APFloat.cpp revision 386f3e9d50507ecbfb01436dee85cb41ac2df530
1//===-- APFloat.cpp - Implement APFloat class -----------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements a class to represent arbitrary precision floating 11// point values and provide a variety of arithmetic operations on them. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/ADT/APFloat.h" 16#include "llvm/ADT/FoldingSet.h" 17#include <cassert> 18#include <cstring> 19#include "llvm/Support/MathExtras.h" 20 21using namespace llvm; 22 23#define convolve(lhs, rhs) ((lhs) * 4 + (rhs)) 24 25/* Assumed in hexadecimal significand parsing, and conversion to 26 hexadecimal strings. */ 27COMPILE_TIME_ASSERT(integerPartWidth % 4 == 0); 28 29namespace llvm { 30 31 /* Represents floating point arithmetic semantics. */ 32 struct fltSemantics { 33 /* The largest E such that 2^E is representable; this matches the 34 definition of IEEE 754. */ 35 exponent_t maxExponent; 36 37 /* The smallest E such that 2^E is a normalized number; this 38 matches the definition of IEEE 754. */ 39 exponent_t minExponent; 40 41 /* Number of bits in the significand. This includes the integer 42 bit. */ 43 unsigned int precision; 44 45 /* True if arithmetic is supported. */ 46 unsigned int arithmeticOK; 47 }; 48 49 const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, true }; 50 const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, true }; 51 const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, true }; 52 const fltSemantics APFloat::x87DoubleExtended = { 16383, -16382, 64, true }; 53 const fltSemantics APFloat::Bogus = { 0, 0, 0, true }; 54 55 // The PowerPC format consists of two doubles. It does not map cleanly 56 // onto the usual format above. For now only storage of constants of 57 // this type is supported, no arithmetic. 58 const fltSemantics APFloat::PPCDoubleDouble = { 1023, -1022, 106, false }; 59 60 /* A tight upper bound on number of parts required to hold the value 61 pow(5, power) is 62 63 power * 815 / (351 * integerPartWidth) + 1 64 65 However, whilst the result may require only this many parts, 66 because we are multiplying two values to get it, the 67 multiplication may require an extra part with the excess part 68 being zero (consider the trivial case of 1 * 1, tcFullMultiply 69 requires two parts to hold the single-part result). So we add an 70 extra one to guarantee enough space whilst multiplying. */ 71 const unsigned int maxExponent = 16383; 72 const unsigned int maxPrecision = 113; 73 const unsigned int maxPowerOfFiveExponent = maxExponent + maxPrecision - 1; 74 const unsigned int maxPowerOfFiveParts = 2 + ((maxPowerOfFiveExponent * 815) 75 / (351 * integerPartWidth)); 76} 77 78/* Put a bunch of private, handy routines in an anonymous namespace. */ 79namespace { 80 81 static inline unsigned int 82 partCountForBits(unsigned int bits) 83 { 84 return ((bits) + integerPartWidth - 1) / integerPartWidth; 85 } 86 87 /* Returns 0U-9U. Return values >= 10U are not digits. */ 88 static inline unsigned int 89 decDigitValue(unsigned int c) 90 { 91 return c - '0'; 92 } 93 94 static unsigned int 95 hexDigitValue(unsigned int c) 96 { 97 unsigned int r; 98 99 r = c - '0'; 100 if(r <= 9) 101 return r; 102 103 r = c - 'A'; 104 if(r <= 5) 105 return r + 10; 106 107 r = c - 'a'; 108 if(r <= 5) 109 return r + 10; 110 111 return -1U; 112 } 113 114 static inline void 115 assertArithmeticOK(const llvm::fltSemantics &semantics) { 116 assert(semantics.arithmeticOK 117 && "Compile-time arithmetic does not support these semantics"); 118 } 119 120 /* Return the value of a decimal exponent of the form 121 [+-]ddddddd. 122 123 If the exponent overflows, returns a large exponent with the 124 appropriate sign. */ 125 static int 126 readExponent(const char *p) 127 { 128 bool isNegative; 129 unsigned int absExponent; 130 const unsigned int overlargeExponent = 24000; /* FIXME. */ 131 132 isNegative = (*p == '-'); 133 if (*p == '-' || *p == '+') 134 p++; 135 136 absExponent = decDigitValue(*p++); 137 assert (absExponent < 10U); 138 139 for (;;) { 140 unsigned int value; 141 142 value = decDigitValue(*p); 143 if (value >= 10U) 144 break; 145 146 p++; 147 value += absExponent * 10; 148 if (absExponent >= overlargeExponent) { 149 absExponent = overlargeExponent; 150 break; 151 } 152 absExponent = value; 153 } 154 155 if (isNegative) 156 return -(int) absExponent; 157 else 158 return (int) absExponent; 159 } 160 161 /* This is ugly and needs cleaning up, but I don't immediately see 162 how whilst remaining safe. */ 163 static int 164 totalExponent(const char *p, int exponentAdjustment) 165 { 166 int unsignedExponent; 167 bool negative, overflow; 168 int exponent; 169 170 /* Move past the exponent letter and sign to the digits. */ 171 p++; 172 negative = *p == '-'; 173 if(*p == '-' || *p == '+') 174 p++; 175 176 unsignedExponent = 0; 177 overflow = false; 178 for(;;) { 179 unsigned int value; 180 181 value = decDigitValue(*p); 182 if(value >= 10U) 183 break; 184 185 p++; 186 unsignedExponent = unsignedExponent * 10 + value; 187 if(unsignedExponent > 65535) 188 overflow = true; 189 } 190 191 if(exponentAdjustment > 65535 || exponentAdjustment < -65536) 192 overflow = true; 193 194 if(!overflow) { 195 exponent = unsignedExponent; 196 if(negative) 197 exponent = -exponent; 198 exponent += exponentAdjustment; 199 if(exponent > 65535 || exponent < -65536) 200 overflow = true; 201 } 202 203 if(overflow) 204 exponent = negative ? -65536: 65535; 205 206 return exponent; 207 } 208 209 static const char * 210 skipLeadingZeroesAndAnyDot(const char *p, const char **dot) 211 { 212 *dot = 0; 213 while(*p == '0') 214 p++; 215 216 if(*p == '.') { 217 *dot = p++; 218 while(*p == '0') 219 p++; 220 } 221 222 return p; 223 } 224 225 /* Given a normal decimal floating point number of the form 226 227 dddd.dddd[eE][+-]ddd 228 229 where the decimal point and exponent are optional, fill out the 230 structure D. Exponent is appropriate if the significand is 231 treated as an integer, and normalizedExponent if the significand 232 is taken to have the decimal point after a single leading 233 non-zero digit. 234 235 If the value is zero, V->firstSigDigit points to a non-digit, and 236 the return exponent is zero. 237 */ 238 struct decimalInfo { 239 const char *firstSigDigit; 240 const char *lastSigDigit; 241 int exponent; 242 int normalizedExponent; 243 }; 244 245 static void 246 interpretDecimal(const char *p, decimalInfo *D) 247 { 248 const char *dot; 249 250 p = skipLeadingZeroesAndAnyDot (p, &dot); 251 252 D->firstSigDigit = p; 253 D->exponent = 0; 254 D->normalizedExponent = 0; 255 256 for (;;) { 257 if (*p == '.') { 258 assert(dot == 0); 259 dot = p++; 260 } 261 if (decDigitValue(*p) >= 10U) 262 break; 263 p++; 264 } 265 266 /* If number is all zerooes accept any exponent. */ 267 if (p != D->firstSigDigit) { 268 if (*p == 'e' || *p == 'E') 269 D->exponent = readExponent(p + 1); 270 271 /* Implied decimal point? */ 272 if (!dot) 273 dot = p; 274 275 /* Drop insignificant trailing zeroes. */ 276 do 277 do 278 p--; 279 while (*p == '0'); 280 while (*p == '.'); 281 282 /* Adjust the exponents for any decimal point. */ 283 D->exponent += static_cast<exponent_t>((dot - p) - (dot > p)); 284 D->normalizedExponent = (D->exponent + 285 static_cast<exponent_t>((p - D->firstSigDigit) 286 - (dot > D->firstSigDigit && dot < p))); 287 } 288 289 D->lastSigDigit = p; 290 } 291 292 /* Return the trailing fraction of a hexadecimal number. 293 DIGITVALUE is the first hex digit of the fraction, P points to 294 the next digit. */ 295 static lostFraction 296 trailingHexadecimalFraction(const char *p, unsigned int digitValue) 297 { 298 unsigned int hexDigit; 299 300 /* If the first trailing digit isn't 0 or 8 we can work out the 301 fraction immediately. */ 302 if(digitValue > 8) 303 return lfMoreThanHalf; 304 else if(digitValue < 8 && digitValue > 0) 305 return lfLessThanHalf; 306 307 /* Otherwise we need to find the first non-zero digit. */ 308 while(*p == '0') 309 p++; 310 311 hexDigit = hexDigitValue(*p); 312 313 /* If we ran off the end it is exactly zero or one-half, otherwise 314 a little more. */ 315 if(hexDigit == -1U) 316 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf; 317 else 318 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf; 319 } 320 321 /* Return the fraction lost were a bignum truncated losing the least 322 significant BITS bits. */ 323 static lostFraction 324 lostFractionThroughTruncation(const integerPart *parts, 325 unsigned int partCount, 326 unsigned int bits) 327 { 328 unsigned int lsb; 329 330 lsb = APInt::tcLSB(parts, partCount); 331 332 /* Note this is guaranteed true if bits == 0, or LSB == -1U. */ 333 if(bits <= lsb) 334 return lfExactlyZero; 335 if(bits == lsb + 1) 336 return lfExactlyHalf; 337 if(bits <= partCount * integerPartWidth 338 && APInt::tcExtractBit(parts, bits - 1)) 339 return lfMoreThanHalf; 340 341 return lfLessThanHalf; 342 } 343 344 /* Shift DST right BITS bits noting lost fraction. */ 345 static lostFraction 346 shiftRight(integerPart *dst, unsigned int parts, unsigned int bits) 347 { 348 lostFraction lost_fraction; 349 350 lost_fraction = lostFractionThroughTruncation(dst, parts, bits); 351 352 APInt::tcShiftRight(dst, parts, bits); 353 354 return lost_fraction; 355 } 356 357 /* Combine the effect of two lost fractions. */ 358 static lostFraction 359 combineLostFractions(lostFraction moreSignificant, 360 lostFraction lessSignificant) 361 { 362 if(lessSignificant != lfExactlyZero) { 363 if(moreSignificant == lfExactlyZero) 364 moreSignificant = lfLessThanHalf; 365 else if(moreSignificant == lfExactlyHalf) 366 moreSignificant = lfMoreThanHalf; 367 } 368 369 return moreSignificant; 370 } 371 372 /* The error from the true value, in half-ulps, on multiplying two 373 floating point numbers, which differ from the value they 374 approximate by at most HUE1 and HUE2 half-ulps, is strictly less 375 than the returned value. 