( a, b, aSign STATUS_VAR ); 4987 } 4988 4989 } 4990 4991 /*---------------------------------------------------------------------------- 4992 | Returns the result of multiplying the quadruple-precision floating-point 4993 | values `a' and `b'. The operation is performed according to the IEC/IEEE 4994 | Standard for Binary Floating-Point Arithmetic. 4995 *----------------------------------------------------------------------------*/ 4996 4997 float128 float128_mul( float128 a, float128 b STATUS_PARAM ) 4998 { 4999 flag aSign, bSign, zSign; 5000 int32 aExp, bExp, zExp; 5001 bits64 aSig0, aSig1, bSig0, bSig1, zSig0, zSig1, zSig2, zSig3; 5002 float128 z; 5003 5004 aSig1 = extractFloat128Frac1( a ); 5005 aSig0 = extractFloat128Frac0( a ); 5006 aExp = extractFloat128Exp( a ); 5007 aSign = extractFloat128Sign( a ); 5008 bSig1 = extractFloat128Frac1( b ); 5009 bSig0 = extractFloat128Frac0( b ); 5010 bExp = extractFloat128Exp( b ); 5011 bSign = extractFloat128Sign( b ); 5012 zSign = aSign ^ bSign; 5013 if ( aExp == 0x7FFF ) { 5014 if ( ( aSig0 | aSig1 ) 5015 || ( ( bExp == 0x7FFF ) && ( bSig0 | bSig1 ) ) ) { 5016 return propagateFloat128NaN( a, b STATUS_VAR ); 5017 } 5018 if ( ( bExp | bSig0 | bSig1 ) == 0 ) goto invalid; 5019 return packFloat128( zSign, 0x7FFF, 0, 0 ); 5020 } 5021 if ( bExp == 0x7FFF ) { 5022 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 5023 if ( ( aExp | aSig0 | aSig1 ) == 0 ) { 5024 invalid: 5025 float_raise( float_flag_invalid STATUS_VAR); 5026 z.low = float128_default_nan_low; 5027 z.high = float128_default_nan_high; 5028 return z; 5029 } 5030 return packFloat128( zSign, 0x7FFF, 0, 0 ); 5031 } 5032 if ( aExp == 0 ) { 5033 if ( ( aSig0 | aSig1 ) == 0 ) return packFloat128( zSign, 0, 0, 0 ); 5034 normalizeFloat128Subnormal( aSig0, aSig1, &aExp, &aSig0, &aSig1 ); 5035 } 5036 if ( bExp == 0 ) { 5037 if ( ( bSig0 | bSig1 ) == 0 ) return packFloat128( zSign, 0, 0, 0 ); 5038 normalizeFloat128Subnormal( bSig0, bSig1, &bExp, &bSig0, &bSig1 ); 5039 } 5040 zExp = aExp + bExp - 0x4000; 5041 aSig0 |= LIT64( 0x0001000000000000 ); 5042 shortShift128Left( bSig0, bSig1, 16, &bSig0, &bSig1 ); 5043 mul128To256( aSig0, aSig1, bSig0, bSig1, &zSig0, &zSig1, &zSig2, &zSig3 ); 5044 add128( zSig0, zSig1, aSig0, aSig1, &zSig0, &zSig1 ); 5045 zSig2 |= ( zSig3 != 0 ); 5046 if ( LIT64( 0x0002000000000000 ) <= zSig0 ) { 5047 shift128ExtraRightJamming( 5048 zSig0, zSig1, zSig2, 1, &zSig0, &zSig1, &zSig2 ); 5049 ++zExp; 5050 } 5051 return roundAndPackFloat128( zSign, zExp, zSig0, zSig1, zSig2 STATUS_VAR ); 5052 5053 } 5054 5055 /*---------------------------------------------------------------------------- 5056 | Returns the result of dividing the quadruple-precision floating-point value 5057 | `a' by the corresponding value `b'. The operation is performed according to 5058 | the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5059 *----------------------------------------------------------------------------*/ 5060 5061 float128 float128_div( float128 a, float128 b STATUS_PARAM ) 5062 { 5063 flag aSign, bSign, zSign; 5064 int32 aExp, bExp, zExp; 5065 bits64 aSig0, aSig1, bSig0, bSig1, zSig0, zSig1, zSig2; 5066 bits64 rem0, rem1, rem2, rem3, term0, term1, term2, term3; 5067 float128 z; 5068 5069 aSig1 = extractFloat128Frac1( a ); 5070 aSig0 = extractFloat128Frac0( a ); 5071 aExp = extractFloat128Exp( a ); 5072 aSign = extractFloat128Sign( a ); 5073 bSig1 = extractFloat128Frac1( b ); 5074 bSig0 = extractFloat128Frac0( b ); 5075 bExp = extractFloat128Exp( b ); 5076 bSign = extractFloat128Sign( b ); 5077 zSign = aSign ^ bSign; 5078 if ( aExp == 0x7FFF ) { 5079 if ( aSig0 | aSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 5080 if ( bExp == 0x7FFF ) { 5081 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 5082 goto invalid; 5083 } 5084 return packFloat128( zSign, 0x7FFF, 0, 0 ); 5085 } 5086 if ( bExp == 0x7FFF ) { 5087 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 5088 return packFloat128( zSign, 0, 0, 0 ); 5089 } 5090 if ( bExp == 0 ) { 5091 if ( ( bSig0 | bSig1 ) == 0 ) { 5092 if ( ( aExp | aSig0 | aSig1 ) == 0 ) { 5093 invalid: 5094 float_raise( float_flag_invalid STATUS_VAR); 5095 z.low = float128_default_nan_low; 5096 z.high = float128_default_nan_high; 5097 return z; 5098 } 5099 float_raise( float_flag_divbyzero STATUS_VAR); 5100 return packFloat128( zSign, 0x7FFF, 0, 0 ); 5101 } 5102 normalizeFloat128Subnormal( bSig0, bSig1, &bExp, &bSig0, &bSig1 ); 5103 } 5104 if ( aExp == 0 ) { 5105 if ( ( aSig0 | aSig1 ) == 0 ) return packFloat128( zSign, 0, 0, 0 ); 5106 normalizeFloat128Subnormal( aSig0, aSig1, &aExp, &aSig0, &aSig1 ); 5107 } 5108 zExp = aExp - bExp + 0x3FFD; 5109 shortShift128Left( 5110 aSig0 | LIT64( 0x0001000000000000 ), aSig1, 15, &aSig0, &aSig1 ); 5111 shortShift128Left( 5112 bSig0 | LIT64( 0x0001000000000000 ), bSig1, 15, &bSig0, &bSig1 ); 5113 if ( le128( bSig0, bSig1, aSig0, aSig1 ) ) { 5114 shift128Right( aSig0, aSig1, 1, &aSig0, &aSig1 ); 5115 ++zExp; 5116 } 5117 zSig0 = estimateDiv128To64( aSig0, aSig1, bSig0 ); 5118 mul128By64To192( bSig0, bSig1, zSig0, &term0, &term1, &term2 ); 5119 sub192( aSig0, aSig1, 0, term0, term1, term2, &rem0, &rem1, &rem2 ); 5120 while ( (sbits64) rem0 < 0 ) { 5121 --zSig0; 5122 add192( rem0, rem1, rem2, 0, bSig0, bSig1, &rem0, &rem1, &rem2 ); 5123 } 5124 zSig1 = estimateDiv128To64( rem1, rem2, bSig0 ); 5125 if ( ( zSig1 & 0x3FFF ) <= 4 ) { 5126 mul128By64To192( bSig0, bSig1, zSig1, &term1, &term2, &term3 ); 5127 sub192( rem1, rem2, 0, term1, term2, term3, &rem1, &rem2, &rem3 ); 5128 while ( (sbits64) rem1 < 0 ) { 5129 --zSig1; 5130 add192( rem1, rem2, rem3, 0, bSig0, bSig1, &rem1, &rem2, &rem3 ); 5131 } 5132 zSig1 |= ( ( rem1 | rem2 | rem3 ) != 0 ); 5133 } 5134 shift128ExtraRightJamming( zSig0, zSig1, 0, 15, &zSig0, &zSig1, &zSig2 ); 5135 return roundAndPackFloat128( zSign, zExp, zSig0, zSig1, zSig2 STATUS_VAR ); 5136 5137 } 5138 5139 /*---------------------------------------------------------------------------- 5140 | Returns the remainder of the quadruple-precision floating-point value `a' 5141 | with respect to the corresponding value `b'. The operation is performed 5142 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5143 *----------------------------------------------------------------------------*/ 5144 5145 float128 float128_rem( float128 a, float128 b STATUS_PARAM ) 5146 { 5147 flag aSign, bSign, zSign; 5148 int32 aExp, bExp, expDiff; 5149 bits64 aSig0, aSig1, bSig0, bSig1, q, term0, term1, term2; 5150 bits64 allZero, alternateASig0, alternateASig1, sigMean1; 5151 sbits64 sigMean0; 5152 float128 z; 5153 5154 aSig1 = extractFloat128Frac1( a ); 5155 aSig0 = extractFloat128Frac0( a ); 5156 aExp = extractFloat128Exp( a ); 5157 aSign = extractFloat128Sign( a ); 5158 bSig1 = extractFloat128Frac1( b ); 5159 bSig0 = extractFloat128Frac0( b ); 5160 bExp = extractFloat128Exp( b ); 5161 bSign = extractFloat128Sign( b ); 5162 if ( aExp == 0x7FFF ) { 5163 if ( ( aSig0 | aSig1 ) 5164 || ( ( bExp == 0x7FFF ) && ( bSig0 | bSig1 ) ) ) { 5165 return propagateFloat128NaN( a, b STATUS_VAR ); 5166 } 5167 goto invalid; 5168 } 5169 if ( bExp == 0x7FFF ) { 5170 if ( bSig0 | bSig1 ) return propagateFloat128NaN( a, b STATUS_VAR ); 5171 return a; 5172 } 5173 if ( bExp == 0 ) { 5174 if ( ( bSig0 | bSig1 ) == 0 ) { 5175 invalid: 5176 float_raise( float_flag_invalid STATUS_VAR); 5177 z.low = float128_default_nan_low; 5178 z.high = float128_default_nan_high; 5179 return z; 5180 } 5181 normalizeFloat128Subnormal( bSig0, bSig1, &bExp, &bSig0, &bSig1 ); 5182 } 5183 if ( aExp == 0 ) { 5184 if ( ( aSig0 | aSig1 ) == 0 ) return a; 5185 normalizeFloat128Subnormal( aSig0, aSig1, &aExp, &aSig0, &aSig1 ); 5186 } 5187 expDiff = aExp - bExp; 5188 if ( expDiff < -1 ) return a; 5189 shortShift128Left( 5190 aSig0 | LIT64( 0x0001000000000000 ), 5191 aSig1, 5192 15 - ( expDiff < 0 ), 5193 &aSig0, 5194 &aSig1 5195 ); 5196 shortShift128Left( 5197 bSig0 | LIT64( 0x0001000000000000 ), bSig1, 15, &bSig0, &bSig1 ); 5198 q = le128( bSig0, bSig1, aSig0, aSig1 ); 5199 if ( q ) sub128( aSig0, aSig1, bSig0, bSig1, &aSig0, &aSig1 ); 5200 expDiff -= 64; 5201 while ( 0 < expDiff ) { 5202 q = estimateDiv128To64( aSig0, aSig1, bSig0 ); 5203 q = ( 4 < q ) ? q - 4 : 0; 5204 mul128By64To192( bSig0, bSig1, q, &term0, &term1, &term2 ); 5205 shortShift192Left( term0, term1, term2, 61, &term1, &term2, &allZero ); 5206 shortShift128Left( aSig0, aSig1, 61, &aSig0, &allZero ); 5207 sub128( aSig0, 0, term1, term2, &aSig0, &aSig1 ); 5208 expDiff -= 61; 5209 } 5210 if ( -64 < expDiff ) { 5211 q = estimateDiv128To64( aSig0, aSig1, bSig0 ); 5212 q = ( 4 < q ) ? q - 4 : 0; 5213 q >>= - expDiff; 5214 shift128Right( bSig0, bSig1, 12, &bSig0, &bSig1 ); 5215 expDiff += 52; 5216 if ( expDiff < 0 ) { 5217 shift128Right( aSig0, aSig1, - expDiff, &aSig0, &aSig1 ); 5218 } 5219 else { 5220 shortShift128Left( aSig0, aSig1, expDiff, &aSig0, &aSig1 ); 5221 } 5222 mul128By64To192( bSig0, bSig1, q, &term0, &term1, &term2 ); 5223 sub128( aSig0, aSig1, term1, term2, &aSig0, &aSig1 ); 5224 } 5225 else { 5226 shift128Right( aSig0, aSig1, 12, &aSig0, &aSig1 ); 5227 shift128Right( bSig0, bSig1, 12, &bSig0, &bSig1 ); 5228 } 5229 do { 5230 alternateASig0 = aSig0; 5231 alternateASig1 = aSig1; 5232 ++q; 5233 sub128( aSig0, aSig1, bSig0, bSig1, &aSig0, &aSig1 ); 5234 } while ( 0 <= (sbits64) aSig0 ); 5235 add128( 5236 aSig0, aSig1, alternateASig0, alternateASig1, (bits64 *)&sigMean0, &sigMean1 ); 5237 if ( ( sigMean0 < 0 ) 5238 || ( ( ( sigMean0 | sigMean1 ) == 0 ) && ( q & 1 ) ) ) { 5239 aSig0 = alternateASig0; 5240 aSig1 = alternateASig1; 5241 } 5242 zSign = ( (sbits64) aSig0 < 0 ); 5243 if ( zSign ) sub128( 0, 0, aSig0, aSig1, &aSig0, &aSig1 ); 5244 return 5245 normalizeRoundAndPackFloat128( aSign ^ zSign, bExp - 4, aSig0, aSig1 STATUS_VAR ); 5246 5247 } 5248 5249 /*---------------------------------------------------------------------------- 5250 | Returns the square root of the quadruple-precision floating-point value `a'. 5251 | The operation is performed according to the IEC/IEEE Standard for Binary 5252 | Floating-Point Arithmetic. 