1 //===-- APFloat.cpp - Implement APFloat class -----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a class to represent arbitrary precision floating
11 // point values and provide a variety of arithmetic operations on them.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/ADT/APSInt.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/FoldingSet.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/MathExtras.h"
26 #define convolve(lhs, rhs) ((lhs) * 4 + (rhs))
28 /* Assumed in hexadecimal significand parsing, and conversion to
29 hexadecimal strings. */
30 #define COMPILE_TIME_ASSERT(cond) extern int CTAssert[(cond) ? 1 : -1]
31 COMPILE_TIME_ASSERT(integerPartWidth % 4 == 0);
35 /* Represents floating point arithmetic semantics. */
37 /* The largest E such that 2^E is representable; this matches the
38 definition of IEEE 754. */
39 exponent_t maxExponent;
41 /* The smallest E such that 2^E is a normalized number; this
42 matches the definition of IEEE 754. */
43 exponent_t minExponent;
45 /* Number of bits in the significand. This includes the integer
47 unsigned int precision;
49 /* True if arithmetic is supported. */
50 unsigned int arithmeticOK;
53 const fltSemantics APFloat::IEEEhalf = { 15, -14, 11, true };
54 const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, true };
55 const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, true };
56 const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, true };
57 const fltSemantics APFloat::x87DoubleExtended = { 16383, -16382, 64, true };
58 const fltSemantics APFloat::Bogus = { 0, 0, 0, true };
60 // The PowerPC format consists of two doubles. It does not map cleanly
61 // onto the usual format above. For now only storage of constants of
62 // this type is supported, no arithmetic.
63 const fltSemantics APFloat::PPCDoubleDouble = { 1023, -1022, 106, false };
65 /* A tight upper bound on number of parts required to hold the value
68 power * 815 / (351 * integerPartWidth) + 1
70 However, whilst the result may require only this many parts,
71 because we are multiplying two values to get it, the
72 multiplication may require an extra part with the excess part
73 being zero (consider the trivial case of 1 * 1, tcFullMultiply
74 requires two parts to hold the single-part result). So we add an
75 extra one to guarantee enough space whilst multiplying. */
76 const unsigned int maxExponent = 16383;
77 const unsigned int maxPrecision = 113;
78 const unsigned int maxPowerOfFiveExponent = maxExponent + maxPrecision - 1;
79 const unsigned int maxPowerOfFiveParts = 2 + ((maxPowerOfFiveExponent * 815)
80 / (351 * integerPartWidth));
83 /* A bunch of private, handy routines. */
85 static inline unsigned int
86 partCountForBits(unsigned int bits)
88 return ((bits) + integerPartWidth - 1) / integerPartWidth;
91 /* Returns 0U-9U. Return values >= 10U are not digits. */
92 static inline unsigned int
93 decDigitValue(unsigned int c)
99 hexDigitValue(unsigned int c)
119 assertArithmeticOK(const llvm::fltSemantics &semantics) {
120 assert(semantics.arithmeticOK &&
121 "Compile-time arithmetic does not support these semantics");
124 /* Return the value of a decimal exponent of the form
127 If the exponent overflows, returns a large exponent with the
130 readExponent(StringRef::iterator begin, StringRef::iterator end)
133 unsigned int absExponent;
134 const unsigned int overlargeExponent = 24000; /* FIXME. */
135 StringRef::iterator p = begin;
137 assert(p != end && "Exponent has no digits");
139 isNegative = (*p == '-');
140 if (*p == '-' || *p == '+') {
142 assert(p != end && "Exponent has no digits");
145 absExponent = decDigitValue(*p++);
146 assert(absExponent < 10U && "Invalid character in exponent");
148 for (; p != end; ++p) {
151 value = decDigitValue(*p);
152 assert(value < 10U && "Invalid character in exponent");
154 value += absExponent * 10;
155 if (absExponent >= overlargeExponent) {
156 absExponent = overlargeExponent;
157 p = end; /* outwit assert below */
163 assert(p == end && "Invalid exponent in exponent");
166 return -(int) absExponent;
168 return (int) absExponent;
171 /* This is ugly and needs cleaning up, but I don't immediately see
172 how whilst remaining safe. */
174 totalExponent(StringRef::iterator p, StringRef::iterator end,
175 int exponentAdjustment)
177 int unsignedExponent;
178 bool negative, overflow;
181 assert(p != end && "Exponent has no digits");
183 negative = *p == '-';
184 if (*p == '-' || *p == '+') {
186 assert(p != end && "Exponent has no digits");
189 unsignedExponent = 0;
191 for (; p != end; ++p) {
194 value = decDigitValue(*p);
195 assert(value < 10U && "Invalid character in exponent");
197 unsignedExponent = unsignedExponent * 10 + value;
198 if (unsignedExponent > 32767)
202 if (exponentAdjustment > 32767 || exponentAdjustment < -32768)
206 exponent = unsignedExponent;
208 exponent = -exponent;
209 exponent += exponentAdjustment;
210 if (exponent > 32767 || exponent < -32768)
215 exponent = negative ? -32768: 32767;
220 static StringRef::iterator
221 skipLeadingZeroesAndAnyDot(StringRef::iterator begin, StringRef::iterator end,
222 StringRef::iterator *dot)
224 StringRef::iterator p = begin;
226 while (*p == '0' && p != end)
232 assert(end - begin != 1 && "Significand has no digits");
234 while (*p == '0' && p != end)
241 /* Given a normal decimal floating point number of the form
245 where the decimal point and exponent are optional, fill out the
246 structure D. Exponent is appropriate if the significand is
247 treated as an integer, and normalizedExponent if the significand
248 is taken to have the decimal point after a single leading
251 If the value is zero, V->firstSigDigit points to a non-digit, and
252 the return exponent is zero.
255 const char *firstSigDigit;
256 const char *lastSigDigit;
258 int normalizedExponent;
262 interpretDecimal(StringRef::iterator begin, StringRef::iterator end,
265 StringRef::iterator dot = end;
266 StringRef::iterator p = skipLeadingZeroesAndAnyDot (begin, end, &dot);
268 D->firstSigDigit = p;
270 D->normalizedExponent = 0;
272 for (; p != end; ++p) {
274 assert(dot == end && "String contains multiple dots");
279 if (decDigitValue(*p) >= 10U)
284 assert((*p == 'e' || *p == 'E') && "Invalid character in significand");
285 assert(p != begin && "Significand has no digits");
286 assert((dot == end || p - begin != 1) && "Significand has no digits");
288 /* p points to the first non-digit in the string */
289 D->exponent = readExponent(p + 1, end);
291 /* Implied decimal point? */
296 /* If number is all zeroes accept any exponent. */
297 if (p != D->firstSigDigit) {
298 /* Drop insignificant trailing zeroes. */
303 while (p != begin && *p == '0');
304 while (p != begin && *p == '.');
307 /* Adjust the exponents for any decimal point. */
308 D->exponent += static_cast<exponent_t>((dot - p) - (dot > p));
309 D->normalizedExponent = (D->exponent +
310 static_cast<exponent_t>((p - D->firstSigDigit)
311 - (dot > D->firstSigDigit && dot < p)));
317 /* Return the trailing fraction of a hexadecimal number.
318 DIGITVALUE is the first hex digit of the fraction, P points to
321 trailingHexadecimalFraction(StringRef::iterator p, StringRef::iterator end,
322 unsigned int digitValue)
324 unsigned int hexDigit;
326 /* If the first trailing digit isn't 0 or 8 we can work out the
327 fraction immediately. */
329 return lfMoreThanHalf;
330 else if (digitValue < 8 && digitValue > 0)
331 return lfLessThanHalf;
333 /* Otherwise we need to find the first non-zero digit. */
337 assert(p != end && "Invalid trailing hexadecimal fraction!");
339 hexDigit = hexDigitValue(*p);
341 /* If we ran off the end it is exactly zero or one-half, otherwise
344 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf;
346 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf;
349 /* Return the fraction lost were a bignum truncated losing the least
350 significant BITS bits. */
352 lostFractionThroughTruncation(const integerPart *parts,
353 unsigned int partCount,
358 lsb = APInt::tcLSB(parts, partCount);
360 /* Note this is guaranteed true if bits == 0, or LSB == -1U. */
362 return lfExactlyZero;
364 return lfExactlyHalf;
365 if (bits <= partCount * integerPartWidth &&
366 APInt::tcExtractBit(parts, bits - 1))
367 return lfMoreThanHalf;
369 return lfLessThanHalf;
372 /* Shift DST right BITS bits noting lost fraction. */
374 shiftRight(integerPart *dst, unsigned int parts, unsigned int bits)
376 lostFraction lost_fraction;
378 lost_fraction = lostFractionThroughTruncation(dst, parts, bits);
380 APInt::tcShiftRight(dst, parts, bits);
382 return lost_fraction;
385 /* Combine the effect of two lost fractions. */
387 combineLostFractions(lostFraction moreSignificant,
388 lostFraction lessSignificant)
390 if (lessSignificant != lfExactlyZero) {
391 if (moreSignificant == lfExactlyZero)
392 moreSignificant = lfLessThanHalf;
393 else if (moreSignificant == lfExactlyHalf)
394 moreSignificant = lfMoreThanHalf;
397 return moreSignificant;
400 /* The error from the true value, in half-ulps, on multiplying two
401 floating point numbers, which differ from the value they
402 approximate by at most HUE1 and HUE2 half-ulps, is strictly less
403 than the returned value.
405 See "How to Read Floating Point Numbers Accurately" by William D
408 HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2)
410 assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8));
412 if (HUerr1 + HUerr2 == 0)
413 return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */
415 return inexactMultiply + 2 * (HUerr1 + HUerr2);
418 /* The number of ulps from the boundary (zero, or half if ISNEAREST)
419 when the least significant BITS are truncated. BITS cannot be
422 ulpsFromBoundary(const integerPart *parts, unsigned int bits, bool isNearest)
424 unsigned int count, partBits;
425 integerPart part, boundary;
430 count = bits / integerPartWidth;
431 partBits = bits % integerPartWidth + 1;
433 part = parts[count] & (~(integerPart) 0 >> (integerPartWidth - partBits));
436 boundary = (integerPart) 1 << (partBits - 1);
441 if (part - boundary <= boundary - part)
442 return part - boundary;
444 return boundary - part;
447 if (part == boundary) {
450 return ~(integerPart) 0; /* A lot. */
453 } else if (part == boundary - 1) {
456 return ~(integerPart) 0; /* A lot. */
461 return ~(integerPart) 0; /* A lot. */
464 /* Place pow(5, power) in DST, and return the number of parts used.