376 377 See "How to Read Floating Point Numbers Accurately" by William D 378 Clinger. */ 379 static unsigned int 380 HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2) 381 { 382 assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8)); 383 384 if (HUerr1 + HUerr2 == 0) 385 return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */ 386 else 387 return inexactMultiply + 2 * (HUerr1 + HUerr2); 388 } 389 390 /* The number of ulps from the boundary (zero, or half if ISNEAREST) 391 when the least significant BITS are truncated. BITS cannot be 392 zero. */ 393 static integerPart 394 ulpsFromBoundary(const integerPart *parts, unsigned int bits, bool isNearest) 395 { 396 unsigned int count, partBits; 397 integerPart part, boundary; 398 399 assert (bits != 0); 400 401 bits--; 402 count = bits / integerPartWidth; 403 partBits = bits % integerPartWidth + 1; 404 405 part = parts[count] & (~(integerPart) 0 >> (integerPartWidth - partBits)); 406 407 if (isNearest) 408 boundary = (integerPart) 1 << (partBits - 1); 409 else 410 boundary = 0; 411 412 if (count == 0) { 413 if (part - boundary <= boundary - part) 414 return part - boundary; 415 else 416 return boundary - part; 417 } 418 419 if (part == boundary) { 420 while (--count) 421 if (parts[count]) 422 return ~(integerPart) 0; /* A lot. */ 423 424 return parts[0]; 425 } else if (part == boundary - 1) { 426 while (--count) 427 if (~parts[count]) 428 return ~(integerPart) 0; /* A lot. */ 429 430 return -parts[0]; 431 } 432 433 return ~(integerPart) 0; /* A lot. */ 434 } 435 436 /* Place pow(5, power) in DST, and return the number of parts used. 437 DST must be at least one part larger than size of the answer. */ 438 static unsigned int 439 powerOf5(integerPart *dst, unsigned int power) 440 { 441 static const integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125, 442 15625, 78125 }; 443 static integerPart pow5s[maxPowerOfFiveParts * 2 + 5] = { 78125 * 5 }; 444 static unsigned int partsCount[16] = { 1 }; 445 446 integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5; 447 unsigned int result; 448 449 assert(power <= maxExponent); 450 451 p1 = dst; 452 p2 = scratch; 453 454 *p1 = firstEightPowers[power & 7]; 455 power >>= 3; 456 457 result = 1; 458 pow5 = pow5s; 459 460 for (unsigned int n = 0; power; power >>= 1, n++) { 461 unsigned int pc; 462 463 pc = partsCount[n]; 464 465 /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */ 466 if (pc == 0) { 467 pc = partsCount[n - 1]; 468 APInt::tcFullMultiply(pow5, pow5 - pc, pow5 - pc, pc, pc); 469 pc *= 2; 470 if (pow5[pc - 1] == 0) 471 pc--; 472 partsCount[n] = pc; 473 } 474 475 if (power & 1) { 476 integerPart *tmp; 477 478 APInt::tcFullMultiply(p2, p1, pow5, result, pc); 479 result += pc; 480 if (p2[result - 1] == 0) 481 result--; 482 483 /* Now result is in p1 with partsCount parts and p2 is scratch 484 space. */ 485 tmp = p1, p1 = p2, p2 = tmp; 486 } 487 488 pow5 += pc; 489 } 490 491 if (p1 != dst) 492 APInt::tcAssign(dst, p1, result); 493 494 return result; 495 } 496 497 /* Zero at the end to avoid modular arithmetic when adding one; used 498 when rounding up during hexadecimal output. */ 499 static const char hexDigitsLower[] = "0123456789abcdef0"; 500 static const char hexDigitsUpper[] = "0123456789ABCDEF0"; 501 static const char infinityL[] = "infinity"; 502 static const char infinityU[] = "INFINITY"; 503 static const char NaNL[] = "nan"; 504 static const char NaNU[] = "NAN"; 505 506 /* Write out an integerPart in hexadecimal, starting with the most 507 significant nibble. Write out exactly COUNT hexdigits, return 508 COUNT. */ 509 static unsigned int 510 partAsHex (char *dst, integerPart part, unsigned int count, 511 const char *hexDigitChars) 512 { 513 unsigned int result = count; 514 515 assert (count != 0 && count <= integerPartWidth / 4); 516 517 part >>= (integerPartWidth - 4 * count); 518 while (count--) { 519 dst[count] = hexDigitChars[part & 0xf]; 520 part >>= 4; 521 } 522 523 return result; 524 } 525 526 /* Write out an unsigned decimal integer. */ 527 static char * 528 writeUnsignedDecimal (char *dst, unsigned int n) 529 { 530 char buff[40], *p; 531 532 p = buff; 533 do 534 *p++ = '0' + n % 10; 535 while (n /= 10); 536 537 do 538 *dst++ = *--p; 539 while (p != buff); 540 541 return dst; 542 } 543 544 /* Write out a signed decimal integer. */ 545 static char * 546 writeSignedDecimal (char *dst, int value) 547 { 548 if (value < 0) { 549 *dst++ = '-'; 550 dst = writeUnsignedDecimal(dst, -(unsigned) value); 551 } else 552 dst = writeUnsignedDecimal(dst, value); 553 554 return dst; 555 } 556} 557 558/* Constructors. */ 559void 560APFloat::initialize(const fltSemantics *ourSemantics) 561{ 562 unsigned int count; 563 564 semantics = ourSemantics; 565 count = partCount(); 566 if(count > 1) 567 significand.parts = new integerPart[count]; 568} 569 570void 571APFloat::freeSignificand() 572{ 573 if(partCount() > 1) 574 delete [] significand.parts; 575} 576 577void 578APFloat::assign(const APFloat &rhs) 579{ 580 assert(semantics == rhs.semantics); 581 582 sign = rhs.sign; 583 category = rhs.category; 584 exponent = rhs.exponent; 585 sign2 = rhs.sign2; 586 exponent2 = rhs.exponent2; 587 if(category == fcNormal || category == fcNaN) 588 copySignificand(rhs); 589} 590 591void 592APFloat::copySignificand(const APFloat &rhs) 593{ 594 assert(category == fcNormal || category == fcNaN); 595 assert(rhs.partCount() >= partCount()); 596 597 APInt::tcAssign(significandParts(), rhs.significandParts(), 598 partCount()); 599} 600 601/* Make this number a NaN, with an arbitrary but deterministic value 602 for the significand. */ 603void 604APFloat::makeNaN(void) 605{ 606 category = fcNaN; 607 APInt::tcSet(significandParts(), ~0U, partCount()); 608} 609 610APFloat & 611APFloat::operator=(const APFloat &rhs) 612{ 613 if(this != &rhs) { 614 if(semantics != rhs.semantics) { 615 freeSignificand(); 616 initialize(rhs.semantics); 617 } 618 assign(rhs); 619 } 620 621 return *this; 622} 623 624bool 625APFloat::bitwiseIsEqual(const APFloat &rhs) const { 626 if (this == &rhs) 627 return true; 628 if (semantics != rhs.semantics || 629 category != rhs.category || 630 sign != rhs.sign) 631 return false; 632 if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble && 633 sign2 != rhs.sign2) 634 return false; 635 if (category==fcZero || category==fcInfinity) 636 return true; 637 else if (category==fcNormal && exponent!=rhs.exponent) 638 return false; 639 else if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble && 640 exponent2!=rhs.exponent2) 641 return false; 642 else { 643 int i= partCount(); 644 const integerPart* p=significandParts(); 645 const integerPart* q=rhs.significandParts(); 646 for (; i>0; i--, p++, q++) { 647 if (*p != *q) 648 return false; 649 } 650 return true; 651 } 652} 653 654APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value) 655{ 656 assertArithmeticOK(ourSemantics); 657 initialize(&ourSemantics); 658 sign = 0; 659 zeroSignificand(); 660 exponent = ourSemantics.precision - 1; 661 significandParts()[0] = value; 662 normalize(rmNearestTiesToEven, lfExactlyZero); 663} 664 665APFloat::APFloat(const fltSemantics &ourSemantics, 666 fltCategory ourCategory, bool negative) 667{ 668 assertArithmeticOK(ourSemantics); 669 initialize(&ourSemantics); 670 category = ourCategory; 671 sign = negative; 672 if(category == fcNormal) 673 category = fcZero; 674 else if (ourCategory == fcNaN) 675 makeNaN(); 676} 677 678APFloat::APFloat(const fltSemantics &ourSemantics, const char *text) 679{ 680 assertArithmeticOK(ourSemantics); 681 initialize(&ourSemantics); 682 convertFromString(text, rmNearestTiesToEven); 683} 684 685APFloat::APFloat(const APFloat &rhs) 686{ 687 initialize(rhs.semantics); 688 assign(rhs); 689} 690 691APFloat::~APFloat() 692{ 693 freeSignificand(); 694} 695 696// Profile - This method 'profiles' an APFloat for use with FoldingSet. 697void APFloat::Profile(FoldingSetNodeID& ID) const { 698 ID.Add(convertToAPInt()); 699} 700 701unsigned int 702APFloat::partCount() const 703{ 704 return partCountForBits(semantics->precision + 1); 705} 706 707unsigned int 708APFloat::semanticsPrecision(const fltSemantics &semantics) 709{ 710 return semantics.precision; 711} 712 713const integerPart * 714APFloat::significandParts() const 715{ 716 return const_cast<APFloat *>(this)->significandParts(); 717} 718 719integerPart * 720APFloat::significandParts() 721{ 722 assert(category == fcNormal || category == fcNaN); 723 724 if(partCount() > 1) 725 return significand.parts; 726 else 727 return &significand.part; 728} 729 730void 731APFloat::zeroSignificand() 732{ 733 category = fcNormal; 734 APInt::tcSet(significandParts(), 0, partCount()); 735} 736 737/* Increment an fcNormal floating point number's significand. */ 738void 739APFloat::incrementSignificand() 740{ 741 integerPart carry; 742 743 carry = APInt::tcIncrement(significandParts(), partCount()); 744 745 /* Our callers should never cause us to overflow. */ 746 assert(carry == 0); 747} 748 749/* Add the significand of the RHS. Returns the carry flag. */ 750integerPart 751APFloat::addSignificand(const APFloat &rhs) 752{ 753 integerPart *parts; 754 755 parts = significandParts(); 756 757 assert(semantics == rhs.semantics); 758 assert(exponent == rhs.exponent); 759 760 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount()); 761} 762 763/* Subtract the significand of the RHS with a borrow flag. Returns 764 the borrow flag. */ 765integerPart 766APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow) 767{ 768 integerPart *parts; 769 770 parts = significandParts(); 771 772 assert(semantics == rhs.semantics); 773 assert(exponent == rhs.exponent); 774 775 return APInt::tcSubtract(parts, rhs.significandParts(), borrow, 776 partCount()); 777} 778 779/* Multiply the significand of the RHS. If ADDEND is non-NULL, add it 780 on to the full-precision result of the multiplication. Returns the 781 lost fraction. */ 782lostFraction 783APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend) 784{ 785 unsigned int omsb; // One, not zero, based MSB. 786 unsigned int partsCount, newPartsCount, precision; 787 integerPart *lhsSignificand; 788 integerPart scratch[4]; 789 integerPart *fullSignificand; 790 lostFraction lost_fraction; 791 792 assert(semantics == rhs.semantics); 793 794 precision = semantics->precision; 795 newPartsCount = partCountForBits(precision * 2); 796 797 if(newPartsCount > 4) 798 fullSignificand = new integerPart[newPartsCount]; 799 else 800 fullSignificand = scratch; 801 802 lhsSignificand = significandParts(); 803 partsCount = partCount(); 804 805 APInt::tcFullMultiply(fullSignificand, lhsSignificand, 806 