5253 *----------------------------------------------------------------------------*/ 5254 5255 float128 float128_sqrt( float128 a STATUS_PARAM ) 5256 { 5257 flag aSign; 5258 int32 aExp, zExp; 5259 bits64 aSig0, aSig1, zSig0, zSig1, zSig2, doubleZSig0; 5260 bits64 rem0, rem1, rem2, rem3, term0, term1, term2, term3; 5261 float128 z; 5262 5263 aSig1 = extractFloat128Frac1( a ); 5264 aSig0 = extractFloat128Frac0( a ); 5265 aExp = extractFloat128Exp( a ); 5266 aSign = extractFloat128Sign( a ); 5267 if ( aExp == 0x7FFF ) { 5268 if ( aSig0 | aSig1 ) return propagateFloat128NaN( a, a STATUS_VAR ); 5269 if ( ! aSign ) return a; 5270 goto invalid; 5271 } 5272 if ( aSign ) { 5273 if ( ( aExp | aSig0 | aSig1 ) == 0 ) return a; 5274 invalid: 5275 float_raise( float_flag_invalid STATUS_VAR); 5276 z.low = float128_default_nan_low; 5277 z.high = float128_default_nan_high; 5278 return z; 5279 } 5280 if ( aExp == 0 ) { 5281 if ( ( aSig0 | aSig1 ) == 0 ) return packFloat128( 0, 0, 0, 0 ); 5282 normalizeFloat128Subnormal( aSig0, aSig1, &aExp, &aSig0, &aSig1 ); 5283 } 5284 zExp = ( ( aExp - 0x3FFF )>>1 ) + 0x3FFE; 5285 aSig0 |= LIT64( 0x0001000000000000 ); 5286 zSig0 = estimateSqrt32( aExp, aSig0>>17 ); 5287 shortShift128Left( aSig0, aSig1, 13 - ( aExp & 1 ), &aSig0, &aSig1 ); 5288 zSig0 = estimateDiv128To64( aSig0, aSig1, zSig0<<32 ) + ( zSig0<<30 ); 5289 doubleZSig0 = zSig0<<1; 5290 mul64To128( zSig0, zSig0, &term0, &term1 ); 5291 sub128( aSig0, aSig1, term0, term1, &rem0, &rem1 ); 5292 while ( (sbits64) rem0 < 0 ) { 5293 --zSig0; 5294 doubleZSig0 -= 2; 5295 add128( rem0, rem1, zSig0>>63, doubleZSig0 | 1, &rem0, &rem1 ); 5296 } 5297 zSig1 = estimateDiv128To64( rem1, 0, doubleZSig0 ); 5298 if ( ( zSig1 & 0x1FFF ) <= 5 ) { 5299 if ( zSig1 == 0 ) zSig1 = 1; 5300 mul64To128( doubleZSig0, zSig1, &term1, &term2 ); 5301 sub128( rem1, 0, term1, term2, &rem1, &rem2 ); 5302 mul64To128( zSig1, zSig1, &term2, &term3 ); 5303 sub192( rem1, rem2, 0, 0, term2, term3, &rem1, &rem2, &rem3 ); 5304 while ( (sbits64) rem1 < 0 ) { 5305 --zSig1; 5306 shortShift128Left( 0, zSig1, 1, &term2, &term3 ); 5307 term3 |= 1; 5308 term2 |= doubleZSig0; 5309 add192( rem1, rem2, rem3, 0, term2, term3, &rem1, &rem2, &rem3 ); 5310 } 5311 zSig1 |= ( ( rem1 | rem2 | rem3 ) != 0 ); 5312 } 5313 shift128ExtraRightJamming( zSig0, zSig1, 0, 14, &zSig0, &zSig1, &zSig2 ); 5314 return roundAndPackFloat128( 0, zExp, zSig0, zSig1, zSig2 STATUS_VAR ); 5315 5316 } 5317 5318 /*---------------------------------------------------------------------------- 5319 | Returns 1 if the quadruple-precision floating-point value `a' is equal to 5320 | the corresponding value `b', and 0 otherwise. The comparison is performed 5321 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5322 *----------------------------------------------------------------------------*/ 5323 5324 int float128_eq( float128 a, float128 b STATUS_PARAM ) 5325 { 5326 5327 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5328 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5329 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5330 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5331 ) { 5332 if ( float128_is_signaling_nan( a ) 5333 || float128_is_signaling_nan( b ) ) { 5334 float_raise( float_flag_invalid STATUS_VAR); 5335 } 5336 return 0; 5337 } 5338 return 5339 ( a.low == b.low ) 5340 && ( ( a.high == b.high ) 5341 || ( ( a.low == 0 ) 5342 && ( (bits64) ( ( a.high | b.high )<<1 ) == 0 ) ) 5343 ); 5344 5345 } 5346 5347 /*---------------------------------------------------------------------------- 5348 | Returns 1 if the quadruple-precision floating-point value `a' is less than 5349 | or equal to the corresponding value `b', and 0 otherwise. The comparison 5350 | is performed according to the IEC/IEEE Standard for Binary Floating-Point 5351 | Arithmetic. 