465 DST must be at least one part larger than size of the answer. */
467 powerOf5(integerPart *dst, unsigned int power)
469 static const integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125,
471 integerPart pow5s[maxPowerOfFiveParts * 2 + 5];
472 pow5s[0] = 78125 * 5;
474 unsigned int partsCount[16] = { 1 };
475 integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5;
477 assert(power <= maxExponent);
482 *p1 = firstEightPowers[power & 7];
488 for (unsigned int n = 0; power; power >>= 1, n++) {
493 /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */
495 pc = partsCount[n - 1];
496 APInt::tcFullMultiply(pow5, pow5 - pc, pow5 - pc, pc, pc);
498 if (pow5[pc - 1] == 0)
506 APInt::tcFullMultiply(p2, p1, pow5, result, pc);
508 if (p2[result - 1] == 0)
511 /* Now result is in p1 with partsCount parts and p2 is scratch
513 tmp = p1, p1 = p2, p2 = tmp;
520 APInt::tcAssign(dst, p1, result);
525 /* Zero at the end to avoid modular arithmetic when adding one; used
526 when rounding up during hexadecimal output. */
527 static const char hexDigitsLower[] = "0123456789abcdef0";
528 static const char hexDigitsUpper[] = "0123456789ABCDEF0";
529 static const char infinityL[] = "infinity";
530 static const char infinityU[] = "INFINITY";
531 static const char NaNL[] = "nan";
532 static const char NaNU[] = "NAN";
534 /* Write out an integerPart in hexadecimal, starting with the most
535 significant nibble. Write out exactly COUNT hexdigits, return
538 partAsHex (char *dst, integerPart part, unsigned int count,
539 const char *hexDigitChars)
541 unsigned int result = count;
543 assert(count != 0 && count <= integerPartWidth / 4);
545 part >>= (integerPartWidth - 4 * count);
547 dst[count] = hexDigitChars[part & 0xf];
554 /* Write out an unsigned decimal integer. */
556 writeUnsignedDecimal (char *dst, unsigned int n)
572 /* Write out a signed decimal integer. */
574 writeSignedDecimal (char *dst, int value)
578 dst = writeUnsignedDecimal(dst, -(unsigned) value);
580 dst = writeUnsignedDecimal(dst, value);
587 APFloat::initialize(const fltSemantics *ourSemantics)
591 semantics = ourSemantics;
594 significand.parts = new integerPart[count];
598 APFloat::freeSignificand()
601 delete [] significand.parts;
605 APFloat::assign(const APFloat &rhs)
607 assert(semantics == rhs.semantics);
610 category = rhs.category;
611 exponent = rhs.exponent;
613 exponent2 = rhs.exponent2;
614 if (category == fcNormal || category == fcNaN)
615 copySignificand(rhs);
619 APFloat::copySignificand(const APFloat &rhs)
621 assert(category == fcNormal || category == fcNaN);
622 assert(rhs.partCount() >= partCount());
624 APInt::tcAssign(significandParts(), rhs.significandParts(),
628 /* Make this number a NaN, with an arbitrary but deterministic value
629 for the significand. If double or longer, this is a signalling NaN,
630 which may not be ideal. If float, this is QNaN(0). */
631 void APFloat::makeNaN(bool SNaN, bool Negative, const APInt *fill)
636 integerPart *significand = significandParts();
637 unsigned numParts = partCount();
639 // Set the significand bits to the fill.
640 if (!fill || fill->getNumWords() < numParts)
641 APInt::tcSet(significand, 0, numParts);
643 APInt::tcAssign(significand, fill->getRawData(),
644 std::min(fill->getNumWords(), numParts));
646 // Zero out the excess bits of the significand.
647 unsigned bitsToPreserve = semantics->precision - 1;
648 unsigned part = bitsToPreserve / 64;
649 bitsToPreserve %= 64;
650 significand[part] &= ((1ULL << bitsToPreserve) - 1);
651 for (part++; part != numParts; ++part)
652 significand[part] = 0;
655 unsigned QNaNBit = semantics->precision - 2;
658 // We always have to clear the QNaN bit to make it an SNaN.
659 APInt::tcClearBit(significand, QNaNBit);
661 // If there are no bits set in the payload, we have to set
662 // *something* to make it a NaN instead of an infinity;
663 // conventionally, this is the next bit down from the QNaN bit.
664 if (APInt::tcIsZero(significand, numParts))
665 APInt::tcSetBit(significand, QNaNBit - 1);
667 // We always have to set the QNaN bit to make it a QNaN.
668 APInt::tcSetBit(significand, QNaNBit);
671 // For x87 extended precision, we want to make a NaN, not a
672 // pseudo-NaN. Maybe we should expose the ability to make
674 if (semantics == &APFloat::x87DoubleExtended)
675 APInt::tcSetBit(significand, QNaNBit + 1);
678 APFloat APFloat::makeNaN(const fltSemantics &Sem, bool SNaN, bool Negative,
680 APFloat value(Sem, uninitialized);
681 value.makeNaN(SNaN, Negative, fill);
686 APFloat::operator=(const APFloat &rhs)
689 if (semantics != rhs.semantics) {
691 initialize(rhs.semantics);
700 APFloat::bitwiseIsEqual(const APFloat &rhs) const {
703 if (semantics != rhs.semantics ||
704 category != rhs.category ||
707 if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble &&
710 if (category==fcZero || category==fcInfinity)
712 else if (category==fcNormal && exponent!=rhs.exponent)
714 else if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble &&
715 exponent2!=rhs.exponent2)
719 const integerPart* p=significandParts();
720 const integerPart* q=rhs.significandParts();
721 for (; i>0; i--, p++, q++) {
729 APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value)
730 : exponent2(0), sign2(0) {
731 assertArithmeticOK(ourSemantics);
732 initialize(&ourSemantics);
735 exponent = ourSemantics.precision - 1;
736 significandParts()[0] = value;
737 normalize(rmNearestTiesToEven, lfExactlyZero);
740 APFloat::APFloat(const fltSemantics &ourSemantics) : exponent2(0), sign2(0) {
741 assertArithmeticOK(ourSemantics);
742 initialize(&ourSemantics);
747 APFloat::APFloat(const fltSemantics &ourSemantics, uninitializedTag tag)
748 : exponent2(0), sign2(0) {
749 assertArithmeticOK(ourSemantics);
750 // Allocates storage if necessary but does not initialize it.
751 initialize(&ourSemantics);
754 APFloat::APFloat(const fltSemantics &ourSemantics,
755 fltCategory ourCategory, bool negative)
756 : exponent2(0), sign2(0) {
757 assertArithmeticOK(ourSemantics);
758 initialize(&ourSemantics);
759 category = ourCategory;
761 if (category == fcNormal)
763 else if (ourCategory == fcNaN)
767 APFloat::APFloat(const fltSemantics &ourSemantics, StringRef text)
768 : exponent2(0), sign2(0) {
769 assertArithmeticOK(ourSemantics);
770 initialize(&ourSemantics);
771 convertFromString(text, rmNearestTiesToEven);
774 APFloat::APFloat(const APFloat &rhs) : exponent2(0), sign2(0) {
775 initialize(rhs.semantics);
784 // Profile - This method 'profiles' an APFloat for use with FoldingSet.
785 void APFloat::Profile(FoldingSetNodeID& ID) const {
786 ID.Add(bitcastToAPInt());
790 APFloat::partCount() const
792 return partCountForBits(semantics->precision + 1);
796 APFloat::semanticsPrecision(const fltSemantics &semantics)
798 return semantics.precision;
802 APFloat::significandParts() const
804 return const_cast<APFloat *>(this)->significandParts();
808 APFloat::significandParts()
810 assert(category == fcNormal || category == fcNaN);
813 return significand.parts;
815 return &significand.part;
819 APFloat::zeroSignificand()
822 APInt::tcSet(significandParts(), 0, partCount());
825 /* Increment an fcNormal floating point number's significand. */
827 APFloat::incrementSignificand()
831 carry = APInt::tcIncrement(significandParts(), partCount());
833 /* Our callers should never cause us to overflow. */
838 /* Add the significand of the RHS. Returns the carry flag. */
840 APFloat::addSignificand(const APFloat &rhs)
844 parts = significandParts();
846 assert(semantics == rhs.semantics);
847 assert(exponent == rhs.exponent);
849 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount());
852 /* Subtract the significand of the RHS with a borrow flag. Returns
855 APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow)
859 parts = significandParts();
861 assert(semantics == rhs.semantics);
862 assert(exponent == rhs.exponent);
864 return APInt::tcSubtract(parts, rhs.significandParts(), borrow,
868 /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it
869 on to the full-precision result of the multiplication. Returns the
872 APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend)
874 unsigned int omsb; // One, not zero, based MSB.