rhs.significandParts(), partsCount, partsCount); 807 808 lost_fraction = lfExactlyZero; 809 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1; 810 exponent += rhs.exponent; 811 812 if(addend) { 813 Significand savedSignificand = significand; 814 const fltSemantics *savedSemantics = semantics; 815 fltSemantics extendedSemantics; 816 opStatus status; 817 unsigned int extendedPrecision; 818 819 /* Normalize our MSB. */ 820 extendedPrecision = precision + precision - 1; 821 if(omsb != extendedPrecision) 822 { 823 APInt::tcShiftLeft(fullSignificand, newPartsCount, 824 extendedPrecision - omsb); 825 exponent -= extendedPrecision - omsb; 826 } 827 828 /* Create new semantics. */ 829 extendedSemantics = *semantics; 830 extendedSemantics.precision = extendedPrecision; 831 832 if(newPartsCount == 1) 833 significand.part = fullSignificand[0]; 834 else 835 significand.parts = fullSignificand; 836 semantics = &extendedSemantics; 837 838 APFloat extendedAddend(*addend); 839 status = extendedAddend.convert(extendedSemantics, rmTowardZero); 840 assert(status == opOK); 841 lost_fraction = addOrSubtractSignificand(extendedAddend, false); 842 843 /* Restore our state. */ 844 if(newPartsCount == 1) 845 fullSignificand[0] = significand.part; 846 significand = savedSignificand; 847 semantics = savedSemantics; 848 849 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1; 850 } 851 852 exponent -= (precision - 1); 853 854 if(omsb > precision) { 855 unsigned int bits, significantParts; 856 lostFraction lf; 857 858 bits = omsb - precision; 859 significantParts = partCountForBits(omsb); 860 lf = shiftRight(fullSignificand, significantParts, bits); 861 lost_fraction = combineLostFractions(lf, lost_fraction); 862 exponent += bits; 863 } 864 865 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount); 866 867 if(newPartsCount > 4) 868 delete [] fullSignificand; 869 870 return lost_fraction; 871} 872 873/* Multiply the significands of LHS and RHS to DST. */ 874lostFraction 875APFloat::divideSignificand(const APFloat &rhs) 876{ 877 unsigned int bit, i, partsCount; 878 const integerPart *rhsSignificand; 879 integerPart *lhsSignificand, *dividend, *divisor; 880 integerPart scratch[4]; 881 lostFraction lost_fraction; 882 883 assert(semantics == rhs.semantics); 884 885 lhsSignificand = significandParts(); 886 rhsSignificand = rhs.significandParts(); 887 partsCount = partCount(); 888 889 if(partsCount > 2) 890 dividend = new integerPart[partsCount * 2]; 891 else 892 dividend = scratch; 893 894 divisor = dividend + partsCount; 895 896 /* Copy the dividend and divisor as they will be modified in-place. */ 897 for(i = 0; i < partsCount; i++) { 898 dividend[i] = lhsSignificand[i]; 899 divisor[i] = rhsSignificand[i]; 900 lhsSignificand[i] = 0; 901 } 902 903 exponent -= rhs.exponent; 904 905 unsigned int precision = semantics->precision; 906 907 /* Normalize the divisor. */ 908 bit = precision - APInt::tcMSB(divisor, partsCount) - 1; 909 if(bit) { 910 exponent += bit; 911 APInt::tcShiftLeft(divisor, partsCount, bit); 912 } 913 914 /* Normalize the dividend. */ 915 bit = precision - APInt::tcMSB(dividend, partsCount) - 1; 916 if(bit) { 917 exponent -= bit; 918 APInt::tcShiftLeft(dividend, partsCount, bit); 919 } 920 921 /* Ensure the dividend >= divisor initially for the loop below. 922 Incidentally, this means that the division loop below is 923 guaranteed to set the integer bit to one. */ 924 if(APInt::tcCompare(dividend, divisor, partsCount) < 0) { 925 exponent--; 926 APInt::tcShiftLeft(dividend, partsCount, 1); 927 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0); 928 } 929 930 /* Long division. */ 931 for(bit = precision; bit; bit -= 1) { 932 if(APInt::tcCompare(dividend, divisor, partsCount) >= 0) { 933 APInt::tcSubtract(dividend, divisor, 0, partsCount); 934 APInt::tcSetBit(lhsSignificand, bit - 1); 935 } 936 937 APInt::tcShiftLeft(dividend, partsCount, 1); 938 } 939 940 /* Figure out the lost fraction. */ 941 int cmp = APInt::tcCompare(dividend, divisor, partsCount); 942 943 if(cmp > 0) 944 lost_fraction = lfMoreThanHalf; 945 else if(cmp == 0) 946 lost_fraction = lfExactlyHalf; 947 else if(APInt::tcIsZero(dividend, partsCount)) 948 lost_fraction = lfExactlyZero; 949 else 950 lost_fraction = lfLessThanHalf; 951 952 if(partsCount > 2) 953 delete [] dividend; 954 955 return lost_fraction; 956} 957 958unsigned int 959APFloat::significandMSB() const 960{ 961 return APInt::tcMSB(significandParts(), partCount()); 962} 963 964unsigned int 965APFloat::significandLSB() const 966{ 967 return APInt::tcLSB(significandParts(), partCount()); 968} 969 970/* Note that a zero result is NOT normalized to fcZero. */ 971lostFraction 972APFloat::shiftSignificandRight(unsigned int bits) 973{ 974 /* Our exponent should not overflow. */ 975 assert((exponent_t) (exponent + bits) >= exponent); 976 977 exponent += bits; 978 979 return shiftRight(significandParts(), partCount(), bits); 980} 981 982/* Shift the significand left BITS bits, subtract BITS from its exponent. */ 983void 984APFloat::shiftSignificandLeft(unsigned int bits) 985{ 986 assert(bits < semantics->precision); 987 988 if(bits) { 989 unsigned int partsCount = partCount(); 990 991 APInt::tcShiftLeft(significandParts(), partsCount, bits); 992 exponent -= bits; 993 994 assert(!APInt::tcIsZero(significandParts(), partsCount)); 995 } 996} 997 998APFloat::cmpResult 999APFloat::compareAbsoluteValue(const APFloat &rhs) const 1000{ 1001 int compare; 1002 1003 assert(semantics == rhs.semantics); 1004 assert(category == fcNormal); 1005 assert(rhs.category == fcNormal); 1006 1007 compare = exponent - rhs.exponent; 1008 1009 /* If exponents are equal, do an unsigned bignum comparison of the 1010 significands. */ 1011 if(compare == 0) 1012 compare = APInt::tcCompare(significandParts(), rhs.significandParts(), 1013 partCount()); 1014 1015 if(compare > 0) 1016 return cmpGreaterThan; 1017 else if(compare < 0) 1018 return cmpLessThan; 1019 else 1020 return cmpEqual; 1021} 1022 1023/* Handle overflow. Sign is preserved. We either become infinity or 1024 the largest finite number. */ 1025APFloat::opStatus 1026APFloat::handleOverflow(roundingMode rounding_mode) 1027{ 1028 /* Infinity? */ 1029 if(rounding_mode == rmNearestTiesToEven 1030 || rounding_mode == rmNearestTiesToAway 1031 || (rounding_mode == rmTowardPositive && !sign) 1032 || (rounding_mode == rmTowardNegative && sign)) 1033 { 1034 category = fcInfinity; 1035 return (opStatus) (opOverflow | opInexact); 1036 } 1037 1038 /* Otherwise we become the largest finite number. */ 1039 category = fcNormal; 1040 exponent = semantics->maxExponent; 1041 APInt::tcSetLeastSignificantBits(significandParts(), partCount(), 1042 semantics->precision); 1043 1044 return opInexact; 1045} 1046 1047/* Returns TRUE if, when truncating the current number, with BIT the 1048 new LSB, with the given lost fraction and rounding mode, the result 1049 would need to be rounded away from zero (i.e., by increasing the 1050 signficand). This routine must work for fcZero of both signs, and 1051 fcNormal numbers. */ 1052bool 1053APFloat::roundAwayFromZero(roundingMode rounding_mode, 1054 lostFraction lost_fraction, 1055 unsigned int bit) const 1056{ 1057 /* NaNs and infinities should not have lost fractions. */ 1058 assert(category == fcNormal || category == fcZero); 1059 1060 /* Current callers never pass this so we don't handle it. */ 1061 assert(lost_fraction != lfExactlyZero); 1062 1063 switch(rounding_mode) { 1064 default: 1065 assert(0); 1066 1067 case rmNearestTiesToAway: 1068 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf; 1069 1070 case rmNearestTiesToEven: 1071 if(lost_fraction == lfMoreThanHalf) 1072 return true; 1073 1074 /* Our zeroes don't have a significand to test. */ 1075 if(lost_fraction == lfExactlyHalf && category != fcZero) 1076 return APInt::tcExtractBit(significandParts(), bit); 1077 1078 return false; 1079 1080 case rmTowardZero: 1081 return false; 1082 1083 case rmTowardPositive: 1084 return sign == false; 1085 1086 case rmTowardNegative: 1087 return sign == true; 1088 } 1089} 1090 1091APFloat::opStatus 1092APFloat::normalize(roundingMode rounding_mode, 1093 lostFraction lost_fraction) 1094{ 1095 unsigned int omsb; /* One, not zero, based MSB. */ 1096 int exponentChange; 1097 1098 if(category != fcNormal) 1099 return opOK; 1100 1101 /* Before rounding normalize the exponent of fcNormal numbers. */ 1102 omsb = significandMSB() + 1; 1103 1104 if(omsb) { 1105 /* OMSB is numbered from 1. We want to place it in the integer 1106 bit numbered PRECISON if possible, with a compensating change in 1107 the exponent. */ 1108 exponentChange = omsb - semantics->precision; 1109 1110 /* If the resulting exponent is too high, overflow according to 1111 the rounding mode. */ 1112 if(exponent + exponentChange > semantics->maxExponent) 1113 return handleOverflow(rounding_mode); 1114 1115 /* Subnormal numbers have exponent minExponent, and their MSB 1116 is forced based on that. */ 1117 if(exponent + exponentChange < semantics->minExponent) 1118 exponentChange = semantics->minExponent - exponent; 1119 1120 /* Shifting left is easy as we don't lose precision. */ 1121 if(exponentChange < 0) { 1122 assert(lost_fraction == lfExactlyZero); 1123 1124 shiftSignificandLeft(-exponentChange); 1125 1126 return opOK; 1127 } 1128 1129 if(exponentChange > 0) { 1130 lostFraction lf; 1131 1132 /* Shift right and capture any new lost fraction. */ 1133 lf = shiftSignificandRight(exponentChange); 1134 1135 lost_fraction = combineLostFractions(lf, lost_fraction); 1136 1137 /* Keep OMSB up-to-date. */ 1138 if(omsb > (unsigned) exponentChange) 1139 omsb -= exponentChange; 1140 else 1141 omsb = 0; 1142 } 1143 } 1144 1145 /* Now round the number according to rounding_mode given the lost 1146 fraction. */ 1147 1148 /* As specified in IEEE 754, since we do not trap we do not report 1149 underflow for exact results. */ 1150 if(lost_fraction == lfExactlyZero) { 1151 /* Canonicalize zeroes. */ 1152 if(omsb == 0) 1153 category = fcZero; 1154 1155 return opOK; 1156 } 1157 1158 /* Increment the significand if we're rounding away from zero. */ 1159 if(roundAwayFromZero(rounding_mode, lost_fraction, 0)) { 1160 if(omsb == 0) 1161 exponent = semantics->minExponent; 1162 1163 incrementSignificand(); 1164 omsb = significandMSB() + 1; 1165 1166 /* Did