5352 *----------------------------------------------------------------------------*/ 5353 5354 int float128_le( float128 a, float128 b STATUS_PARAM ) 5355 { 5356 flag aSign, bSign; 5357 5358 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5359 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5360 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5361 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5362 ) { 5363 float_raise( float_flag_invalid STATUS_VAR); 5364 return 0; 5365 } 5366 aSign = extractFloat128Sign( a ); 5367 bSign = extractFloat128Sign( b ); 5368 if ( aSign != bSign ) { 5369 return 5370 aSign 5371 || ( ( ( (bits64) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 5372 == 0 ); 5373 } 5374 return 5375 aSign ? le128( b.high, b.low, a.high, a.low ) 5376 : le128( a.high, a.low, b.high, b.low ); 5377 5378 } 5379 5380 /*---------------------------------------------------------------------------- 5381 | Returns 1 if the quadruple-precision floating-point value `a' is less than 5382 | the corresponding value `b', and 0 otherwise. The comparison is performed 5383 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5384 *----------------------------------------------------------------------------*/ 5385 5386 int float128_lt( float128 a, float128 b STATUS_PARAM ) 5387 { 5388 flag aSign, bSign; 5389 5390 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5391 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5392 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5393 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5394 ) { 5395 float_raise( float_flag_invalid STATUS_VAR); 5396 return 0; 5397 } 5398 aSign = extractFloat128Sign( a ); 5399 bSign = extractFloat128Sign( b ); 5400 if ( aSign != bSign ) { 5401 return 5402 aSign 5403 && ( ( ( (bits64) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 5404 != 0 ); 5405 } 5406 return 5407 aSign ? lt128( b.high, b.low, a.high, a.low ) 5408 : lt128( a.high, a.low, b.high, b.low ); 5409 5410 } 5411 5412 /*---------------------------------------------------------------------------- 5413 | Returns 1 if the quadruple-precision floating-point value `a' is equal to 5414 | the corresponding value `b', and 0 otherwise. The invalid exception is 5415 | raised if either operand is a NaN. Otherwise, the comparison is performed 5416 | according to the IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5417 *----------------------------------------------------------------------------*/ 5418 5419 int float128_eq_signaling( float128 a, float128 b STATUS_PARAM ) 5420 { 5421 5422 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5423 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5424 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5425 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5426 ) { 5427 float_raise( float_flag_invalid STATUS_VAR); 5428 return 0; 5429 } 5430 return 5431 ( a.low == b.low ) 5432 && ( ( a.high == b.high ) 5433 || ( ( a.low == 0 ) 5434 && ( (bits64) ( ( a.high | b.high )<<1 ) == 0 ) ) 5435 ); 5436 5437 } 5438 5439 /*---------------------------------------------------------------------------- 5440 | Returns 1 if the quadruple-precision floating-point value `a' is less than 5441 | or equal to the corresponding value `b', and 0 otherwise. Quiet NaNs do not 5442 | cause an exception. Otherwise, the comparison is performed according to the 5443 | IEC/IEEE Standard for Binary Floating-Point Arithmetic. 