875 unsigned int partsCount, newPartsCount, precision;
876 integerPart *lhsSignificand;
877 integerPart scratch[4];
878 integerPart *fullSignificand;
879 lostFraction lost_fraction;
882 assert(semantics == rhs.semantics);
884 precision = semantics->precision;
885 newPartsCount = partCountForBits(precision * 2);
887 if (newPartsCount > 4)
888 fullSignificand = new integerPart[newPartsCount];
890 fullSignificand = scratch;
892 lhsSignificand = significandParts();
893 partsCount = partCount();
895 APInt::tcFullMultiply(fullSignificand, lhsSignificand,
896 rhs.significandParts(), partsCount, partsCount);
898 lost_fraction = lfExactlyZero;
899 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
900 exponent += rhs.exponent;
903 Significand savedSignificand = significand;
904 const fltSemantics *savedSemantics = semantics;
905 fltSemantics extendedSemantics;
907 unsigned int extendedPrecision;
909 /* Normalize our MSB. */
910 extendedPrecision = precision + precision - 1;
911 if (omsb != extendedPrecision) {
912 APInt::tcShiftLeft(fullSignificand, newPartsCount,
913 extendedPrecision - omsb);
914 exponent -= extendedPrecision - omsb;
917 /* Create new semantics. */
918 extendedSemantics = *semantics;
919 extendedSemantics.precision = extendedPrecision;
921 if (newPartsCount == 1)
922 significand.part = fullSignificand[0];
924 significand.parts = fullSignificand;
925 semantics = &extendedSemantics;
927 APFloat extendedAddend(*addend);
928 status = extendedAddend.convert(extendedSemantics, rmTowardZero, &ignored);
929 assert(status == opOK);
931 lost_fraction = addOrSubtractSignificand(extendedAddend, false);
933 /* Restore our state. */
934 if (newPartsCount == 1)
935 fullSignificand[0] = significand.part;
936 significand = savedSignificand;
937 semantics = savedSemantics;
939 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
942 exponent -= (precision - 1);
944 if (omsb > precision) {
945 unsigned int bits, significantParts;
948 bits = omsb - precision;
949 significantParts = partCountForBits(omsb);
950 lf = shiftRight(fullSignificand, significantParts, bits);
951 lost_fraction = combineLostFractions(lf, lost_fraction);
955 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount);
957 if (newPartsCount > 4)
958 delete [] fullSignificand;
960 return lost_fraction;
963 /* Multiply the significands of LHS and RHS to DST. */
965 APFloat::divideSignificand(const APFloat &rhs)
967 unsigned int bit, i, partsCount;
968 const integerPart *rhsSignificand;
969 integerPart *lhsSignificand, *dividend, *divisor;
970 integerPart scratch[4];
971 lostFraction lost_fraction;
973 assert(semantics == rhs.semantics);
975 lhsSignificand = significandParts();
976 rhsSignificand = rhs.significandParts();
977 partsCount = partCount();
980 dividend = new integerPart[partsCount * 2];
984 divisor = dividend + partsCount;
986 /* Copy the dividend and divisor as they will be modified in-place. */
987 for (i = 0; i < partsCount; i++) {
988 dividend[i] = lhsSignificand[i];
989 divisor[i] = rhsSignificand[i];
990 lhsSignificand[i] = 0;
993 exponent -= rhs.exponent;
995 unsigned int precision = semantics->precision;
997 /* Normalize the divisor. */
998 bit = precision - APInt::tcMSB(divisor, partsCount) - 1;
1001 APInt::tcShiftLeft(divisor, partsCount, bit);
1004 /* Normalize the dividend. */
1005 bit = precision - APInt::tcMSB(dividend, partsCount) - 1;
1008 APInt::tcShiftLeft(dividend, partsCount, bit);
1011 /* Ensure the dividend >= divisor initially for the loop below.
1012 Incidentally, this means that the division loop below is
1013 guaranteed to set the integer bit to one. */
1014 if (APInt::tcCompare(dividend, divisor, partsCount) < 0) {
1016 APInt::tcShiftLeft(dividend, partsCount, 1);
1017 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0);
1020 /* Long division. */
1021 for (bit = precision; bit; bit -= 1) {
1022 if (APInt::tcCompare(dividend, divisor, partsCount) >= 0) {
1023 APInt::tcSubtract(dividend, divisor, 0, partsCount);
1024 APInt::tcSetBit(lhsSignificand, bit - 1);
1027 APInt::tcShiftLeft(dividend, partsCount, 1);
1030 /* Figure out the lost fraction. */
1031 int cmp = APInt::tcCompare(dividend, divisor, partsCount);
1034 lost_fraction = lfMoreThanHalf;
1036 lost_fraction = lfExactlyHalf;
1037 else if (APInt::tcIsZero(dividend, partsCount))
1038 lost_fraction = lfExactlyZero;
1040 lost_fraction = lfLessThanHalf;
1045 return lost_fraction;
1049 APFloat::significandMSB() const
1051 return APInt::tcMSB(significandParts(), partCount());
1055 APFloat::significandLSB() const
1057 return APInt::tcLSB(significandParts(), partCount());
1060 /* Note that a zero result is NOT normalized to fcZero. */
1062 APFloat::shiftSignificandRight(unsigned int bits)
1064 /* Our exponent should not overflow. */
1065 assert((exponent_t) (exponent + bits) >= exponent);
1069 return shiftRight(significandParts(), partCount(), bits);
1072 /* Shift the significand left BITS bits, subtract BITS from its exponent. */
1074 APFloat::shiftSignificandLeft(unsigned int bits)
1076 assert(bits < semantics->precision);
1079 unsigned int partsCount = partCount();
1081 APInt::tcShiftLeft(significandParts(), partsCount, bits);
1084 assert(!APInt::tcIsZero(significandParts(), partsCount));
1089 APFloat::compareAbsoluteValue(const APFloat &rhs) const
1093 assert(semantics == rhs.semantics);
1094 assert(category == fcNormal);
1095 assert(rhs.category == fcNormal);
1097 compare = exponent - rhs.exponent;
1099 /* If exponents are equal, do an unsigned bignum comparison of the
1102 compare = APInt::tcCompare(significandParts(), rhs.significandParts(),
1106 return cmpGreaterThan;
1107 else if (compare < 0)
1113 /* Handle overflow. Sign is preserved. We either become infinity or
1114 the largest finite number. */
1116 APFloat::handleOverflow(roundingMode rounding_mode)
1119 if (rounding_mode == rmNearestTiesToEven ||
1120 rounding_mode == rmNearestTiesToAway ||
1121 (rounding_mode == rmTowardPositive && !sign) ||
1122 (rounding_mode == rmTowardNegative && sign)) {
1123 category = fcInfinity;
1124 return (opStatus) (opOverflow | opInexact);
1127 /* Otherwise we become the largest finite number. */
1128 category = fcNormal;
1129 exponent = semantics->maxExponent;
1130 APInt::tcSetLeastSignificantBits(significandParts(), partCount(),
1131 semantics->precision);
1136 /* Returns TRUE if, when truncating the current number, with BIT the
1137 new LSB, with the given lost fraction and rounding mode, the result
1138 would need to be rounded away from zero (i.e., by increasing the
1139 signficand). This routine must work for fcZero of both signs, and
1140 fcNormal numbers. */
1142 APFloat::roundAwayFromZero(roundingMode rounding_mode,
1143 lostFraction lost_fraction,
1144 unsigned int bit) const
1146 /* NaNs and infinities should not have lost fractions. */
1147 assert(category == fcNormal || category == fcZero);
1149 /* Current callers never pass this so we don't handle it. */
1150 assert(lost_fraction != lfExactlyZero);
1152 switch (rounding_mode) {
1153 case rmNearestTiesToAway:
1154 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf;
1156 case rmNearestTiesToEven:
1157 if (lost_fraction == lfMoreThanHalf)
1160 /* Our zeroes don't have a significand to test. */
1161 if (lost_fraction == lfExactlyHalf && category != fcZero)
1162 return APInt::tcExtractBit(significandParts(), bit);
1169 case rmTowardPositive:
1170 return sign == false;
1172 case rmTowardNegative:
1173 return sign == true;
1178 APFloat::normalize(roundingMode rounding_mode,
1179 lostFraction lost_fraction)
1181 unsigned int omsb; /* One, not zero, based MSB. */
1184 if (category != fcNormal)
1187 /* Before rounding normalize the exponent of fcNormal numbers. */
1188 omsb = significandMSB() + 1;
1191 /* OMSB is numbered from 1. We want to place it in the integer
1192 bit numbered PRECISION if possible, with a compensating change in
1194 exponentChange = omsb - semantics->precision;
1196 /* If the resulting exponent is too high, overflow according to
1197 the rounding mode. */
1198 if (exponent + exponentChange > semantics->maxExponent)
1199 return handleOverflow(rounding_mode);
1201 /* Subnormal numbers have exponent minExponent, and their MSB
1202 is forced based on that. */
1203 if (exponent + exponentChange < semantics->minExponent)
1204 exponentChange = semantics->minExponent - exponent;
1206 /* Shifting left is easy as we don't lose precision. */
1207 if (exponentChange < 0) {
1208 assert(lost_fraction == lfExactlyZero);
1210 shiftSignificandLeft(-exponentChange);
1215 if (exponentChange > 0) {
1218 /* Shift right and capture any new lost fraction. */
1219 lf = shiftSignificandRight(exponentChange);
1221 lost_fraction = combineLostFractions(lf, lost_fraction);
1223 /* Keep OMSB up-to-date. */
1224 if (omsb > (unsigned) exponentChange)
1225 omsb -= exponentChange;
1231 /* Now round the number according to rounding_mode given the lost
1234 /* As specified in IEEE 754, since we do not trap we do not report
1235 underflow for exact results. */
1236 if (lost_fraction == lfExactlyZero) {
1237 /* Canonicalize zeroes. */