the significand increment overflow? */ 1167 if(omsb == (unsigned) semantics->precision + 1) { 1168 /* Renormalize by incrementing the exponent and shifting our 1169 significand right one. However if we already have the 1170 maximum exponent we overflow to infinity. */ 1171 if(exponent == semantics->maxExponent) { 1172 category = fcInfinity; 1173 1174 return (opStatus) (opOverflow | opInexact); 1175 } 1176 1177 shiftSignificandRight(1); 1178 1179 return opInexact; 1180 } 1181 } 1182 1183 /* The normal case - we were and are not denormal, and any 1184 significand increment above didn't overflow. */ 1185 if(omsb == semantics->precision) 1186 return opInexact; 1187 1188 /* We have a non-zero denormal. */ 1189 assert(omsb < semantics->precision); 1190 1191 /* Canonicalize zeroes. */ 1192 if(omsb == 0) 1193 category = fcZero; 1194 1195 /* The fcZero case is a denormal that underflowed to zero. */ 1196 return (opStatus) (opUnderflow | opInexact); 1197} 1198 1199APFloat::opStatus 1200APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract) 1201{ 1202 switch(convolve(category, rhs.category)) { 1203 default: 1204 assert(0); 1205 1206 case convolve(fcNaN, fcZero): 1207 case convolve(fcNaN, fcNormal): 1208 case convolve(fcNaN, fcInfinity): 1209 case convolve(fcNaN, fcNaN): 1210 case convolve(fcNormal, fcZero): 1211 case convolve(fcInfinity, fcNormal): 1212 case convolve(fcInfinity, fcZero): 1213 return opOK; 1214 1215 case convolve(fcZero, fcNaN): 1216 case convolve(fcNormal, fcNaN): 1217 case convolve(fcInfinity, fcNaN): 1218 category = fcNaN; 1219 copySignificand(rhs); 1220 return opOK; 1221 1222 case convolve(fcNormal, fcInfinity): 1223 case convolve(fcZero, fcInfinity): 1224 category = fcInfinity; 1225 sign = rhs.sign ^ subtract; 1226 return opOK; 1227 1228 case convolve(fcZero, fcNormal): 1229 assign(rhs); 1230 sign = rhs.sign ^ subtract; 1231 return opOK; 1232 1233 case convolve(fcZero, fcZero): 1234 /* Sign depends on rounding mode; handled by caller. */ 1235 return opOK; 1236 1237 case convolve(fcInfinity, fcInfinity): 1238 /* Differently signed infinities can only be validly 1239 subtracted. */ 1240 if((sign ^ rhs.sign) != subtract) { 1241 makeNaN(); 1242 return opInvalidOp; 1243 } 1244 1245 return opOK; 1246 1247 case convolve(fcNormal, fcNormal): 1248 return opDivByZero; 1249 } 1250} 1251 1252/* Add or subtract two normal numbers. */ 1253lostFraction 1254APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract) 1255{ 1256 integerPart carry; 1257 lostFraction lost_fraction; 1258 int bits; 1259 1260 /* Determine if the operation on the absolute values is effectively 1261 an addition or subtraction. */ 1262 subtract ^= (sign ^ rhs.sign) ? true : false; 1263 1264 /* Are we bigger exponent-wise than the RHS? */ 1265 bits = exponent - rhs.exponent; 1266 1267 /* Subtraction is more subtle than one might naively expect. */ 1268 if(subtract) { 1269 APFloat temp_rhs(rhs); 1270 bool reverse; 1271 1272 if (bits == 0) { 1273 reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan; 1274 lost_fraction = lfExactlyZero; 1275 } else if (bits > 0) { 1276 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1); 1277 shiftSignificandLeft(1); 1278 reverse = false; 1279 } else { 1280 lost_fraction = shiftSignificandRight(-bits - 1); 1281 temp_rhs.shiftSignificandLeft(1); 1282 reverse = true; 1283 } 1284 1285 if (reverse) { 1286 carry = temp_rhs.subtractSignificand 1287 (*this, lost_fraction != lfExactlyZero); 1288 copySignificand(temp_rhs); 1289 sign = !sign; 1290 } else { 1291 carry = subtractSignificand 1292 (temp_rhs, lost_fraction != lfExactlyZero); 1293 } 1294 1295 /* Invert the lost fraction - it was on the RHS and 1296 subtracted. */ 1297 if(lost_fraction == lfLessThanHalf) 1298 lost_fraction = lfMoreThanHalf; 1299 else if(lost_fraction == lfMoreThanHalf) 1300 lost_fraction = lfLessThanHalf; 1301 1302 /* The code above is intended to ensure that no borrow is 1303 necessary. */ 1304 assert(!carry); 1305 } else { 1306 if(bits > 0) { 1307 APFloat temp_rhs(rhs); 1308 1309 lost_fraction = temp_rhs.shiftSignificandRight(bits); 1310 carry = addSignificand(temp_rhs); 1311 } else { 1312 lost_fraction = shiftSignificandRight(-bits); 1313 carry = addSignificand(rhs); 1314 } 1315 1316 /* We have a guard bit; generating a carry cannot happen. */ 1317 assert(!carry); 1318 } 1319 1320 return lost_fraction; 1321} 1322 1323APFloat::opStatus 1324APFloat::multiplySpecials(const APFloat &rhs) 1325{ 1326 switch(convolve(category, rhs.category)) { 1327 default: 1328 assert(0); 1329 1330 case convolve(fcNaN, fcZero): 1331 case convolve(fcNaN, fcNormal): 1332 case convolve(fcNaN, fcInfinity): 1333 case convolve(fcNaN, fcNaN): 1334 return opOK; 1335 1336 case convolve(fcZero, fcNaN): 1337 case convolve(fcNormal, fcNaN): 1338 case convolve(fcInfinity, fcNaN): 1339 category = fcNaN; 1340 copySignificand(rhs); 1341 return opOK; 1342 1343 case convolve(fcNormal, fcInfinity): 1344 case convolve(fcInfinity, fcNormal): 1345 case convolve(fcInfinity, fcInfinity): 1346 category = fcInfinity; 1347 return opOK; 1348 1349 case convolve(fcZero, fcNormal): 1350 case convolve(fcNormal, fcZero): 1351 case convolve(fcZero, fcZero): 1352 category = fcZero; 1353 return opOK; 1354 1355 case convolve(fcZero, fcInfinity): 1356 case convolve(fcInfinity, fcZero): 1357 makeNaN(); 1358 return opInvalidOp; 1359 1360 case convolve(fcNormal, fcNormal): 1361 return opOK; 1362 } 1363} 1364 1365APFloat::opStatus 1366APFloat::divideSpecials(const APFloat &rhs) 1367{ 1368 switch(convolve(category, rhs.category)) { 1369 default: 1370 assert(0); 1371 1372 case convolve(fcNaN, fcZero): 1373 case convolve(fcNaN, fcNormal): 1374 case convolve(fcNaN, fcInfinity): 1375 case convolve(fcNaN, fcNaN): 1376 case convolve(fcInfinity, fcZero): 1377 case convolve(fcInfinity, fcNormal): 1378 case convolve(fcZero, fcInfinity): 1379 case convolve(fcZero, fcNormal): 1380 return opOK; 1381 1382 case convolve(fcZero, fcNaN): 1383 case convolve(fcNormal, fcNaN): 1384 case convolve(fcInfinity, fcNaN): 1385 category = fcNaN; 1386 copySignificand(rhs); 1387 return opOK; 1388 1389 case convolve(fcNormal, fcInfinity): 1390 category = fcZero; 1391 return opOK; 1392 1393 case convolve(fcNormal, fcZero): 1394 category = fcInfinity; 1395 return opDivByZero; 1396 1397 case convolve(fcInfinity, fcInfinity): 1398 case convolve(fcZero, fcZero): 1399 makeNaN(); 1400 return opInvalidOp; 1401 1402 case convolve(fcNormal, fcNormal): 1403 return opOK; 1404 } 1405} 1406 1407/* Change sign. */ 1408void 1409APFloat::changeSign() 1410{ 1411 /* Look mummy, this one's easy. */ 1412 sign = !sign; 1413} 1414 1415void 1416APFloat::clearSign() 1417{ 1418 /* So is this one. */ 1419 sign = 0; 1420} 1421 1422void 1423APFloat::copySign(const APFloat &rhs) 1424{ 1425 /* And this one. */ 1426 sign = rhs.sign; 1427} 1428 1429/* Normalized addition or subtraction. */ 1430APFloat::opStatus 1431APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode, 1432 bool subtract) 1433{ 1434 opStatus fs; 1435 1436 assertArithmeticOK(*semantics); 1437 1438 fs = addOrSubtractSpecials(rhs, subtract); 1439 1440 /* This return code means it was not a simple case. */ 1441 if(fs == opDivByZero) { 1442 lostFraction lost_fraction; 1443 1444 lost_fraction = addOrSubtractSignificand(rhs, subtract); 1445 fs = normalize(rounding_mode, lost_fraction); 1446 1447 /* Can only be zero if we lost no fraction. */ 1448 assert(category != fcZero || lost_fraction == lfExactlyZero); 1449 } 1450 1451 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a 1452 positive zero unless rounding to minus infinity, except that 1453 adding two like-signed zeroes gives that zero. */ 1454 if(category == fcZero) { 1455 if(rhs.category != fcZero || (sign == rhs.sign) == subtract) 1456 sign = (rounding_mode == rmTowardNegative); 1457 } 1458 1459 return fs; 1460} 1461 1462/* Normalized addition. */ 1463APFloat::opStatus 1464APFloat::add(const APFloat &rhs, roundingMode rounding_mode) 1465{ 1466 return addOrSubtract(rhs, rounding_mode, false); 1467} 1468 1469/* Normalized subtraction. */ 1470APFloat::opStatus 1471APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode) 1472{ 1473 return addOrSubtract(rhs, rounding_mode, true); 1474} 1475 1476/* Normalized multiply. */ 1477APFloat::opStatus 1478APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode) 1479{ 1480 opStatus fs; 1481 1482 assertArithmeticOK(*semantics); 1483 sign ^= rhs.sign; 1484 fs = multiplySpecials(rhs); 1485 1486 if(category == fcNormal) { 1487 lostFraction lost_fraction = multiplySignificand(rhs, 0); 1488 fs = normalize(rounding_mode, lost_fraction); 1489 if(lost_fraction != lfExactlyZero) 1490 fs = (opStatus) (fs | opInexact); 1491 } 1492 1493 return fs; 1494} 1495 1496/* Normalized divide. */ 1497APFloat::opStatus 1498APFloat::divide(const APFloat &rhs, roundingMode rounding_mode) 1499{ 1500 opStatus fs; 1501 1502 assertArithmeticOK(*semantics); 1503 sign ^= rhs.sign; 1504 fs = divideSpecials(rhs); 1505 1506 if(category == fcNormal) { 1507 lostFraction lost_fraction = divideSignificand(rhs); 1508 fs = normalize(rounding_mode, lost_fraction); 1509 if(lost_fraction != lfExactlyZero) 1510 fs = (opStatus) (fs | opInexact); 1511 } 1512 1513 return fs; 1514} 1515 1516/* Normalized remainder. This is not currently doing TRT. */ 1517APFloat::opStatus 1518APFloat::mod(const APFloat &rhs, roundingMode rounding_mode) 1519{ 1520 opStatus fs; 1521 APFloat V = *this; 1522 unsigned int origSign = sign; 1523 1524 assertArithmeticOK(*semantics); 1525 fs = V.divide(rhs, rmNearestTiesToEven); 1526 if (fs == opDivByZero) 1527 return fs; 1528 1529 int parts = partCount(); 1530 integerPart *x = new integerPart[parts]; 1531 fs = V.convertToInteger(x, parts * integerPartWidth, true, 1532 rmNearestTiesToEven); 1533 if (fs==opInvalidOp) 1534 return fs; 1535 1536 fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true, 1537 rmNearestTiesToEven); 1538 assert(fs==opOK); // should always work 1539 1540 fs = V.multiply(rhs, rounding_mode); 1541 assert(fs==opOK || fs==opInexact); // should not overflow or underflow 1542 1543 fs = subtract(V, rounding_mode); 1544 assert(fs==opOK || fs==opInexact); // likewise 1545 1546 if (isZero()) 1547 sign = origSign; // IEEE754 requires this 1548 delete[] x; 1549 return fs; 1550} 1551 1552/* Normalized fused-multiply-add. */ 