5444 *----------------------------------------------------------------------------*/ 5445 5446 int float128_le_quiet( float128 a, float128 b STATUS_PARAM ) 5447 { 5448 flag aSign, bSign; 5449 5450 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5451 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5452 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5453 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5454 ) { 5455 if ( float128_is_signaling_nan( a ) 5456 || float128_is_signaling_nan( b ) ) { 5457 float_raise( float_flag_invalid STATUS_VAR); 5458 } 5459 return 0; 5460 } 5461 aSign = extractFloat128Sign( a ); 5462 bSign = extractFloat128Sign( b ); 5463 if ( aSign != bSign ) { 5464 return 5465 aSign 5466 || ( ( ( (bits64) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 5467 == 0 ); 5468 } 5469 return 5470 aSign ? le128( b.high, b.low, a.high, a.low ) 5471 : le128( a.high, a.low, b.high, b.low ); 5472 5473 } 5474 5475 /*---------------------------------------------------------------------------- 5476 | Returns 1 if the quadruple-precision floating-point value `a' is less than 5477 | the corresponding value `b', and 0 otherwise. Quiet NaNs do not cause an 5478 | exception. Otherwise, the comparison is performed according to the IEC/IEEE 5479 | Standard for Binary Floating-Point Arithmetic. 5480 *----------------------------------------------------------------------------*/ 5481 5482 int float128_lt_quiet( float128 a, float128 b STATUS_PARAM ) 5483 { 5484 flag aSign, bSign; 5485 5486 if ( ( ( extractFloat128Exp( a ) == 0x7FFF ) 5487 && ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) 5488 || ( ( extractFloat128Exp( b ) == 0x7FFF ) 5489 && ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) ) 5490 ) { 5491 if ( float128_is_signaling_nan( a ) 5492 || float128_is_signaling_nan( b ) ) { 5493 float_raise( float_flag_invalid STATUS_VAR); 5494 } 5495 return 0; 5496 } 5497 aSign = extractFloat128Sign( a ); 5498 bSign = extractFloat128Sign( b ); 5499 if ( aSign != bSign ) { 5500 return 5501 aSign 5502 && ( ( ( (bits64) ( ( a.high | b.high )<<1 ) ) | a.low | b.low ) 5503 != 0 ); 5504 } 5505 return 5506 aSign ? lt128( b.high, b.low, a.high, a.low ) 5507 : lt128( a.high, a.low, b.high, b.low ); 5508 5509 } 5510 5511 #endif 5512 5513 /* misc functions */ 5514 float32 uint32_to_float32( unsigned int a STATUS_PARAM ) 5515 { 5516 return int64_to_float32(a STATUS_VAR); 5517 } 5518 5519 float64 uint32_to_float64( unsigned int a STATUS_PARAM ) 5520 { 5521 return int64_to_float64(a STATUS_VAR); 5522 } 5523 5524 unsigned int float32_to_uint32( float32 a STATUS_PARAM ) 5525 { 5526 int64_t v; 5527 unsigned int res; 5528 5529 v = float32_to_int64(a STATUS_VAR); 5530 if (v < 0) { 5531 res = 0; 5532 float_raise( float_flag_invalid STATUS_VAR); 5533 } else if (v > 0xffffffff) { 5534 res = 0xffffffff; 5535 float_raise( float_flag_invalid STATUS_VAR); 5536 } else { 5537 res = v; 5538 } 5539 return res; 5540 } 5541 5542 unsigned int float32_to_uint32_round_to_zero( float32 a STATUS_PARAM ) 5543 { 5544 int64_t v; 5545 unsigned int res; 5546 5547 v = float32_to_int64_round_to_zero(a STATUS_VAR); 5548 if (v < 0) { 