1244 /* Increment the significand if we're rounding away from zero. */
1245 if (roundAwayFromZero(rounding_mode, lost_fraction, 0)) {
1247 exponent = semantics->minExponent;
1249 incrementSignificand();
1250 omsb = significandMSB() + 1;
1252 /* Did the significand increment overflow? */
1253 if (omsb == (unsigned) semantics->precision + 1) {
1254 /* Renormalize by incrementing the exponent and shifting our
1255 significand right one. However if we already have the
1256 maximum exponent we overflow to infinity. */
1257 if (exponent == semantics->maxExponent) {
1258 category = fcInfinity;
1260 return (opStatus) (opOverflow | opInexact);
1263 shiftSignificandRight(1);
1269 /* The normal case - we were and are not denormal, and any
1270 significand increment above didn't overflow. */
1271 if (omsb == semantics->precision)
1274 /* We have a non-zero denormal. */
1275 assert(omsb < semantics->precision);
1277 /* Canonicalize zeroes. */
1281 /* The fcZero case is a denormal that underflowed to zero. */
1282 return (opStatus) (opUnderflow | opInexact);
1286 APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract)
1288 switch (convolve(category, rhs.category)) {
1290 llvm_unreachable(0);
1292 case convolve(fcNaN, fcZero):
1293 case convolve(fcNaN, fcNormal):
1294 case convolve(fcNaN, fcInfinity):
1295 case convolve(fcNaN, fcNaN):
1296 case convolve(fcNormal, fcZero):
1297 case convolve(fcInfinity, fcNormal):
1298 case convolve(fcInfinity, fcZero):
1301 case convolve(fcZero, fcNaN):
1302 case convolve(fcNormal, fcNaN):
1303 case convolve(fcInfinity, fcNaN):
1305 copySignificand(rhs);
1308 case convolve(fcNormal, fcInfinity):
1309 case convolve(fcZero, fcInfinity):
1310 category = fcInfinity;
1311 sign = rhs.sign ^ subtract;
1314 case convolve(fcZero, fcNormal):
1316 sign = rhs.sign ^ subtract;
1319 case convolve(fcZero, fcZero):
1320 /* Sign depends on rounding mode; handled by caller. */
1323 case convolve(fcInfinity, fcInfinity):
1324 /* Differently signed infinities can only be validly
1326 if (((sign ^ rhs.sign)!=0) != subtract) {
1333 case convolve(fcNormal, fcNormal):
1338 /* Add or subtract two normal numbers. */
1340 APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract)
1343 lostFraction lost_fraction;
1346 /* Determine if the operation on the absolute values is effectively
1347 an addition or subtraction. */
1348 subtract ^= (sign ^ rhs.sign) ? true : false;
1350 /* Are we bigger exponent-wise than the RHS? */
1351 bits = exponent - rhs.exponent;
1353 /* Subtraction is more subtle than one might naively expect. */
1355 APFloat temp_rhs(rhs);
1359 reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan;
1360 lost_fraction = lfExactlyZero;
1361 } else if (bits > 0) {
1362 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1);
1363 shiftSignificandLeft(1);
1366 lost_fraction = shiftSignificandRight(-bits - 1);
1367 temp_rhs.shiftSignificandLeft(1);
1372 carry = temp_rhs.subtractSignificand
1373 (*this, lost_fraction != lfExactlyZero);
1374 copySignificand(temp_rhs);
1377 carry = subtractSignificand
1378 (temp_rhs, lost_fraction != lfExactlyZero);
1381 /* Invert the lost fraction - it was on the RHS and
1383 if (lost_fraction == lfLessThanHalf)
1384 lost_fraction = lfMoreThanHalf;
1385 else if (lost_fraction == lfMoreThanHalf)
1386 lost_fraction = lfLessThanHalf;
1388 /* The code above is intended to ensure that no borrow is
1394 APFloat temp_rhs(rhs);
1396 lost_fraction = temp_rhs.shiftSignificandRight(bits);
1397 carry = addSignificand(temp_rhs);
1399 lost_fraction = shiftSignificandRight(-bits);
1400 carry = addSignificand(rhs);
1403 /* We have a guard bit; generating a carry cannot happen. */
1408 return lost_fraction;
1412 APFloat::multiplySpecials(const APFloat &rhs)
1414 switch (convolve(category, rhs.category)) {
1416 llvm_unreachable(0);
1418 case convolve(fcNaN, fcZero):
1419 case convolve(fcNaN, fcNormal):
1420 case convolve(fcNaN, fcInfinity):
1421 case convolve(fcNaN, fcNaN):
1424 case convolve(fcZero, fcNaN):
1425 case convolve(fcNormal, fcNaN):
1426 case convolve(fcInfinity, fcNaN):
1428 copySignificand(rhs);
1431 case convolve(fcNormal, fcInfinity):
1432 case convolve(fcInfinity, fcNormal):
1433 case convolve(fcInfinity, fcInfinity):
1434 category = fcInfinity;
1437 case convolve(fcZero, fcNormal):
1438 case convolve(fcNormal, fcZero):
1439 case convolve(fcZero, fcZero):
1443 case convolve(fcZero, fcInfinity):
1444 case convolve(fcInfinity, fcZero):
1448 case convolve(fcNormal, fcNormal):
1454 APFloat::divideSpecials(const APFloat &rhs)
1456 switch (convolve(category, rhs.category)) {
1458 llvm_unreachable(0);
1460 case convolve(fcNaN, fcZero):
1461 case convolve(fcNaN, fcNormal):
1462 case convolve(fcNaN, fcInfinity):
1463 case convolve(fcNaN, fcNaN):
1464 case convolve(fcInfinity, fcZero):
1465 case convolve(fcInfinity, fcNormal):
1466 case convolve(fcZero, fcInfinity):
1467 case convolve(fcZero, fcNormal):
1470 case convolve(fcZero, fcNaN):
1471 case convolve(fcNormal, fcNaN):
1472 case convolve(fcInfinity, fcNaN):
1474 copySignificand(rhs);
1477 case convolve(fcNormal, fcInfinity):
1481 case convolve(fcNormal, fcZero):
1482 category = fcInfinity;
1485 case convolve(fcInfinity, fcInfinity):
1486 case convolve(fcZero, fcZero):
1490 case convolve(fcNormal, fcNormal):
1496 APFloat::modSpecials(const APFloat &rhs)
1498 switch (convolve(category, rhs.category)) {
1500 llvm_unreachable(0);
1502 case convolve(fcNaN, fcZero):
1503 case convolve(fcNaN, fcNormal):
1504 case convolve(fcNaN, fcInfinity):
1505 case convolve(fcNaN, fcNaN):
1506 case convolve(fcZero, fcInfinity):
1507 case convolve(fcZero, fcNormal):
1508 case convolve(fcNormal, fcInfinity):
1511 case convolve(fcZero, fcNaN):
1512 case convolve(fcNormal, fcNaN):
1513 case convolve(fcInfinity, fcNaN):
1515 copySignificand(rhs);
1518 case convolve(fcNormal, fcZero):
1519 case convolve(fcInfinity, fcZero):
1520 case convolve(fcInfinity, fcNormal):
1521 case convolve(fcInfinity, fcInfinity):
1522 case convolve(fcZero, fcZero):
1526 case convolve(fcNormal, fcNormal):
1533 APFloat::changeSign()
1535 /* Look mummy, this one's easy. */
1540 APFloat::clearSign()
1542 /* So is this one. */
1547 APFloat::copySign(const APFloat &rhs)
1553 /* Normalized addition or subtraction. */
1555 APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode,
1560 assertArithmeticOK(*semantics);
1562 fs = addOrSubtractSpecials(rhs, subtract);
1564 /* This return code means it was not a simple case. */
1565 if (fs == opDivByZero) {
1566 lostFraction lost_fraction;
1568 lost_fraction = addOrSubtractSignificand(rhs, subtract);
1569 fs = normalize(rounding_mode, lost_fraction);
1571 /* Can only be zero if we lost no fraction. */
1572 assert(category != fcZero || lost_fraction == lfExactlyZero);
1575 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1576 positive zero unless rounding to minus infinity, except that
1577 adding two like-signed zeroes gives that zero. */
1578 if (category == fcZero) {
1579 if (rhs.category != fcZero || (sign == rhs.sign) == subtract)
1580 sign = (rounding_mode == rmTowardNegative);
1586 /* Normalized addition. */
1588 APFloat::add(const APFloat &rhs, roundingMode rounding_mode)
1590 return addOrSubtract(rhs, rounding_mode, false);
1593 /* Normalized subtraction. */
1595 APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode)
1597 return addOrSubtract(rhs, rounding_mode, true);
1600 /* Normalized multiply. */
1602 APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode)
1606 assertArithmeticOK(*semantics);
1608 fs = multiplySpecials(rhs);
1610 if (category == fcNormal) {
1611 lostFraction lost_fraction = multiplySignificand(rhs, 0);
1612 fs = normalize(rounding_mode, lost_fraction);
1613 if (lost_fraction != lfExactlyZero)
1614 fs = (opStatus) (fs | opInexact);
1620 /* Normalized divide. */
1622 APFloat::divide(const APFloat &rhs, roundingMode rounding_mode)
1626 assertArithmeticOK(*semantics);
1628 fs = divideSpecials(rhs);
1630 if (category == fcNormal) {
1631 lostFraction lost_fraction = divideSignificand(rhs);
1632 fs = normalize(rounding_mode, lost_fraction);
1633 if (lost_fraction != lfExactlyZero)
1634 fs = (opStatus) (fs | opInexact);
1640 /* Normalized remainder. This is not currently correct in all cases. */
1642 APFloat::remainder(const APFloat &rhs)
1646 unsigned int origSign = sign;
1648 assertArithmeticOK(*semantics);
1649 fs = V.divide(rhs, rmNearestTiesToEven);
1650 if (fs == opDivByZero)
1653 int parts = partCount();
1654 integerPart *x = new integerPart[parts];
1656 fs = V.convertToInteger(x, parts * integerPartWidth, true,
1657 rmNearestTiesToEven, &ignored);
1658 if (fs==opInvalidOp)
1661 fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true,
1662 rmNearestTiesToEven);
1663 assert(fs==opOK); // should always work
1665 fs = V.multiply(rhs, rmNearestTiesToEven);
1666 assert(fs==opOK || fs==opInexact); // should not overflow or underflow
1668 fs = subtract(V, rmNearestTiesToEven);
1669 assert(fs==opOK || fs==opInexact); // likewise
1672 sign = origSign; // IEEE754 requires this
1677 /* Normalized llvm frem (C fmod).