1553APFloat::opStatus 1554APFloat::fusedMultiplyAdd(const APFloat &multiplicand, 1555 const APFloat &addend, 1556 roundingMode rounding_mode) 1557{ 1558 opStatus fs; 1559 1560 assertArithmeticOK(*semantics); 1561 1562 /* Post-multiplication sign, before addition. */ 1563 sign ^= multiplicand.sign; 1564 1565 /* If and only if all arguments are normal do we need to do an 1566 extended-precision calculation. */ 1567 if(category == fcNormal 1568 && multiplicand.category == fcNormal 1569 && addend.category == fcNormal) { 1570 lostFraction lost_fraction; 1571 1572 lost_fraction = multiplySignificand(multiplicand, &addend); 1573 fs = normalize(rounding_mode, lost_fraction); 1574 if(lost_fraction != lfExactlyZero) 1575 fs = (opStatus) (fs | opInexact); 1576 1577 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a 1578 positive zero unless rounding to minus infinity, except that 1579 adding two like-signed zeroes gives that zero. */ 1580 if(category == fcZero && sign != addend.sign) 1581 sign = (rounding_mode == rmTowardNegative); 1582 } else { 1583 fs = multiplySpecials(multiplicand); 1584 1585 /* FS can only be opOK or opInvalidOp. There is no more work 1586 to do in the latter case. The IEEE-754R standard says it is 1587 implementation-defined in this case whether, if ADDEND is a 1588 quiet NaN, we raise invalid op; this implementation does so. 1589 1590 If we need to do the addition we can do so with normal 1591 precision. */ 1592 if(fs == opOK) 1593 fs = addOrSubtract(addend, rounding_mode, false); 1594 } 1595 1596 return fs; 1597} 1598 1599/* Comparison requires normalized numbers. */ 1600APFloat::cmpResult 1601APFloat::compare(const APFloat &rhs) const 1602{ 1603 cmpResult result; 1604 1605 assertArithmeticOK(*semantics); 1606 assert(semantics == rhs.semantics); 1607 1608 switch(convolve(category, rhs.category)) { 1609 default: 1610 assert(0); 1611 1612 case convolve(fcNaN, fcZero): 1613 case convolve(fcNaN, fcNormal): 1614 case convolve(fcNaN, fcInfinity): 1615 case convolve(fcNaN, fcNaN): 1616 case convolve(fcZero, fcNaN): 1617 case convolve(fcNormal, fcNaN): 1618 case convolve(fcInfinity, fcNaN): 1619 return cmpUnordered; 1620 1621 case convolve(fcInfinity, fcNormal): 1622 case convolve(fcInfinity, fcZero): 1623 case convolve(fcNormal, fcZero): 1624 if(sign) 1625 return cmpLessThan; 1626 else 1627 return cmpGreaterThan; 1628 1629 case convolve(fcNormal, fcInfinity): 1630 case convolve(fcZero, fcInfinity): 1631 case convolve(fcZero, fcNormal): 1632 if(rhs.sign) 1633 return cmpGreaterThan; 1634 else 1635 return cmpLessThan; 1636 1637 case convolve(fcInfinity, fcInfinity): 1638 if(sign == rhs.sign) 1639 return cmpEqual; 1640 else if(sign) 1641 return cmpLessThan; 1642 else 1643 return cmpGreaterThan; 1644 1645 case convolve(fcZero, fcZero): 1646 return cmpEqual; 1647 1648 case convolve(fcNormal, fcNormal): 1649 break; 1650 } 1651 1652 /* Two normal numbers. Do they have the same sign? */ 1653 if(sign != rhs.sign) { 1654 if(sign) 1655 result = cmpLessThan; 1656 else 1657 result = cmpGreaterThan; 1658 } else { 1659 /* Compare absolute values; invert result if negative. */ 1660 result = compareAbsoluteValue(rhs); 1661 1662 if(sign) { 1663 if(result == cmpLessThan) 1664 result = cmpGreaterThan; 1665 else if(result == cmpGreaterThan) 1666 result = cmpLessThan; 1667 } 1668 } 1669 1670 return result; 1671} 1672 1673APFloat::opStatus 1674APFloat::convert(const fltSemantics &toSemantics, 1675 roundingMode rounding_mode) 1676{ 1677 lostFraction lostFraction; 1678 unsigned int newPartCount, oldPartCount; 1679 opStatus fs; 1680 1681 assertArithmeticOK(*semantics); 1682 assertArithmeticOK(toSemantics); 1683 lostFraction = lfExactlyZero; 1684 newPartCount = partCountForBits(toSemantics.precision + 1); 1685 oldPartCount = partCount(); 1686 1687 /* Handle storage complications. If our new form is wider, 1688 re-allocate our bit pattern into wider storage. If it is 1689 narrower, we ignore the excess parts, but if narrowing to a 1690 single part we need to free the old storage. 1691 Be careful not to reference significandParts for zeroes 1692 and infinities, since it aborts. */ 1693 if (newPartCount > oldPartCount) { 1694 integerPart *newParts; 1695 newParts = new integerPart[newPartCount]; 1696 APInt::tcSet(newParts, 0, newPartCount); 1697 if (category==fcNormal || category==fcNaN) 1698 APInt::tcAssign(newParts, significandParts(), oldPartCount); 1699 freeSignificand(); 1700 significand.parts = newParts; 1701 } else if (newPartCount < oldPartCount) { 1702 /* Capture any lost fraction through truncation of parts so we get 1703 correct rounding whilst normalizing. */ 1704 if (category==fcNormal) 1705 lostFraction = lostFractionThroughTruncation 1706 (significandParts(), oldPartCount, toSemantics.precision); 1707 if (newPartCount == 1) { 1708 integerPart newPart = 0; 1709 if (category==fcNormal || category==fcNaN) 1710 newPart = significandParts()[0]; 1711 freeSignificand(); 1712 significand.part = newPart; 1713 } 1714 } 1715 1716 if(category == fcNormal) { 1717 /* Re-interpret our bit-pattern. */ 1718 exponent += toSemantics.precision - semantics->precision; 1719 semantics = &toSemantics; 1720 fs = normalize(rounding_mode, lostFraction); 1721 } else if (category == fcNaN) { 1722 int shift = toSemantics.precision - semantics->precision; 1723 // Do this now so significandParts gets the right answer 1724 semantics = &toSemantics; 1725 // No normalization here, just truncate 1726 if (shift>0) 1727 APInt::tcShiftLeft(significandParts(), newPartCount, shift); 1728 else if (shift < 0) 1729 APInt::tcShiftRight(significandParts(), newPartCount, -shift); 1730 // gcc forces the Quiet bit on, which means (float)(double)(float_sNan) 1731 // does not give you back the same bits. This is dubious, and we 1732 // don't currently do it. You're really supposed to get 1733 // an invalid operation signal at runtime, but nobody does that. 1734 fs = opOK; 1735 } else { 1736 semantics = &toSemantics; 1737 fs = opOK; 1738 } 1739 1740 return fs; 1741} 1742 1743/* Convert a floating point number to an integer according to the 1744 rounding mode. If the rounded integer value is out of range this 1745 returns an invalid operation exception and the contents of the 1746 destination parts are unspecified. If the rounded value is in 1747 range but the floating point number is not the exact integer, the C 1748 standard doesn't require an inexact exception to be raised. IEEE 1749 854 does require it so we do that. 1750 1751 Note that for conversions to integer type the C standard requires 1752 round-to-zero to always be used. */ 1753APFloat::opStatus 1754APFloat::convertToSignExtendedInteger(integerPart *parts, unsigned int width, 1755 bool isSigned, 1756 roundingMode rounding_mode) const 1757{ 1758 lostFraction lost_fraction; 1759 const integerPart *src; 1760 unsigned int dstPartsCount, truncatedBits; 1761 1762 assertArithmeticOK(*semantics); 1763 1764 /* Handle the three special cases first. */ 1765 if(category == fcInfinity || category == fcNaN) 1766 return opInvalidOp; 1767 1768 dstPartsCount = partCountForBits(width); 1769 1770 if(category == fcZero) { 1771 APInt::tcSet(parts, 0, dstPartsCount); 1772 return opOK; 1773 } 1774 1775 src = significandParts(); 1776 1777 /* Step 1: place our absolute value, with any fraction truncated, in 1778 the destination. */ 1779 if (exponent < 0) { 1780 /* Our absolute value is less than one; truncate everything. */ 1781 APInt::tcSet(parts, 0, dstPartsCount); 1782 truncatedBits = semantics->precision; 1783 } else { 1784 /* We want the most significant (exponent + 1) bits; the rest are 1785 truncated. */ 1786 unsigned int bits = exponent + 1U; 1787 1788 /* Hopelessly large in magnitude? */ 1789 if (bits > width) 1790 return opInvalidOp; 1791 1792 if (bits < semantics->precision) { 1793 /* We truncate (semantics->precision - bits) bits. */ 1794 truncatedBits = semantics->precision - bits; 1795 APInt::tcExtract(parts, dstPartsCount, src, bits, truncatedBits); 1796 } else { 1797 /* We want at least as many bits as are available. */ 1798 APInt::tcExtract(parts, dstPartsCount, src, semantics->precision, 0); 1799 APInt::tcShiftLeft(parts, dstPartsCount, bits - semantics->precision); 1800 truncatedBits = 0; 1801 } 1802 } 1803 1804 /* Step 2: work out any lost fraction, and increment the absolute 1805 value if we would round away from zero. */ 1806 if (truncatedBits) { 1807 lost_fraction = lostFractionThroughTruncation(src, partCount(), 1808 truncatedBits); 1809 if (lost_fraction != lfExactlyZero 1810 && roundAwayFromZero(rounding_mode, lost_fraction, truncatedBits)) { 1811 if (APInt::tcIncrement(parts, dstPartsCount)) 1812 return opInvalidOp; /* Overflow. */ 1813 } 1814 } else { 1815 lost_fraction = lfExactlyZero; 1816 } 1817 1818 /* Step 3: check if we fit in the destination. */ 1819 unsigned int omsb = APInt::tcMSB(parts, dstPartsCount) + 1; 1820 1821 if (sign) { 1822 if (!isSigned) { 1823 /* Negative numbers cannot be represented as unsigned. */ 1824 if (omsb != 0) 1825 return opInvalidOp; 1826 } else { 1827 /* It takes omsb bits to represent the unsigned integer value. 