5549 res = 0; 5550 float_raise( float_flag_invalid STATUS_VAR); 5551 } else if (v > 0xffffffff) { 5552 res = 0xffffffff; 5553 float_raise( float_flag_invalid STATUS_VAR); 5554 } else { 5555 res = v; 5556 } 5557 return res; 5558 } 5559 5560 unsigned int float64_to_uint32( float64 a STATUS_PARAM ) 5561 { 5562 int64_t v; 5563 unsigned int res; 5564 5565 v = float64_to_int64(a STATUS_VAR); 5566 if (v < 0) { 5567 res = 0; 5568 float_raise( float_flag_invalid STATUS_VAR); 5569 } else if (v > 0xffffffff) { 5570 res = 0xffffffff; 5571 float_raise( float_flag_invalid STATUS_VAR); 5572 } else { 5573 res = v; 5574 } 5575 return res; 5576 } 5577 5578 unsigned int float64_to_uint32_round_to_zero( float64 a STATUS_PARAM ) 5579 { 5580 int64_t v; 5581 unsigned int res; 5582 5583 v = float64_to_int64_round_to_zero(a STATUS_VAR); 5584 if (v < 0) { 5585 res = 0; 5586 float_raise( float_flag_invalid STATUS_VAR); 5587 } else if (v > 0xffffffff) { 5588 res = 0xffffffff; 5589 float_raise( float_flag_invalid STATUS_VAR); 5590 } else { 5591 res = v; 5592 } 5593 return res; 5594 } 5595 5596 /* FIXME: This looks broken. */ 5597 uint64_t float64_to_uint64 (float64 a STATUS_PARAM) 5598 { 5599 int64_t v; 5600 5601 v = float64_val(int64_to_float64(INT64_MIN STATUS_VAR)); 5602 v += float64_val(a); 5603 v = float64_to_int64(make_float64(v) STATUS_VAR); 5604 5605 return v - INT64_MIN; 5606 } 5607 5608 uint64_t float64_to_uint64_round_to_zero (float64 a STATUS_PARAM) 5609 { 5610 int64_t v; 5611 5612 v = float64_val(int64_to_float64(INT64_MIN STATUS_VAR)); 5613 v += float64_val(a); 5614 v = float64_to_int64_round_to_zero(make_float64(v) STATUS_VAR); 5615 5616 return v - INT64_MIN; 5617 } 5618 5619 #define COMPARE(s, nan_exp) \ 5620 INLINE int float ## s ## _compare_internal( float ## s a, float ## s b, \ 5621 int is_quiet STATUS_PARAM ) \ 5622 { \ 5623 flag aSign, bSign; \ 5624 bits ## s av, bv; \ 5625 \ 5626 if (( ( extractFloat ## s ## Exp( a ) == nan_exp ) && \ 5627 extractFloat ## s ## Frac( a ) ) || \ 5628 ( ( extractFloat ## s ## Exp( b ) == nan_exp ) && \ 5629 extractFloat ## s ## Frac( b ) )) { \ 5630 if (!is_quiet || \ 5631 float ## s ## _is_signaling_nan( a ) || \ 5632 float ## s ## _is_signaling_nan( b ) ) { \ 5633 float_raise( float_flag_invalid STATUS_VAR); \ 5634 } \ 5635 return float_relation_unordered; \ 5636 } \ 5637 aSign = extractFloat ## s ## Sign( a ); \ 5638 bSign = extractFloat ## s ## Sign( b ); \ 5639 av = float ## s ## _val(a); \ 5640 bv = float ## s ## _val(b); \ 5641 if ( aSign != bSign ) { \ 5642 if ( (bits ## s) ( ( av | bv )<<1 ) == 0 ) { \ 5643 /* zero case */ \ 5644 return float_relation_equal; \ 5645 } else { \ 5646 return 1 - (2 * aSign); \ 5647 } \ 5648 } else { \ 5649 if (av == bv) { \ 5650 return float_relation_equal; \ 5651 } else { \ 5652 return 1 - 2 * (aSign ^ ( av < bv )); \ 5653 } \ 5654 } \ 5655 } \ 5656 \ 5657 int float ## s ## _compare( float ## s a, float ## s b STATUS_PARAM ) \ 5658 { \ 5659 return float ## s ## _compare_internal(a, b, 0 STATUS_VAR); \ 5660 } \ 5661 \ 5662 int float ## s ## _compare_quiet( float ## s a, float ## s b STATUS_PARAM ) \ 5663 { \ 5664 return float ## s ## _compare_internal(a, b, 1 STATUS_VAR); \ 5665 } 5666 5667 COMPARE(32, 0xff) 5668 COMPARE(64, 0x7ff) 5669 5670 INLINE int float128_compare_internal( float128 a, float128 b, 5671 int is_quiet STATUS_PARAM ) 5672 { 5673 flag aSign, bSign; 5674 5675 if (( ( extractFloat128Exp( a ) == 0x7fff ) && 5676 ( extractFloat128Frac0( a ) | extractFloat128Frac1( a ) ) ) || 5677 ( ( extractFloat128Exp( b ) == 0x7fff ) && 5678 ( extractFloat128Frac0( b ) | extractFloat128Frac1( b ) ) )) { 5679 if (!is_quiet || 5680 float128_is_signaling_nan( a ) || 5681 float128_is_signaling_nan( b ) ) { 5682 float_raise( float_flag_invalid STATUS_VAR); 5683 } 5684 return float_relation_unordered; 5685 } 5686 aSign = extractFloat128Sign( a ); 5687 bSign = extractFloat128Sign( b ); 5688 if ( aSign != bSign ) { 5689 if ( ( ( ( a.high | b.high )<<1 ) | a.low | b.low ) == 0 ) { 5690 /* zero case */ 5691 return float_relation_equal; 5692 } else { 5693 return 1 - (2 * aSign); 5694 } 5695 } else { 5696 if (a.low == b.low && a.high == b.high) { 5697 return float_relation_equal; 5698 } else { 5699 return 1 - 2 * (aSign ^ ( lt128( a.high, a.low, b.high, b.low ) )); 5700 } 5701 } 5702 } 5703 5704 int float128_compare( float128 a, float128 b STATUS_PARAM ) 5705 { 5706 return float128_compare_internal(a, b, 0 STATUS_VAR); 5707 } 5708 5709 int float128_compare_quiet( float128 a, float128 b STATUS_PARAM ) 5710 { 5711 return float128_compare_internal(a, b, 1 STATUS_VAR); 5712 } 5713 5714 /* Multiply A by 2 raised to the power N. */ 5715 float32 float32_scalbn( float32 a, int n STATUS_PARAM ) 5716 { 5717 flag aSign; 5718 int16 aExp; 5719 bits32 aSig; 5720 5721 aSig = extractFloat32Frac( a ); 5722 aExp = extractFloat32Exp( a ); 5723 aSign = extractFloat32Sign( a ); 5724 5725 if ( aExp == 0xFF ) { 5726 return a; 5727 } 5728 if ( aExp != 0 ) 5729 aSig |= 0x00800000; 5730 else if ( aSig == 0 ) 5731 return a; 5732 5733 aExp += n - 1; 5734 aSig <<= 7; 5735 return normalizeRoundAndPackFloat32( aSign, aExp, aSig STATUS_VAR ); 5736 } 5737 5738 float64 float64_scalbn( float64 a, int n STATUS_PARAM ) 5739 { 5740 flag aSign; 5741 int16 aExp; 5742 bits64 aSig; 5743 5744 aSig = extractFloat64Frac( a ); 5745 aExp = extractFloat64Exp( a ); 5746 aSign = extractFloat64Sign( a ); 5747 5748 if ( aExp == 0x7FF ) { 5749 return a; 5750 } 5751 if ( aExp != 0 ) 5752 aSig |= LIT64( 0x0010000000000000 ); 5753 else if ( aSig == 0 ) 5754 return a; 5755 5756 aExp += n - 1; 5757 aSig <<= 10; 5758 return normalizeRoundAndPackFloat64( aSign, aExp, aSig STATUS_VAR ); 5759 } 5760 5761 #ifdef FLOATX80 5762 floatx80 floatx80_scalbn( floatx80 a, int n STATUS_PARAM ) 5763 { 5764 flag aSign; 5765 int16 aExp; 5766 bits64 aSig; 5767 5768 aSig = extractFloatx80Frac( a ); 5769 aExp = extractFloatx80Exp( a ); 5770 aSign = extractFloatx80Sign( a ); 5771 5772 if ( aExp == 0x7FF ) { 5773 return a; 5774 } 5775 if (aExp == 0 && aSig == 0) 5776 return a; 5777 5778 aExp += n; 5779 return normalizeRoundAndPackFloatx80( STATUS(floatx80_rounding_precision), 5780 aSign, aExp, aSig, 0 STATUS_VAR ); 5781 } 5782 #endif 5783 5784 #ifdef FLOAT128 5785 float128 float128_scalbn( float128 a, int n STATUS_PARAM ) 5786 { 5787 flag aSign; 5788 int32 aExp; 5789 bits64 aSig0, aSig1; 5790 5791 aSig1 = extractFloat128Frac1( a ); 5792 aSig0 = extractFloat128Frac0( a ); 5793 aExp = extractFloat128Exp( a ); 5794 aSign = extractFloat128Sign( a ); 5795 if ( aExp == 0x7FFF ) { 5796 return a; 5797 } 5798 if ( aExp != 0 ) 5799 aSig0 |= LIT64( 0x0001000000000000 ); 5800 else if ( aSig0 == 0 && aSig1 == 0 ) 5801 return a; 5802 5803 aExp += n - 1; 5804 return normalizeRoundAndPackFloat128( aSign, aExp, aSig0, aSig1 5805 STATUS_VAR ); 5806 5807 } 5808 #endif 5809