1678 This is not currently correct in all cases. */
1680 APFloat::mod(const APFloat &rhs, roundingMode rounding_mode)
1683 assertArithmeticOK(*semantics);
1684 fs = modSpecials(rhs);
1686 if (category == fcNormal && rhs.category == fcNormal) {
1688 unsigned int origSign = sign;
1690 fs = V.divide(rhs, rmNearestTiesToEven);
1691 if (fs == opDivByZero)
1694 int parts = partCount();
1695 integerPart *x = new integerPart[parts];
1697 fs = V.convertToInteger(x, parts * integerPartWidth, true,
1698 rmTowardZero, &ignored);
1699 if (fs==opInvalidOp)
1702 fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true,
1703 rmNearestTiesToEven);
1704 assert(fs==opOK); // should always work
1706 fs = V.multiply(rhs, rounding_mode);
1707 assert(fs==opOK || fs==opInexact); // should not overflow or underflow
1709 fs = subtract(V, rounding_mode);
1710 assert(fs==opOK || fs==opInexact); // likewise
1713 sign = origSign; // IEEE754 requires this
1719 /* Normalized fused-multiply-add. */
1721 APFloat::fusedMultiplyAdd(const APFloat &multiplicand,
1722 const APFloat &addend,
1723 roundingMode rounding_mode)
1727 assertArithmeticOK(*semantics);
1729 /* Post-multiplication sign, before addition. */
1730 sign ^= multiplicand.sign;
1732 /* If and only if all arguments are normal do we need to do an
1733 extended-precision calculation. */
1734 if (category == fcNormal &&
1735 multiplicand.category == fcNormal &&
1736 addend.category == fcNormal) {
1737 lostFraction lost_fraction;
1739 lost_fraction = multiplySignificand(multiplicand, &addend);
1740 fs = normalize(rounding_mode, lost_fraction);
1741 if (lost_fraction != lfExactlyZero)
1742 fs = (opStatus) (fs | opInexact);
1744 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1745 positive zero unless rounding to minus infinity, except that
1746 adding two like-signed zeroes gives that zero. */
1747 if (category == fcZero && sign != addend.sign)
1748 sign = (rounding_mode == rmTowardNegative);
1750 fs = multiplySpecials(multiplicand);
1752 /* FS can only be opOK or opInvalidOp. There is no more work
1753 to do in the latter case. The IEEE-754R standard says it is
1754 implementation-defined in this case whether, if ADDEND is a
1755 quiet NaN, we raise invalid op; this implementation does so.
1757 If we need to do the addition we can do so with normal
1760 fs = addOrSubtract(addend, rounding_mode, false);
1766 /* Comparison requires normalized numbers. */
1768 APFloat::compare(const APFloat &rhs) const
1772 assertArithmeticOK(*semantics);
1773 assert(semantics == rhs.semantics);
1775 switch (convolve(category, rhs.category)) {
1777 llvm_unreachable(0);
1779 case convolve(fcNaN, fcZero):
1780 case convolve(fcNaN, fcNormal):
1781 case convolve(fcNaN, fcInfinity):
1782 case convolve(fcNaN, fcNaN):
1783 case convolve(fcZero, fcNaN):
1784 case convolve(fcNormal, fcNaN):
1785 case convolve(fcInfinity, fcNaN):
1786 return cmpUnordered;
1788 case convolve(fcInfinity, fcNormal):
1789 case convolve(fcInfinity, fcZero):
1790 case convolve(fcNormal, fcZero):
1794 return cmpGreaterThan;
1796 case convolve(fcNormal, fcInfinity):
1797 case convolve(fcZero, fcInfinity):
1798 case convolve(fcZero, fcNormal):
1800 return cmpGreaterThan;
1804 case convolve(fcInfinity, fcInfinity):
1805 if (sign == rhs.sign)
1810 return cmpGreaterThan;
1812 case convolve(fcZero, fcZero):
1815 case convolve(fcNormal, fcNormal):
1819 /* Two normal numbers. Do they have the same sign? */
1820 if (sign != rhs.sign) {
1822 result = cmpLessThan;
1824 result = cmpGreaterThan;
1826 /* Compare absolute values; invert result if negative. */
1827 result = compareAbsoluteValue(rhs);
1830 if (result == cmpLessThan)
1831 result = cmpGreaterThan;
1832 else if (result == cmpGreaterThan)
1833 result = cmpLessThan;
1840 /// APFloat::convert - convert a value of one floating point type to another.
1841 /// The return value corresponds to the IEEE754 exceptions. *losesInfo
1842 /// records whether the transformation lost information, i.e. whether
1843 /// converting the result back to the original type will produce the
1844 /// original value (this is almost the same as return value==fsOK, but there
1845 /// are edge cases where this is not so).
1848 APFloat::convert(const fltSemantics &toSemantics,
1849 roundingMode rounding_mode, bool *losesInfo)
1851 lostFraction lostFraction;
1852 unsigned int newPartCount, oldPartCount;
1855 const fltSemantics &fromSemantics = *semantics;
1857 assertArithmeticOK(fromSemantics);
1858 assertArithmeticOK(toSemantics);
1859 lostFraction = lfExactlyZero;
1860 newPartCount = partCountForBits(toSemantics.precision + 1);
1861 oldPartCount = partCount();
1862 shift = toSemantics.precision - fromSemantics.precision;
1864 bool X86SpecialNan = false;
1865 if (&fromSemantics == &APFloat::x87DoubleExtended &&
1866 &toSemantics != &APFloat::x87DoubleExtended && category == fcNaN &&
1867 (!(*significandParts() & 0x8000000000000000ULL) ||
1868 !(*significandParts() & 0x4000000000000000ULL))) {
1869 // x86 has some unusual NaNs which cannot be represented in any other
1870 // format; note them here.
1871 X86SpecialNan = true;
1874 // If this is a truncation, perform the shift before we narrow the storage.
1875 if (shift < 0 && (category==fcNormal || category==fcNaN))
1876 lostFraction = shiftRight(significandParts(), oldPartCount, -shift);
1878 // Fix the storage so it can hold to new value.
1879 if (newPartCount > oldPartCount) {
1880 // The new type requires more storage; make it available.
1881 integerPart *newParts;
1882 newParts = new integerPart[newPartCount];
1883 APInt::tcSet(newParts, 0, newPartCount);
1884 if (category==fcNormal || category==fcNaN)
1885 APInt::tcAssign(newParts, significandParts(), oldPartCount);
1887 significand.parts = newParts;
1888 } else if (newPartCount == 1 && oldPartCount != 1) {
1889 // Switch to built-in storage for a single part.
1890 integerPart newPart = 0;
1891 if (category==fcNormal || category==fcNaN)
1892 newPart = significandParts()[0];
1894 significand.part = newPart;
1897 // Now that we have the right storage, switch the semantics.
1898 semantics = &toSemantics;
1900 // If this is an extension, perform the shift now that the storage is
1902 if (shift > 0 && (category==fcNormal || category==fcNaN))
1903 APInt::tcShiftLeft(significandParts(), newPartCount, shift);
1905 if (category == fcNormal) {
1906 fs = normalize(rounding_mode, lostFraction);
1907 *losesInfo = (fs != opOK);
1908 } else if (category == fcNaN) {
1909 *losesInfo = lostFraction != lfExactlyZero || X86SpecialNan;
1910 // gcc forces the Quiet bit on, which means (float)(double)(float_sNan)
1911 // does not give you back the same bits. This is dubious, and we
1912 // don't currently do it. You're really supposed to get
1913 // an invalid operation signal at runtime, but nobody does that.
1923 /* Convert a floating point number to an integer according to the
1924 rounding mode. If the rounded integer value is out of range this
1925 returns an invalid operation exception and the contents of the
1926 destination parts are unspecified. If the rounded value is in
1927 range but the floating point number is not the exact integer, the C
1928 standard doesn't require an inexact exception to be raised. IEEE
1929 854 does require it so we do that.
1931 Note that for conversions to integer type the C standard requires
1932 round-to-zero to always be used. */
1934 APFloat::convertToSignExtendedInteger(integerPart *parts, unsigned int width,
1936 roundingMode rounding_mode,
1937 bool *isExact) const
1939 lostFraction lost_fraction;
1940 const integerPart *src;
1941 unsigned int dstPartsCount, truncatedBits;
1943 assertArithmeticOK(*semantics);
1947 /* Handle the three special cases first. */
1948 if (category == fcInfinity || category == fcNaN)
1951 dstPartsCount = partCountForBits(width);
1953 if (category == fcZero) {
1954 APInt::tcSet(parts, 0, dstPartsCount);
1955 // Negative zero can't be represented as an int.
1960 src = significandParts();
1962 /* Step 1: place our absolute value, with any fraction truncated, in
1965 /* Our absolute value is less than one; truncate everything. */
1966 APInt::tcSet(parts, 0, dstPartsCount);
1967 /* For exponent -1 the integer bit represents .5, look at that.
1968 For smaller exponents leftmost truncated bit is 0. */
1969 truncatedBits = semantics->precision -1U - exponent;
1971 /* We want the most significant (exponent + 1) bits; the rest are
1973 unsigned int bits = exponent + 1U;
1975 /* Hopelessly large in magnitude? */
1979 if (bits < semantics->precision) {
1980 /* We truncate (semantics->precision - bits) bits. */
1981 truncatedBits = semantics->precision - bits;
1982 APInt::tcExtract(parts, dstPartsCount, src, bits, truncatedBits);
1984 /* We want at least as many bits as are available. */
1985 APInt::tcExtract(parts, dstPartsCount, src, semantics->precision, 0);
1986 APInt::tcShiftLeft(parts, dstPartsCount, bits - semantics->precision);
1991 /* Step 2: work out any lost fraction, and increment the absolute
1992 value if we would round away from zero. */
1993 if (truncatedBits) {
1994 lost_fraction = lostFractionThroughTruncation(src, partCount(),
1996 if (lost_fraction != lfExactlyZero &&
1997 roundAwayFromZero(rounding_mode, lost_fraction, truncatedBits)) {
1998 if (APInt::tcIncrement(parts, dstPartsCount))
1999 return opInvalidOp; /* Overflow. */
2002 lost_fraction = lfExactlyZero;
2005 /* Step 3: check if we fit in the destination. */
2006 unsigned int omsb = APInt::tcMSB(parts, dstPartsCount) + 1;
2010 /* Negative numbers cannot be represented as unsigned. */
2014 /* It takes omsb bits to represent the unsigned integer value.