1828 We lose a bit for the sign, but care is needed as the 1829 maximally negative integer is a special case. */ 1830 if (omsb == width && APInt::tcLSB(parts, dstPartsCount) + 1 != omsb) 1831 return opInvalidOp; 1832 1833 /* This case can happen because of rounding. */ 1834 if (omsb > width) 1835 return opInvalidOp; 1836 } 1837 1838 APInt::tcNegate (parts, dstPartsCount); 1839 } else { 1840 if (omsb >= width + !isSigned) 1841 return opInvalidOp; 1842 } 1843 1844 if (lost_fraction == lfExactlyZero) 1845 return opOK; 1846 else 1847 return opInexact; 1848} 1849 1850/* Same as convertToSignExtendedInteger, except we provide 1851 deterministic values in case of an invalid operation exception, 1852 namely zero for NaNs and the minimal or maximal value respectively 1853 for underflow or overflow. */ 1854APFloat::opStatus 1855APFloat::convertToInteger(integerPart *parts, unsigned int width, 1856 bool isSigned, 1857 roundingMode rounding_mode) const 1858{ 1859 opStatus fs; 1860 1861 fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode); 1862 1863 if (fs == opInvalidOp) { 1864 unsigned int bits, dstPartsCount; 1865 1866 dstPartsCount = partCountForBits(width); 1867 1868 if (category == fcNaN) 1869 bits = 0; 1870 else if (sign) 1871 bits = isSigned; 1872 else 1873 bits = width - isSigned; 1874 1875 APInt::tcSetLeastSignificantBits(parts, dstPartsCount, bits); 1876 if (sign && isSigned) 1877 APInt::tcShiftLeft(parts, dstPartsCount, width - 1); 1878 } 1879 1880 return fs; 1881} 1882 1883/* Convert an unsigned integer SRC to a floating point number, 1884 rounding according to ROUNDING_MODE. The sign of the floating 1885 point number is not modified. */ 1886APFloat::opStatus 1887APFloat::convertFromUnsignedParts(const integerPart *src, 1888 unsigned int srcCount, 1889 roundingMode rounding_mode) 1890{ 1891 unsigned int omsb, precision, dstCount; 1892 integerPart *dst; 1893 lostFraction lost_fraction; 1894 1895 assertArithmeticOK(*semantics); 1896 category = fcNormal; 1897 omsb = APInt::tcMSB(src, srcCount) + 1; 1898 dst = significandParts(); 1899 dstCount = partCount(); 1900 precision = semantics->precision; 1901 1902 /* We want the most significant PRECISON bits of SRC. There may not 1903 be that many; extract what we can. */ 1904 if (precision <= omsb) { 1905 exponent = omsb - 1; 1906 lost_fraction = lostFractionThroughTruncation(src, srcCount, 1907 omsb - precision); 1908 APInt::tcExtract(dst, dstCount, src, precision, omsb - precision); 1909 } else { 1910 exponent = precision - 1; 1911 lost_fraction = lfExactlyZero; 1912 APInt::tcExtract(dst, dstCount, src, omsb, 0); 1913 } 1914 1915 return normalize(rounding_mode, lost_fraction); 1916} 1917 1918APFloat::opStatus 1919APFloat::convertFromAPInt(const APInt &Val, 1920 bool isSigned, 1921 roundingMode rounding_mode) 1922{ 1923 unsigned int partCount = Val.getNumWords(); 1924 APInt api = Val; 1925 1926 sign = false; 1927 if (isSigned && api.isNegative()) { 1928 sign = true; 1929 api = -api; 1930 } 1931 1932 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode); 1933} 1934 1935/* Convert a two's complement integer SRC to a floating point number, 1936 rounding according to ROUNDING_MODE. ISSIGNED is true if the 1937 integer is signed, in which case it must be sign-extended. */ 1938APFloat::opStatus 1939APFloat::convertFromSignExtendedInteger(const integerPart *src, 1940 unsigned int srcCount, 1941 bool isSigned, 1942 roundingMode rounding_mode) 1943{ 1944 opStatus status; 1945 1946 assertArithmeticOK(*semantics); 1947 if (isSigned 1948 && APInt::tcExtractBit(src, srcCount * integerPartWidth - 1)) { 1949 integerPart *copy; 1950 1951 /* If we're signed and negative negate a copy. */ 1952 sign = true; 1953 copy = new integerPart[srcCount]; 1954 APInt::tcAssign(copy, src, srcCount); 1955 APInt::tcNegate(copy, srcCount); 1956 status = convertFromUnsignedParts(copy, srcCount, rounding_mode); 1957 delete [] copy; 1958 } else { 1959 sign = false; 1960 status = convertFromUnsignedParts(src, srcCount, rounding_mode); 1961 } 1962 1963 return status; 1964} 1965 1966/* FIXME: should this just take a const APInt reference? */ 1967APFloat::opStatus 1968APFloat::convertFromZeroExtendedInteger(const integerPart *parts, 1969 unsigned int width, bool isSigned, 1970 roundingMode rounding_mode) 1971{ 1972 unsigned int partCount = partCountForBits(width); 1973 APInt api = APInt(width, partCount, parts); 1974 1975 sign = false; 1976 if(isSigned && APInt::tcExtractBit(parts, width - 1)) { 1977 sign = true; 1978 api = -api; 1979 } 1980 1981 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode); 1982} 1983 1984APFloat::opStatus 1985APFloat::convertFromHexadecimalString(const char *p, 1986 roundingMode rounding_mode) 1987{ 1988 lostFraction lost_fraction; 1989 integerPart *significand; 1990 unsigned int bitPos, partsCount; 1991 const char *dot, *firstSignificantDigit; 1992 1993 zeroSignificand(); 1994 exponent = 0; 1995 category = fcNormal; 1996 1997 significand = significandParts(); 1998 partsCount = partCount(); 1999 bitPos = partsCount * integerPartWidth; 2000 2001 /* Skip leading zeroes and any (hexa)decimal point. */ 2002 p = skipLeadingZeroesAndAnyDot(p, &dot); 2003 firstSignificantDigit = p; 2004 2005 for(;;) { 2006 integerPart hex_value; 2007 2008 if(*p == '.') { 2009 assert(dot == 0); 2010 dot = p++; 2011 } 2012 2013 hex_value = hexDigitValue(*p); 2014 if(hex_value == -1U) { 2015 lost_fraction = lfExactlyZero; 2016 break; 2017 } 2018 2019 p++; 2020 2021 /* Store the number whilst 4-bit nibbles remain. */ 2022 if(bitPos) { 2023 bitPos -= 4; 2024 hex_value <<= bitPos % integerPartWidth; 2025 significand[bitPos / integerPartWidth] |= hex_value; 2026 } else { 2027 lost_fraction = trailingHexadecimalFraction(p, hex_value); 2028 while(hexDigitValue(*p) != -1U) 2029 p++; 2030 break; 2031 } 2032 } 2033 2034 /* Hex floats require an exponent but not a hexadecimal point. */ 2035 assert(*p == 'p' || *p == 'P'); 2036 2037 /* Ignore the exponent if we are zero. */ 2038 if(p != firstSignificantDigit) { 2039 int expAdjustment; 2040 2041 /* Implicit hexadecimal point? */ 2042 if(!dot) 2043 dot = p; 2044 2045 /* Calculate the exponent adjustment implicit in the number of 2046 significant digits. */ 2047 expAdjustment = static_cast<int>(dot - firstSignificantDigit); 2048 if(expAdjustment < 0) 2049 expAdjustment++; 2050 expAdjustment = expAdjustment * 4 - 1; 2051 2052 /* Adjust for writing the significand starting at the most 2053 significant nibble. */ 2054 expAdjustment += semantics->precision; 2055 expAdjustment -= partsCount * integerPartWidth; 2056 2057 /* Adjust for the given exponent. */ 2058 exponent = totalExponent(p, expAdjustment); 2059 } 2060 2061 return normalize(rounding_mode, lost_fraction); 2062} 2063 2064APFloat::opStatus 2065APFloat::roundSignificandWithExponent(const integerPart *decSigParts, 2066 unsigned sigPartCount, int exp, 2067 roundingMode rounding_mode) 2068{ 2069 unsigned int parts, pow5PartCount; 2070 fltSemantics calcSemantics = { 32767, -32767, 0, true }; 2071 integerPart pow5Parts[maxPowerOfFiveParts]; 2072 bool isNearest; 2073 2074 isNearest = (rounding_mode == rmNearestTiesToEven 2075 || rounding_mode == rmNearestTiesToAway); 2076 2077 parts = partCountForBits(semantics->precision + 11); 2078 2079 /* Calculate pow(5, abs(exp)). */ 2080 pow5PartCount = powerOf5(pow5Parts, exp >= 0 ? exp: -exp); 2081 2082 for (;; parts *= 2) { 2083 opStatus sigStatus, powStatus; 2084 unsigned int excessPrecision, truncatedBits; 2085 2086 calcSemantics.precision = parts * integerPartWidth - 1; 2087 excessPrecision = calcSemantics.precision - semantics->precision; 2088 truncatedBits = excessPrecision; 2089 2090 APFloat decSig(calcSemantics, fcZero, sign); 2091 APFloat pow5(calcSemantics, fcZero, false); 2092 2093 sigStatus = decSig.convertFromUnsignedParts(decSigParts, sigPartCount, 2094 rmNearestTiesToEven); 2095 powStatus = pow5.convertFromUnsignedParts(pow5Parts, pow5PartCount, 2096 rmNearestTiesToEven); 2097 /* Add exp, as 10^n = 5^n * 2^n. */ 2098 decSig.exponent += exp; 2099 2100 lostFraction calcLostFraction; 2101 integerPart HUerr, HUdistance; 2102 unsigned int powHUerr; 2103 2104 if (exp >= 0) { 2105 /* multiplySignificand leaves the precision-th bit set to 1. */ 2106 calcLostFraction = decSig.multiplySignificand(pow5, NULL); 2107 powHUerr = powStatus != opOK; 2108 } else { 2109 calcLostFraction = decSig.divideSignificand(pow5); 2110 /* Denormal numbers have less precision. */ 2111 if (decSig.exponent < semantics->minExponent) { 2112 excessPrecision += (semantics->minExponent - decSig.exponent); 2113 truncatedBits = excessPrecision; 2114 if (excessPrecision > calcSemantics.precision) 2115 excessPrecision = calcSemantics.precision; 2116 } 2117 /* Extra half-ulp lost in reciprocal of exponent. */ 2118 powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2; 2119 } 2120 2121 /* Both multiplySignificand and divideSignificand return the 2122 result with the integer bit set. */ 2123 assert (APInt::tcExtractBit 2124 (decSig.significandParts(), calcSemantics.precision - 1) == 1); 2125 2126 HUerr = HUerrBound(calcLostFraction != lfExactlyZero, sigStatus != opOK, 2127 powHUerr); 2128 HUdistance = 2 * ulpsFromBoundary(decSig.significandParts(), 2129 excessPrecision, isNearest); 2130 2131 /* Are we guaranteed to round correctly if we truncate? */ 2132 if (HUdistance >= HUerr) { 2133 APInt::tcExtract(significandParts(), partCount(), decSig.significandParts(), 2134 calcSemantics.precision - excessPrecision, 2135 excessPrecision); 2136 /* Take the exponent of decSig. If we tcExtract-ed less bits 2137 above we must adjust our exponent to compensate for the 2138 implicit right shift. */ 2139 exponent = (decSig.exponent + semantics->precision 2140 - (calcSemantics.precision - excessPrecision)); 2141 calcLostFraction = lostFractionThroughTruncation(decSig.significandParts(), 2142 decSig.partCount(), 2143 truncatedBits); 2144 return normalize(rounding_mode, calcLostFraction); 2145 } 2146 } 2147} 2148 2149APFloat::opStatus 2150APFloat::convertFromDecimalString(const char *p, roundingMode rounding_mode) 2151{ 2152 decimalInfo D; 2153 opStatus fs; 2154 2155 /* Scan the text. */ 2156 interpretDecimal(p, &D); 2157 2158 /* Handle the quick cases. First the case of no significant digits, 2159 i.e. zero, and then exponents that are obviously too large or too 2160 small. Writing L for log 10 / log 2, a number d.ddddd*10^exp 2161 definitely overflows if 2162 2163 (exp - 1) * L >= maxExponent 2164 2165 and definitely underflows to zero where 2166 2167 (exp + 1) * L <= minExponent - precision 2168 2169 With integer arithmetic the tightest bounds for L are 2170 2171 93/28 < L < 196/59 [ numerator <= 256 ] 2172 42039/12655 < L < 28738/8651 [ numerator <= 65536 ] 2173 */ 2174 2175 if (decDigitValue(*D.firstSigDigit) >= 10U) { 2176 category = fcZero; 2177 fs = opOK; 2178 } else if ((D.normalizedExponent + 1) * 28738 2179 <= 8651 * (semantics->minExponent - (int) semantics->precision)) { 2180 /* Underflow to zero and round. */ 2181 zeroSignificand(); 2182 fs = normalize(rounding_mode, lfLessThanHalf); 2183 } else if ((D.normalizedExponent - 1) * 42039 2184 >= 12655 * semantics->maxExponent) { 2185 /* Overflow and round. */ 2186 fs = handleOverflow(rounding_mode); 2187 } else { 2188 integerPart *decSignificand; 2189 unsigned int partCount; 2190 2191 /* A tight upper bound on number of bits required to hold an 2192 N-digit decimal integer is N * 196 / 59. Allocate enough space 2193 to hold the full significand, and an extra part required by 2194 tcMultiplyPart. */ 2195 partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1; 2196 partCount = partCountForBits(1 + 196 * partCount / 59); 2197 decSignificand = new integerPart[partCount + 1]; 2198 partCount = 0; 2199 2200 /* Convert to binary efficiently - we do almost all multiplication 2201 in an integerPart. When this would overflow do we do a single 2202 bignum multiplication, and then revert again to multiplication 2203 in an integerPart. */ 2204 do { 2205 integerPart decValue, val, multiplier; 2206 2207 val = 0; 2208 multiplier = 1; 2209 2210 do { 2211 if (*p == '.') 