2015 We lose a bit for the sign, but care is needed as the
2016 maximally negative integer is a special case. */
2017 if (omsb == width && APInt::tcLSB(parts, dstPartsCount) + 1 != omsb)
2020 /* This case can happen because of rounding. */
2025 APInt::tcNegate (parts, dstPartsCount);
2027 if (omsb >= width + !isSigned)
2031 if (lost_fraction == lfExactlyZero) {
2038 /* Same as convertToSignExtendedInteger, except we provide
2039 deterministic values in case of an invalid operation exception,
2040 namely zero for NaNs and the minimal or maximal value respectively
2041 for underflow or overflow.
2042 The *isExact output tells whether the result is exact, in the sense
2043 that converting it back to the original floating point type produces
2044 the original value. This is almost equivalent to result==opOK,
2045 except for negative zeroes.
2048 APFloat::convertToInteger(integerPart *parts, unsigned int width,
2050 roundingMode rounding_mode, bool *isExact) const
2054 fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode,
2057 if (fs == opInvalidOp) {
2058 unsigned int bits, dstPartsCount;
2060 dstPartsCount = partCountForBits(width);
2062 if (category == fcNaN)
2067 bits = width - isSigned;
2069 APInt::tcSetLeastSignificantBits(parts, dstPartsCount, bits);
2070 if (sign && isSigned)
2071 APInt::tcShiftLeft(parts, dstPartsCount, width - 1);
2077 /* Same as convertToInteger(integerPart*, ...), except the result is returned in
2078 an APSInt, whose initial bit-width and signed-ness are used to determine the
2079 precision of the conversion.
2082 APFloat::convertToInteger(APSInt &result,
2083 roundingMode rounding_mode, bool *isExact) const
2085 unsigned bitWidth = result.getBitWidth();
2086 SmallVector<uint64_t, 4> parts(result.getNumWords());
2087 opStatus status = convertToInteger(
2088 parts.data(), bitWidth, result.isSigned(), rounding_mode, isExact);
2089 // Keeps the original signed-ness.
2090 result = APInt(bitWidth, parts);
2094 /* Convert an unsigned integer SRC to a floating point number,
2095 rounding according to ROUNDING_MODE. The sign of the floating
2096 point number is not modified. */
2098 APFloat::convertFromUnsignedParts(const integerPart *src,
2099 unsigned int srcCount,
2100 roundingMode rounding_mode)
2102 unsigned int omsb, precision, dstCount;
2104 lostFraction lost_fraction;
2106 assertArithmeticOK(*semantics);
2107 category = fcNormal;
2108 omsb = APInt::tcMSB(src, srcCount) + 1;
2109 dst = significandParts();
2110 dstCount = partCount();
2111 precision = semantics->precision;
2113 /* We want the most significant PRECISION bits of SRC. There may not
2114 be that many; extract what we can. */
2115 if (precision <= omsb) {
2116 exponent = omsb - 1;
2117 lost_fraction = lostFractionThroughTruncation(src, srcCount,
2119 APInt::tcExtract(dst, dstCount, src, precision, omsb - precision);
2121 exponent = precision - 1;
2122 lost_fraction = lfExactlyZero;
2123 APInt::tcExtract(dst, dstCount, src, omsb, 0);
2126 return normalize(rounding_mode, lost_fraction);
2130 APFloat::convertFromAPInt(const APInt &Val,
2132 roundingMode rounding_mode)
2134 unsigned int partCount = Val.getNumWords();
2138 if (isSigned && api.isNegative()) {
2143 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
2146 /* Convert a two's complement integer SRC to a floating point number,
2147 rounding according to ROUNDING_MODE. ISSIGNED is true if the
2148 integer is signed, in which case it must be sign-extended. */
2150 APFloat::convertFromSignExtendedInteger(const integerPart *src,
2151 unsigned int srcCount,
2153 roundingMode rounding_mode)
2157 assertArithmeticOK(*semantics);
2159 APInt::tcExtractBit(src, srcCount * integerPartWidth - 1)) {
2162 /* If we're signed and negative negate a copy. */
2164 copy = new integerPart[srcCount];
2165 APInt::tcAssign(copy, src, srcCount);
2166 APInt::tcNegate(copy, srcCount);
2167 status = convertFromUnsignedParts(copy, srcCount, rounding_mode);
2171 status = convertFromUnsignedParts(src, srcCount, rounding_mode);
2177 /* FIXME: should this just take a const APInt reference? */
2179 APFloat::convertFromZeroExtendedInteger(const integerPart *parts,
2180 unsigned int width, bool isSigned,
2181 roundingMode rounding_mode)
2183 unsigned int partCount = partCountForBits(width);
2184 APInt api = APInt(width, makeArrayRef(parts, partCount));
2187 if (isSigned && APInt::tcExtractBit(parts, width - 1)) {
2192 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
2196 APFloat::convertFromHexadecimalString(StringRef s, roundingMode rounding_mode)
2198 lostFraction lost_fraction = lfExactlyZero;
2199 integerPart *significand;
2200 unsigned int bitPos, partsCount;
2201 StringRef::iterator dot, firstSignificantDigit;
2205 category = fcNormal;
2207 significand = significandParts();
2208 partsCount = partCount();
2209 bitPos = partsCount * integerPartWidth;
2211 /* Skip leading zeroes and any (hexa)decimal point. */
2212 StringRef::iterator begin = s.begin();
2213 StringRef::iterator end = s.end();
2214 StringRef::iterator p = skipLeadingZeroesAndAnyDot(begin, end, &dot);
2215 firstSignificantDigit = p;
2218 integerPart hex_value;
2221 assert(dot == end && "String contains multiple dots");
2228 hex_value = hexDigitValue(*p);
2229 if (hex_value == -1U) {
2238 /* Store the number whilst 4-bit nibbles remain. */
2241 hex_value <<= bitPos % integerPartWidth;
2242 significand[bitPos / integerPartWidth] |= hex_value;
2244 lost_fraction = trailingHexadecimalFraction(p, end, hex_value);
2245 while (p != end && hexDigitValue(*p) != -1U)
2252 /* Hex floats require an exponent but not a hexadecimal point. */
2253 assert(p != end && "Hex strings require an exponent");
2254 assert((*p == 'p' || *p == 'P') && "Invalid character in significand");
2255 assert(p != begin && "Significand has no digits");
2256 assert((dot == end || p - begin != 1) && "Significand has no digits");
2258 /* Ignore the exponent if we are zero. */
2259 if (p != firstSignificantDigit) {
2262 /* Implicit hexadecimal point? */
2266 /* Calculate the exponent adjustment implicit in the number of
2267 significant digits. */
2268 expAdjustment = static_cast<int>(dot - firstSignificantDigit);
2269 if (expAdjustment < 0)
2271 expAdjustment = expAdjustment * 4 - 1;
2273 /* Adjust for writing the significand starting at the most
2274 significant nibble. */
2275 expAdjustment += semantics->precision;
2276 expAdjustment -= partsCount * integerPartWidth;
2278 /* Adjust for the given exponent. */
2279 exponent = totalExponent(p + 1, end, expAdjustment);
2282 return normalize(rounding_mode, lost_fraction);
2286 APFloat::roundSignificandWithExponent(const integerPart *decSigParts,
2287 unsigned sigPartCount, int exp,
2288 roundingMode rounding_mode)
2290 unsigned int parts, pow5PartCount;
2291 fltSemantics calcSemantics = { 32767, -32767, 0, true };
2292 integerPart pow5Parts[maxPowerOfFiveParts];
2295 isNearest = (rounding_mode == rmNearestTiesToEven ||
2296 rounding_mode == rmNearestTiesToAway);
2298 parts = partCountForBits(semantics->precision + 11);
2300 /* Calculate pow(5, abs(exp)). */
2301 pow5PartCount = powerOf5(pow5Parts, exp >= 0 ? exp: -exp);
2303 for (;; parts *= 2) {
2304 opStatus sigStatus, powStatus;
2305 unsigned int excessPrecision, truncatedBits;
2307 calcSemantics.precision = parts * integerPartWidth - 1;
2308 excessPrecision = calcSemantics.precision - semantics->precision;
2309 truncatedBits = excessPrecision;
2311 APFloat decSig(calcSemantics, fcZero, sign);
2312 APFloat pow5(calcSemantics, fcZero, false);
2314 sigStatus = decSig.convertFromUnsignedParts(decSigParts, sigPartCount,
2315 rmNearestTiesToEven);
2316 powStatus = pow5.convertFromUnsignedParts(pow5Parts, pow5PartCount,
2317 rmNearestTiesToEven);
2318 /* Add exp, as 10^n = 5^n * 2^n. */
2319 decSig.exponent += exp;
2321 lostFraction calcLostFraction;
2322 integerPart HUerr, HUdistance;
2323 unsigned int powHUerr;
2326 /* multiplySignificand leaves the precision-th bit set to 1. */
2327 calcLostFraction = decSig.multiplySignificand(pow5, NULL);
2328 powHUerr = powStatus != opOK;
2330 calcLostFraction = decSig.divideSignificand(pow5);
2331 /* Denormal numbers have less precision. */
2332 if (decSig.exponent < semantics->minExponent) {
2333 excessPrecision += (semantics->minExponent - decSig.exponent);
2334 truncatedBits = excessPrecision;
2335 if (excessPrecision > calcSemantics.precision)
2336 excessPrecision = calcSemantics.precision;
2338 /* Extra half-ulp lost in reciprocal of exponent. */
2339 powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2;
2342 /* Both multiplySignificand and divideSignificand return the
2343 result with the integer bit set. */
2344 assert(APInt::tcExtractBit
2345 (decSig.significandParts(), calcSemantics.precision - 1) == 1);
2347 HUerr = HUerrBound(calcLostFraction != lfExactlyZero, sigStatus != opOK,
2349 HUdistance = 2 * ulpsFromBoundary(decSig.significandParts(),
2350 excessPrecision, isNearest);
2352 /* Are we guaranteed to round correctly if we truncate? */
2353 if (HUdistance >= HUerr) {
2354 APInt::tcExtract(significandParts(), partCount(), decSig.significandParts(),
2355 calcSemantics.precision - excessPrecision,
2357 /* Take the exponent of decSig. If we tcExtract-ed less bits
2358 above we must adjust our exponent to compensate for the
2359 implicit right shift. */
2360 exponent = (decSig.exponent + semantics->precision
2361 - (calcSemantics.precision - excessPrecision));
2362 calcLostFraction = lostFractionThroughTruncation(decSig.significandParts(),
2365 return normalize(rounding_mode, calcLostFraction);
2371 APFloat::convertFromDecimalString(StringRef str, roundingMode rounding_mode)
2376 /* Scan the text. */
2377 StringRef::iterator p = str.begin();
2378 interpretDecimal(p, str.end(), &D);
2380 /* Handle the quick cases. First the case of no significant digits,
2381 i.e. zero, and then exponents that are obviously too large or too
2382 small. Writing L for log 10 / log 2, a number d.ddddd*10^exp
2383 definitely overflows if
2385 (exp - 1) * L >= maxExponent
2387 and definitely underflows to zero where
2389 (exp + 1) * L <= minExponent - precision
2391 With integer arithmetic the tightest bounds for L are
2393 93/28 < L < 196/59 [ numerator <= 256 ]
2394 42039/12655 < L < 28738/8651 [ numerator <= 65536 ]
2397 if (decDigitValue(*D.firstSigDigit) >= 10U) {
2401 /* Check whether the normalized exponent is high enough to overflow
2402 max during the log-rebasing in the max-exponent check below. */
2403 } else if (D.normalizedExponent - 1 > INT_MAX / 42039) {
2404 fs = handleOverflow(rounding_mode);
2406 /* If it wasn't, then it also wasn't high enough to overflow max
2407 during the log-rebasing in the min-exponent check. Check that it
2408 won't overflow min in either check, then perform the min-exponent
2410 } else if (D.normalizedExponent - 1 < INT_MIN / 42039 ||
2411 (D.normalizedExponent + 1) * 28738 <=
2412 8651 * (semantics->minExponent - (int) semantics->precision)) {
2413 /* Underflow to zero and round. */
2415 fs = normalize(rounding_mode, lfLessThanHalf);
2417 /* We can finally safely perform the max-exponent check. */
2418 } else if ((D.normalizedExponent - 1) * 42039
2419 >= 12655 * semantics->maxExponent) {
2420 /* Overflow and round. */
2421 fs = handleOverflow(rounding_mode);
2423 integerPart *decSignificand;
2424 unsigned int partCount;
2426 /* A tight upper bound on number of bits required to hold an
2427 N-digit decimal integer is N * 196 / 59. Allocate enough space
2428 to hold the full significand, and an extra part required by
2430 partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1;
2431 partCount = partCountForBits(1 + 196 * partCount / 59);
2432 decSignificand = new integerPart[partCount + 1];
2435 /* Convert to binary efficiently - we do almost all multiplication
2436 in an integerPart. When this would overflow do we do a single
2437 bignum multiplication, and then revert again to multiplication
2438 in an integerPart. */
2440 integerPart decValue, val, multiplier;
2448 if (p == str.end()) {
2452 decValue = decDigitValue(*p++);
2453 assert(decValue < 10U && "Invalid character in significand");
2455 val = val * 10 + decValue;
2456 /* The maximum number that can be multiplied by ten with any
2457 digit added without overflowing an integerPart. */
2458 } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10);
2460 /* Multiply out the current part. */
2461 APInt::tcMultiplyPart(decSignificand, decSignificand, multiplier, val,
2462 partCount, partCount + 1, false);
2464 /* If we used another part (likely but not guaranteed), increase
2466 if (decSignificand[partCount])
2468 } while (p <= D.lastSigDigit);
2470 category = fcNormal;
2471 fs = roundSignificandWithExponent(decSignificand, partCount,
2472 D.exponent, rounding_mode);
2474 delete [] decSignificand;
2481 APFloat::convertFromString(StringRef str, roundingMode rounding_mode)
2483 assertArithmeticOK(*semantics);
2484 assert(!str.empty() && "Invalid string length");
2486 /* Handle a leading minus sign. */
2487 StringRef::iterator p = str.begin();
2488 size_t slen = str.size();
2489 sign = *p == '-' ? 1 : 0;
2490 if (*p == '-' || *p == '+') {
2493 assert(slen && "String has no digits");
2496 if (slen >= 2 && p[0] == '0' && (p[1] == 'x' || p[1] == 'X')) {
2497 assert(slen - 2 && "Invalid string");
2498 return convertFromHexadecimalString(StringRef(p + 2, slen - 2),
2502 return convertFromDecimalString(StringRef(p, slen), rounding_mode);
2505 /* Write out a hexadecimal representation of the floating point value
2506 to DST, which must be of sufficient size, in the C99 form
2507 [-]0xh.hhhhp[+-]d. Return the number of characters written,
2508 excluding the terminating NUL.
2510 If UPPERCASE, the output is in upper case, otherwise in lower case.
2512 HEXDIGITS digits appear altogether, rounding the value if
2513 necessary. If HEXDIGITS is 0, the minimal precision to display the
2514 number precisely is used instead. If nothing would appear after
2515 the decimal point it is suppressed.
2517 The decimal exponent is always printed and has at least one digit.
2518 Zero values display an exponent of zero. Infinities and NaNs
2519 appear as "infinity" or "nan" respectively.
2521 The above rules are as specified by C99. There is ambiguity about
2522 what the leading hexadecimal digit should be. This implementation
2523 uses whatever is necessary so that the exponent is displayed as
2524 stored. This implies the exponent will fall within the IEEE format
2525 range, and the leading hexadecimal digit will be 0 (for denormals),
2526 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with
2527 any other digits zero).
2530 APFloat::convertToHexString(char *dst, unsigned int hexDigits,
2531 bool upperCase, roundingMode rounding_mode) const
2535 assertArithmeticOK(*semantics);
2543 memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1);
2544 dst += sizeof infinityL - 1;
2548 memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1);
2549 dst += sizeof NaNU - 1;
2554 *dst++ = upperCase ? 'X': 'x';
2556 if (hexDigits > 1) {
2558 memset (dst, '0', hexDigits - 1);
2559 dst += hexDigits - 1;
2561 *dst++ = upperCase ? 'P': 'p';
2566 dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode);
2572 return static_cast<unsigned int>(dst - p);
2575 /* Does the hard work of outputting the correctly rounded hexadecimal
2576 form of a normal floating point number with the specified number of
2577 hexadecimal digits. If HEXDIGITS is zero the minimum number of
2578 digits necessary to print the value precisely is output. */
2580 APFloat::convertNormalToHexString(char *dst, unsigned int hexDigits,
2582 roundingMode rounding_mode) const
2584 unsigned int count, valueBits, shift, partsCount, outputDigits;
2585 const char *hexDigitChars;
2586 const integerPart *significand;
2591 *dst++ = upperCase ? 'X': 'x';
2594 hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower;
2596 significand = significandParts();
2597 partsCount = partCount();
2599 /* +3 because the first digit only uses the single integer bit, so
2600 we have 3 virtual zero most-significant-bits. */
2601 valueBits = semantics->precision + 3;
2602 shift = integerPartWidth - valueBits % integerPartWidth;
2604 /* The natural number of digits required ignoring trailing
2605 insignificant zeroes. */
2606 outputDigits = (valueBits - significandLSB () + 3) / 4;
2608 /* hexDigits of zero means use the required number for the
2609 precision. Otherwise, see if we are truncating. If we are,
2610 find out if we need to round away from zero. */
2612 if (hexDigits < outputDigits) {
2613 /* We are dropping non-zero bits, so need to check how to round.
2614 "bits" is the number of dropped bits. */
2616 lostFraction fraction;
2618 bits = valueBits - hexDigits * 4;
2619 fraction = lostFractionThroughTruncation (significand, partsCount, bits);
2620 roundUp = roundAwayFromZero(rounding_mode, fraction, bits);
2622 outputDigits = hexDigits;
2625 /* Write the digits consecutively, and start writing in the location
2626 of the hexadecimal point. We move the most significant digit
2627 left and add the hexadecimal point later. */
2630 count = (valueBits + integerPartWidth - 1) / integerPartWidth;
2632 while (outputDigits && count) {
2635 /* Put the most significant integerPartWidth bits in "part". */
2636 if (--count == partsCount)
2637 part = 0; /* An imaginary higher zero part. */
2639 part = significand[count] << shift;
2642 part |= significand[count - 1] >> (integerPartWidth - shift);
2644 /* Convert as much of "part" to hexdigits as we can. */
2645 unsigned int curDigits = integerPartWidth / 4;
2647 if (curDigits > outputDigits)
2648 curDigits = outputDigits;
2649 dst += partAsHex (dst, part, curDigits, hexDigitChars);
2650 outputDigits -= curDigits;
2656 /* Note that hexDigitChars has a trailing '0'. */
2659 *q = hexDigitChars[hexDigitValue (*q) + 1];
2660 } while (*q == '0');
2663 /* Add trailing zeroes. */
2664 memset (dst, '0', outputDigits);
2665 dst += outputDigits;
2668 /* Move the most significant digit to before the point, and if there
2669 is something after the decimal point add it. This must come
2670 after rounding above. */
2677 /* Finally output the exponent. */
2678 *dst++ = upperCase ? 'P': 'p';
2680 return writeSignedDecimal (dst, exponent);
2683 // For good performance it is desirable for different APFloats
2684 // to produce different integers.
2686 APFloat::getHashValue() const
2688 if (category==fcZero) return sign<<8 | semantics->precision ;
2689 else if (category==fcInfinity) return sign<<9 | semantics->precision;
2690 else if (category==fcNaN) return 1<<10 | semantics->precision;
2692 uint32_t hash = sign<<11 | semantics->precision | exponent<<12;
2693 const integerPart* p = significandParts();
2694 for (int i=partCount(); i>0; i--, p++)
2695 hash ^= ((uint32_t)*p) ^ (uint32_t)((*p)>>32);
2700 // Conversion from APFloat to/from host float/double. It may eventually be
2701 // possible to eliminate these and have everybody deal with APFloats, but that
2702 // will take a while. This approach will not easily extend to long double.