2212 p++; 2213 2214 decValue = decDigitValue(*p++); 2215 multiplier *= 10; 2216 val = val * 10 + decValue; 2217 /* The maximum number that can be multiplied by ten with any 2218 digit added without overflowing an integerPart. */ 2219 } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10); 2220 2221 /* Multiply out the current part. */ 2222 APInt::tcMultiplyPart(decSignificand, decSignificand, multiplier, val, 2223 partCount, partCount + 1, false); 2224 2225 /* If we used another part (likely but not guaranteed), increase 2226 the count. */ 2227 if (decSignificand[partCount]) 2228 partCount++; 2229 } while (p <= D.lastSigDigit); 2230 2231 category = fcNormal; 2232 fs = roundSignificandWithExponent(decSignificand, partCount, 2233 D.exponent, rounding_mode); 2234 2235 delete [] decSignificand; 2236 } 2237 2238 return fs; 2239} 2240 2241APFloat::opStatus 2242APFloat::convertFromString(const char *p, roundingMode rounding_mode) 2243{ 2244 assertArithmeticOK(*semantics); 2245 2246 /* Handle a leading minus sign. */ 2247 if(*p == '-') 2248 sign = 1, p++; 2249 else 2250 sign = 0; 2251 2252 if(p[0] == '0' && (p[1] == 'x' || p[1] == 'X')) 2253 return convertFromHexadecimalString(p + 2, rounding_mode); 2254 else 2255 return convertFromDecimalString(p, rounding_mode); 2256} 2257 2258/* Write out a hexadecimal representation of the floating point value 2259 to DST, which must be of sufficient size, in the C99 form 2260 [-]0xh.hhhhp[+-]d. Return the number of characters written, 2261 excluding the terminating NUL. 2262 2263 If UPPERCASE, the output is in upper case, otherwise in lower case. 2264 2265 HEXDIGITS digits appear altogether, rounding the value if 2266 necessary. If HEXDIGITS is 0, the minimal precision to display the 2267 number precisely is used instead. If nothing would appear after 2268 the decimal point it is suppressed. 2269 2270 The decimal exponent is always printed and has at least one digit. 2271 Zero values display an exponent of zero. Infinities and NaNs 2272 appear as "infinity" or "nan" respectively. 2273 2274 The above rules are as specified by C99. There is ambiguity about 2275 what the leading hexadecimal digit should be. This implementation 2276 uses whatever is necessary so that the exponent is displayed as 2277 stored. This implies the exponent will fall within the IEEE format 2278 range, and the leading hexadecimal digit will be 0 (for denormals), 2279 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with 2280 any other digits zero). 2281*/ 2282unsigned int 2283APFloat::convertToHexString(char *dst, unsigned int hexDigits, 2284 bool upperCase, roundingMode rounding_mode) const 2285{ 2286 char *p; 2287 2288 assertArithmeticOK(*semantics); 2289 2290 p = dst; 2291 if (sign) 2292 *dst++ = '-'; 2293 2294 switch (category) { 2295 case fcInfinity: 2296 memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1); 2297 dst += sizeof infinityL - 1; 2298 break; 2299 2300 case fcNaN: 2301 memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1); 2302 dst += sizeof NaNU - 1; 2303 break; 2304 2305 case fcZero: 2306 *dst++ = '0'; 2307 *dst++ = upperCase ? 'X': 'x'; 2308 *dst++ = '0'; 2309 if (hexDigits > 1) { 2310 *dst++ = '.'; 2311 memset (dst, '0', hexDigits - 1); 2312 dst += hexDigits - 1; 2313 } 2314 *dst++ = upperCase ? 'P': 'p'; 2315 *dst++ = '0'; 2316 break; 2317 2318 case fcNormal: 2319 dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode); 2320 break; 2321 } 2322 2323 *dst = 0; 2324 2325 return static_cast<unsigned int>(dst - p); 2326} 2327 2328/* Does the hard work of outputting the correctly rounded hexadecimal 2329 form of a normal floating point number with the specified number of 2330 hexadecimal digits. If HEXDIGITS is zero the minimum number of 2331 digits necessary to print the value precisely is output. */ 2332char * 2333APFloat::convertNormalToHexString(char *dst, unsigned int hexDigits, 2334 bool upperCase, 2335 roundingMode rounding_mode) const 2336{ 2337 unsigned int count, valueBits, shift, partsCount, outputDigits; 2338 const char *hexDigitChars; 2339 const integerPart *significand; 2340 char *p; 2341 bool roundUp; 2342 2343 *dst++ = '0'; 2344 *dst++ = upperCase ? 'X': 'x'; 2345 2346 roundUp = false; 2347 hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower; 2348 2349 significand = significandParts(); 2350 partsCount = partCount(); 2351 2352 /* +3 because the first digit only uses the single integer bit, so 2353 we have 3 virtual zero most-significant-bits. */ 2354 valueBits = semantics->precision + 3; 2355 shift = integerPartWidth - valueBits % integerPartWidth; 2356 2357 /* The natural number of digits required ignoring trailing 2358 insignificant zeroes. */ 2359 outputDigits = (valueBits - significandLSB () + 3) / 4; 2360 2361 /* hexDigits of zero means use the required number for the 2362 precision. Otherwise, see if we are truncating. If we are, 2363 find out if we need to round away from zero. */ 2364 if (hexDigits) { 2365 if (hexDigits < outputDigits) { 2366 /* We are dropping non-zero bits, so need to check how to round. 2367 "bits" is the number of dropped bits. */ 2368 unsigned int bits; 2369 lostFraction fraction; 2370 2371 bits = valueBits - hexDigits * 4; 2372 fraction = lostFractionThroughTruncation (significand, partsCount, bits); 2373 roundUp = roundAwayFromZero(rounding_mode, fraction, bits); 2374 } 2375 outputDigits = hexDigits; 2376 } 2377 2378 /* Write the digits consecutively, and start writing in the location 2379 of the hexadecimal point. We move the most significant digit 2380 left and add the hexadecimal point later. */ 2381 p = ++dst; 2382 2383 count = (valueBits + integerPartWidth - 1) / integerPartWidth; 2384 2385 while (outputDigits && count) { 2386 integerPart part; 2387 2388 /* Put the most significant integerPartWidth bits in "part". */ 2389 if (--count == partsCount) 2390 part = 0; /* An imaginary higher zero part. */ 2391 else 2392 part = significand[count] << shift; 2393 2394 if (count && shift) 2395 part |= significand[count - 1] >> (integerPartWidth - shift); 2396 2397 /* Convert as much of "part" to hexdigits as we can. */ 2398 unsigned int curDigits = integerPartWidth / 4; 2399 2400 if (curDigits > outputDigits) 2401 curDigits = outputDigits; 2402 dst += partAsHex (dst, part, curDigits, hexDigitChars); 2403 outputDigits -= curDigits; 2404 } 2405 2406 if (roundUp) { 2407 char *q = dst; 2408 2409 /* Note that hexDigitChars has a trailing '0'. */ 2410 do { 2411 q--; 2412 *q = hexDigitChars[hexDigitValue (*q) + 1]; 2413 } while (*q == '0'); 2414 assert (q >= p); 2415 } else { 2416 /* Add trailing zeroes. */ 2417 memset (dst, '0', outputDigits); 2418 dst += outputDigits; 2419 } 2420 2421 /* Move the most significant digit to before the point, and if there 2422 is something after the decimal point add it. This must come 2423 after rounding above. */ 2424 p[-1] = p[0]; 2425 if (dst -1 == p) 2426 dst--; 2427 else 2428 p[0] = '.'; 2429 2430 /* Finally output the exponent. */ 2431 *dst++ = upperCase ? 'P': 'p'; 2432 2433 return writeSignedDecimal (dst, exponent); 2434} 2435 2436// For good performance it is desirable for different APFloats 2437// to produce different integers. 2438uint32_t 2439APFloat::getHashValue() const 2440{ 2441 if (category==fcZero) return sign<<8 | semantics->precision ; 2442 else if (category==fcInfinity) return sign<<9 | semantics->precision; 2443 else if (category==fcNaN) return 1<<10 | semantics->precision; 2444 else { 2445 uint32_t hash = sign<<11 | semantics->precision | exponent<<12; 2446 const integerPart* p = significandParts(); 2447 for (int i=partCount(); i>0; i--, p++) 2448 hash ^= ((uint32_t)*p) ^ (uint32_t)((*p)>>32); 2449 return hash; 2450 } 2451} 2452 2453// Conversion from APFloat to/from host float/double. It may eventually be 2454// possible to eliminate these and have everybody deal with APFloats, but that 2455// will take a while. This approach will not easily extend to long double. 2456// Current implementation requires integerPartWidth==64, which is correct at 2457// the moment but could be made more general. 2458 2459// Denormals have exponent minExponent in APFloat, but minExponent-1 in 2460// the actual IEEE respresentations. We compensate for that here. 2461 2462APInt 2463APFloat::convertF80LongDoubleAPFloatToAPInt() const 2464{ 2465 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended); 2466 assert (partCount()==2); 2467 2468 uint64_t myexponent, mysignificand; 2469 2470 if (category==fcNormal) { 2471 myexponent = exponent+16383; //bias 2472 mysignificand = significandParts()[0]; 2473 if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL)) 2474 myexponent = 0; // denormal 2475 } else if (category==fcZero) { 2476 myexponent = 0; 2477 mysignificand = 0; 2478 } else if (category==fcInfinity) { 2479 myexponent = 0x7fff; 2480 mysignificand = 0x8000000000000000ULL; 2481 } else { 2482 assert(category == fcNaN && "Unknown category"); 2483 myexponent = 0x7fff; 2484 mysignificand = significandParts()[0]; 2485 } 2486 2487 uint64_t words[2]; 2488 words[0] = ((uint64_t)(sign & 1) << 63) | 2489 ((myexponent & 0x7fffLL) << 48) | 2490 ((mysignificand >>16) & 0xffffffffffffLL); 2491 words[1] = mysignificand & 0xffff; 2492 return APInt(80, 2, words); 2493} 2494 2495APInt 2496APFloat::convertPPCDoubleDoubleAPFloatToAPInt() const 2497{ 2498 assert(semantics == (const llvm::fltSemantics*)&PPCDoubleDouble); 2499 assert (partCount()==2); 2500 2501 uint64_t myexponent, mysignificand, myexponent2, mysignificand2; 2502 2503 if (category==fcNormal) { 2504 myexponent = exponent + 1023; //bias 2505 myexponent2 = exponent2 + 1023; 2506 mysignificand = significandParts()[0]; 2507 mysignificand2 = significandParts()[1]; 2508 if (myexponent==1 && !(mysignificand & 0x10000000000000LL)) 2509 myexponent = 0; // denormal 2510 if (myexponent2==1 && !(mysignificand2 & 0x10000000000000LL)) 2511 myexponent2 = 0; // denormal 2512 } else if (category==fcZero) { 2513 myexponent = 0; 2514 mysignificand = 0; 2515 myexponent2 = 0; 2516 mysignificand2 = 0; 2517 } else if (category==fcInfinity) { 2518 myexponent = 0x7ff; 2519 myexponent2 = 0; 2520 mysignificand = 0; 2521 mysignificand2 = 0; 2522 } else { 2523 assert(category == fcNaN && "Unknown category"); 2524 myexponent = 0x7ff; 2525 mysignificand = significandParts()[0]; 2526 myexponent2 = exponent2; 2527 mysignificand2 = significandParts()[1]; 2528 } 2529 2530 uint64_t words[2]; 2531 words[0] = ((uint64_t)(sign & 1) << 63) | 2532 ((myexponent & 0x7ff) << 52) | 2533 (mysignificand & 0xfffffffffffffLL); 2534 words[1] = ((uint64_t)(sign2 & 1) << 63) | 2535 ((myexponent2 & 0x7ff) << 52) | 2536 (mysignificand2 & 0xfffffffffffffLL); 2537 return APInt(128, 2, words); 2538} 2539 2540APInt 2541APFloat::convertDoubleAPFloatToAPInt() const 2542{ 2543 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble); 2544 assert (partCount()==1); 2545 2546 uint64_t myexponent, mysignificand; 2547 2548 if (category==fcNormal) { 2549 myexponent = exponent+1023; //bias 2550 mysignificand = *significandParts(); 2551 if (myexponent==1 && !(mysignificand & 0x10000000000000LL)) 2552 myexponent = 0; // denormal 2553 } else if (category==fcZero) { 2554 myexponent = 0; 2555 mysignificand = 0; 2556 } else if (category==fcInfinity) { 2557 myexponent = 0x7ff; 2558 mysignificand = 0; 2559 } else { 2560 assert(category == fcNaN && "Unknown category!"); 2561 myexponent = 0x7ff; 2562 mysignificand = *significandParts(); 2563 } 2564 2565 return APInt(64, ((((uint64_t)(sign & 1) << 63) | 2566 ((myexponent & 0x7ff) << 52) | 2567 (mysignificand & 0xfffffffffffffLL)))); 2568} 2569 2570APInt 2571APFloat::convertFloatAPFloatToAPInt() const 2572{ 2573 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle); 2574 assert (partCount()==1); 2575 2576 uint32_t myexponent, mysignificand; 2577 2578 if (category==fcNormal) { 2579 myexponent = exponent+127; //bias 2580 mysignificand = (uint32_t)*significandParts(); 2581 if (myexponent == 1 && !(mysignificand & 0x800000)) 2582 myexponent = 0; // denormal 2583 } else if (category==fcZero) { 2584 myexponent = 0; 2585 mysignificand = 0; 2586 } else if (category==fcInfinity) { 2587 myexponent = 0xff; 2588 mysignificand = 0; 2589 } else { 2590 assert(category == fcNaN && "Unknown category!"); 2591 myexponent = 0xff; 2592 mysignificand = (uint32_t)*significandParts(); 2593 } 2594 2595 return APInt(32, (((sign&1) << 31) | ((myexponent&0xff) << 23) | 2596 (mysignificand & 0x7fffff))); 2597} 2598 2599// This function creates an APInt that is just a bit map of the floating 2600// point constant as it would appear in memory. It is not a conversion, 2601// and treating the result as a normal integer is unlikely to be useful. 2602 2603APInt 2604APFloat::convertToAPInt() const 2605{ 2606 if (semantics == (const llvm::fltSemantics*)&IEEEsingle) 2607 return convertFloatAPFloatToAPInt(); 2608 2609 if (semantics == (const llvm::fltSemantics*)&IEEEdouble) 2610 return convertDoubleAPFloatToAPInt(); 2611 2612 if (semantics == (const llvm::fltSemantics*)&PPCDoubleDouble) 2613 return convertPPCDoubleDoubleAPFloatToAPInt(); 2614 2615 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended && 2616 "unknown format!"); 2617 return convertF80LongDoubleAPFloatToAPInt(); 2618} 2619 2620float 2621APFloat::convertToFloat() const 2622{ 2623 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle); 2624 APInt api = convertToAPInt(); 2625 return api.bitsToFloat(); 2626} 2627 2628double 2629APFloat::convertToDouble() const 2630{ 2631 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble); 2632 APInt api = convertToAPInt(); 2633 return api.bitsToDouble(); 2634} 2635 2636/// Integer bit is explicit in this format. Current Intel book does not 2637/// define meaning of: 2638/// exponent = all 1's, integer bit not set. 2639/// exponent = 0, integer bit set. (formerly "psuedodenormals") 2640/// exponent!=0 nor all 1's, integer bit not set. (formerly "unnormals") 2641void 2642APFloat::initFromF80LongDoubleAPInt(const APInt &api) 2643{ 2644 assert(api.getBitWidth()==80); 2645 uint64_t i1 = api.getRawData()[0]; 2646 uint64_t i2 = api.getRawData()[1]; 2647 uint64_t myexponent = (i1 >> 48) & 0x7fff; 2648 uint64_t mysignificand = ((i1 << 16) & 0xffffffffffff0000ULL) | 2649 (i2 & 0xffff); 2650 2651 initialize(&APFloat::x87DoubleExtended); 2652 assert(partCount()==2); 2653 2654 sign = static_cast<unsigned int>(i1>>63); 2655 if (myexponent==0 && mysignificand==0) { 2656 // exponent, significand meaningless 2657 category = fcZero; 2658 } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) { 2659 // exponent, significand meaningless 2660 category = fcInfinity; 2661 } else if (myexponent==0x7fff && mysignificand!=0x8000000000000000ULL) { 2662 // exponent meaningless 2663 category = fcNaN; 2664 significandParts()[0] = mysignificand; 2665 significandParts()[1] = 0; 2666 } else { 2667 category = fcNormal; 2668 exponent = myexponent - 16383; 2669 significandParts()[0] = mysignificand; 2670 significandParts()[1] = 0; 2671 if (myexponent==0) // denormal 2672 exponent = -16382; 2673 } 2674} 2675 2676void 2677APFloat::initFromPPCDoubleDoubleAPInt(const APInt &api) 2678{ 2679 assert(api.getBitWidth()==128); 2680 uint64_t i1 = api.getRawData()[0]; 2681 uint64_t i2 = api.getRawData()[1]; 2682 uint64_t myexponent = (i1 >> 52) & 0x7ff; 2683 uint64_t mysignificand = i1 & 0xfffffffffffffLL; 2684 uint64_t myexponent2 = (i2 >> 52) & 0x7ff; 2685 uint64_t mysignificand2 = i2 & 0xfffffffffffffLL; 2686 2687 initialize(&APFloat::PPCDoubleDouble); 2688 assert(partCount()==2); 2689 2690 sign = static_cast<unsigned int>(i1>>63); 2691 sign2 = static_cast<unsigned int>(i2>>63); 2692 if (myexponent==0 && mysignificand==0) { 2693 // exponent, significand meaningless 2694 // exponent2 and significand2 are required to be 0; we don't check 2695 category = fcZero; 2696 } else if (myexponent==0x7ff && mysignificand==0) { 2697 // exponent, significand meaningless 2698 // exponent2 and significand2 are required to be 0; we don't check 2699 category = fcInfinity; 2700 } else if (myexponent==0x7ff && mysignificand!=0) { 2701 // exponent meaningless. So is the whole second word, but keep it 2702 // for determinism. 2703 category = fcNaN; 2704 exponent2 = myexponent2; 2705 significandParts()[0] = mysignificand; 2706 significandParts()[1] = mysignificand2; 2707 } else { 2708 category = fcNormal; 2709 // Note there is no category2; the second word is treated as if it is 2710 // fcNormal, although it might be something else considered by itself. 2711 exponent = myexponent - 1023; 2712 exponent2 = myexponent2 - 1023; 2713 significandParts()[0] = mysignificand; 2714 significandParts()[1] = mysignificand2; 2715 if (myexponent==0) // denormal 2716 exponent = -1022; 2717 else 2718 significandParts()[0] |= 0x10000000000000LL; // integer bit 2719 if (myexponent2==0) 2720 exponent2 = -1022; 2721 else 2722 significandParts()[1] |= 0x10000000000000LL; // integer bit 2723 } 2724} 2725 2726void 2727APFloat::initFromDoubleAPInt(const APInt &api) 2728{ 2729 assert(api.getBitWidth()==64); 2730 uint64_t i = *api.getRawData(); 2731 uint64_t myexponent = (i >> 52) & 0x7ff; 2732 uint64_t mysignificand = i & 0xfffffffffffffLL; 2733 2734 initialize(&APFloat::IEEEdouble); 2735 assert(partCount()==1); 2736 2737 sign = static_cast<unsigned int>(i>>63); 2738 if (myexponent==0 && mysignificand==0) { 2739 // exponent, significand meaningless 2740 category = fcZero; 2741 } else if (myexponent==0x7ff && mysignificand==0) { 2742 // exponent, significand meaningless 2743 category = fcInfinity; 2744 } else if (myexponent==0x7ff && mysignificand!=0) { 2745 // exponent meaningless 2746 category = fcNaN; 2747 *significandParts() = mysignificand; 2748 } else { 2749 category = fcNormal; 2750 exponent = myexponent - 1023; 2751 *significandParts() = mysignificand; 2752 if (myexponent==0) // denormal 2753 exponent = -1022; 2754 else 2755 *significandParts() |= 0x10000000000000LL; // integer bit 2756 } 2757} 2758 2759void 2760APFloat::initFromFloatAPInt(const APInt & api) 2761{ 2762 assert(api.getBitWidth()==32); 2763 uint32_t i = (uint32_t)*api.getRawData(); 2764 uint32_t myexponent = (i >> 23) & 0xff; 2765 uint32_t mysignificand = i & 0x7fffff; 2766 2767 initialize(&APFloat::IEEEsingle); 2768 assert(partCount()==1); 2769 2770 sign = i >> 31; 2771 if (myexponent==0 && mysignificand==0) { 2772 // exponent, significand meaningless 2773 category = fcZero; 2774 } else if (myexponent==0xff && mysignificand==0) { 2775 // exponent, significand meaningless 2776 category = fcInfinity; 2777 } else if (myexponent==0xff && mysignificand!=0) { 2778 // sign, exponent, significand meaningless 2779 category = fcNaN; 2780 *significandParts() = mysignificand; 2781 } else { 2782 category = fcNormal; 2783 exponent = myexponent - 127; //bias 2784 *significandParts() = mysignificand; 2785 if (myexponent==0) // denormal 2786 exponent = -126; 2787 else 2788 *significandParts() |= 0x800000; // integer bit 2789 } 2790} 2791 2792/// Treat api as containing the bits of a floating point number. Currently 2793/// we infer the floating point type from the size of the APInt. The 2794/// isIEEE argument distinguishes between PPC128 and IEEE128 (not meaningful 2795/// when the size is anything else). 2796void 2797APFloat::initFromAPInt(const APInt& api, bool isIEEE) 2798{ 2799 if (api.getBitWidth() == 32) 2800 return initFromFloatAPInt(api); 2801 else if (api.getBitWidth()==64) 2802 return initFromDoubleAPInt(api); 2803 else if (api.getBitWidth()==80) 2804 return initFromF80LongDoubleAPInt(api); 2805 else if (api.getBitWidth()==128 && !isIEEE) 2806 return initFromPPCDoubleDoubleAPInt(api); 2807 else 2808 assert(0); 2809} 2810 2811APFloat::APFloat(const APInt& api, bool isIEEE) 2812{ 2813 initFromAPInt(api, isIEEE); 2814} 2815 2816APFloat::APFloat(float f) 2817{ 2818 APInt api = APInt(32, 0); 2819 initFromAPInt(api.floatToBits(f)); 2820} 2821 2822APFloat::APFloat(double d) 2823{ 2824 APInt api = APInt(64, 0); 2825 initFromAPInt(api.doubleToBits(d)); 2826} 2827