2703 // Current implementation requires integerPartWidth==64, which is correct at
2704 // the moment but could be made more general.
2706 // Denormals have exponent minExponent in APFloat, but minExponent-1 in
2707 // the actual IEEE respresentations. We compensate for that here.
2710 APFloat::convertF80LongDoubleAPFloatToAPInt() const
2712 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended);
2713 assert(partCount()==2);
2715 uint64_t myexponent, mysignificand;
2717 if (category==fcNormal) {
2718 myexponent = exponent+16383; //bias
2719 mysignificand = significandParts()[0];
2720 if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL))
2721 myexponent = 0; // denormal
2722 } else if (category==fcZero) {
2725 } else if (category==fcInfinity) {
2726 myexponent = 0x7fff;
2727 mysignificand = 0x8000000000000000ULL;
2729 assert(category == fcNaN && "Unknown category");
2730 myexponent = 0x7fff;
2731 mysignificand = significandParts()[0];
2735 words[0] = mysignificand;
2736 words[1] = ((uint64_t)(sign & 1) << 15) |
2737 (myexponent & 0x7fffLL);
2738 return APInt(80, words);
2742 APFloat::convertPPCDoubleDoubleAPFloatToAPInt() const
2744 assert(semantics == (const llvm::fltSemantics*)&PPCDoubleDouble);
2745 assert(partCount()==2);
2747 uint64_t myexponent, mysignificand, myexponent2, mysignificand2;
2749 if (category==fcNormal) {
2750 myexponent = exponent + 1023; //bias
2751 myexponent2 = exponent2 + 1023;
2752 mysignificand = significandParts()[0];
2753 mysignificand2 = significandParts()[1];
2754 if (myexponent==1 && !(mysignificand & 0x10000000000000LL))
2755 myexponent = 0; // denormal
2756 if (myexponent2==1 && !(mysignificand2 & 0x10000000000000LL))
2757 myexponent2 = 0; // denormal
2758 } else if (category==fcZero) {
2763 } else if (category==fcInfinity) {
2769 assert(category == fcNaN && "Unknown category");
2771 mysignificand = significandParts()[0];
2772 myexponent2 = exponent2;
2773 mysignificand2 = significandParts()[1];
2777 words[0] = ((uint64_t)(sign & 1) << 63) |
2778 ((myexponent & 0x7ff) << 52) |
2779 (mysignificand & 0xfffffffffffffLL);
2780 words[1] = ((uint64_t)(sign2 & 1) << 63) |
2781 ((myexponent2 & 0x7ff) << 52) |
2782 (mysignificand2 & 0xfffffffffffffLL);
2783 return APInt(128, words);
2787 APFloat::convertQuadrupleAPFloatToAPInt() const
2789 assert(semantics == (const llvm::fltSemantics*)&IEEEquad);
2790 assert(partCount()==2);
2792 uint64_t myexponent, mysignificand, mysignificand2;
2794 if (category==fcNormal) {
2795 myexponent = exponent+16383; //bias
2796 mysignificand = significandParts()[0];
2797 mysignificand2 = significandParts()[1];
2798 if (myexponent==1 && !(mysignificand2 & 0x1000000000000LL))
2799 myexponent = 0; // denormal
2800 } else if (category==fcZero) {
2802 mysignificand = mysignificand2 = 0;
2803 } else if (category==fcInfinity) {
2804 myexponent = 0x7fff;
2805 mysignificand = mysignificand2 = 0;
2807 assert(category == fcNaN && "Unknown category!");
2808 myexponent = 0x7fff;
2809 mysignificand = significandParts()[0];
2810 mysignificand2 = significandParts()[1];
2814 words[0] = mysignificand;
2815 words[1] = ((uint64_t)(sign & 1) << 63) |
2816 ((myexponent & 0x7fff) << 48) |
2817 (mysignificand2 & 0xffffffffffffLL);
2819 return APInt(128, words);
2823 APFloat::convertDoubleAPFloatToAPInt() const
2825 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble);
2826 assert(partCount()==1);
2828 uint64_t myexponent, mysignificand;
2830 if (category==fcNormal) {
2831 myexponent = exponent+1023; //bias
2832 mysignificand = *significandParts();
2833 if (myexponent==1 && !(mysignificand & 0x10000000000000LL))
2834 myexponent = 0; // denormal
2835 } else if (category==fcZero) {
2838 } else if (category==fcInfinity) {
2842 assert(category == fcNaN && "Unknown category!");
2844 mysignificand = *significandParts();
2847 return APInt(64, ((((uint64_t)(sign & 1) << 63) |
2848 ((myexponent & 0x7ff) << 52) |
2849 (mysignificand & 0xfffffffffffffLL))));
2853 APFloat::convertFloatAPFloatToAPInt() const
2855 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle);
2856 assert(partCount()==1);
2858 uint32_t myexponent, mysignificand;
2860 if (category==fcNormal) {
2861 myexponent = exponent+127; //bias
2862 mysignificand = (uint32_t)*significandParts();
2863 if (myexponent == 1 && !(mysignificand & 0x800000))
2864 myexponent = 0; // denormal
2865 } else if (category==fcZero) {
2868 } else if (category==fcInfinity) {
2872 assert(category == fcNaN && "Unknown category!");
2874 mysignificand = (uint32_t)*significandParts();
2877 return APInt(32, (((sign&1) << 31) | ((myexponent&0xff) << 23) |
2878 (mysignificand & 0x7fffff)));
2882 APFloat::convertHalfAPFloatToAPInt() const
2884 assert(semantics == (const llvm::fltSemantics*)&IEEEhalf);
2885 assert(partCount()==1);
2887 uint32_t myexponent, mysignificand;
2889 if (category==fcNormal) {
2890 myexponent = exponent+15; //bias
2891 mysignificand = (uint32_t)*significandParts();
2892 if (myexponent == 1 && !(mysignificand & 0x400))
2893 myexponent = 0; // denormal
2894 } else if (category==fcZero) {
2897 } else if (category==fcInfinity) {
2901 assert(category == fcNaN && "Unknown category!");
2903 mysignificand = (uint32_t)*significandParts();
2906 return APInt(16, (((sign&1) << 15) | ((myexponent&0x1f) << 10) |
2907 (mysignificand & 0x3ff)));
2910 // This function creates an APInt that is just a bit map of the floating
2911 // point constant as it would appear in memory. It is not a conversion,
2912 // and treating the result as a normal integer is unlikely to be useful.
2915 APFloat::bitcastToAPInt() const
2917 if (semantics == (const llvm::fltSemantics*)&IEEEhalf)
2918 return convertHalfAPFloatToAPInt();
2920 if (semantics == (const llvm::fltSemantics*)&IEEEsingle)
2921 return convertFloatAPFloatToAPInt();
2923 if (semantics == (const llvm::fltSemantics*)&IEEEdouble)
2924 return convertDoubleAPFloatToAPInt();
2926 if (semantics == (const llvm::fltSemantics*)&IEEEquad)
2927 return convertQuadrupleAPFloatToAPInt();
2929 if (semantics == (const llvm::fltSemantics*)&PPCDoubleDouble)
2930 return convertPPCDoubleDoubleAPFloatToAPInt();
2932 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended &&
2934 return convertF80LongDoubleAPFloatToAPInt();
2938 APFloat::convertToFloat() const
2940 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle &&
2941 "Float semantics are not IEEEsingle");
2942 APInt api = bitcastToAPInt();
2943 return api.bitsToFloat();
2947 APFloat::convertToDouble() const
2949 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble &&
2950 "Float semantics are not IEEEdouble");
2951 APInt api = bitcastToAPInt();
2952 return api.bitsToDouble();
2955 /// Integer bit is explicit in this format. Intel hardware (387 and later)
2956 /// does not support these bit patterns:
2957 /// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity")
2958 /// exponent = all 1's, integer bit 0, significand nonzero ("pseudoNaN")
2959 /// exponent = 0, integer bit 1 ("pseudodenormal")
2960 /// exponent!=0 nor all 1's, integer bit 0 ("unnormal")
2961 /// At the moment, the first two are treated as NaNs, the second two as Normal.
2963 APFloat::initFromF80LongDoubleAPInt(const APInt &api)
2965 assert(api.getBitWidth()==80);
2966 uint64_t i1 = api.getRawData()[0];
2967 uint64_t i2 = api.getRawData()[1];
2968 uint64_t myexponent = (i2 & 0x7fff);
2969 uint64_t mysignificand = i1;
2971 initialize(&APFloat::x87DoubleExtended);
2972 assert(partCount()==2);
2974 sign = static_cast<unsigned int>(i2>>15);
2975 if (myexponent==0 && mysignificand==0) {
2976 // exponent, significand meaningless
2978 } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) {
2979 // exponent, significand meaningless
2980 category = fcInfinity;
2981 } else if (myexponent==0x7fff && mysignificand!=0x8000000000000000ULL) {
2982 // exponent meaningless
2984 significandParts()[0] = mysignificand;
2985 significandParts()[1] = 0;
2987 category = fcNormal;
2988 exponent = myexponent - 16383;
2989 significandParts()[0] = mysignificand;
2990 significandParts()[1] = 0;
2991 if (myexponent==0) // denormal
2997 APFloat::initFromPPCDoubleDoubleAPInt(const APInt &api)
2999 assert(api.getBitWidth()==128);
3000 uint64_t i1 = api.getRawData()[0];
3001 uint64_t i2 = api.getRawData()[1];
3002 uint64_t myexponent = (i1 >> 52) & 0x7ff;
3003 uint64_t mysignificand = i1 & 0xfffffffffffffLL;
3004 uint64_t myexponent2 = (i2 >> 52) & 0x7ff;
3005 uint64_t mysignificand2 = i2 & 0xfffffffffffffLL;
3007 initialize(&APFloat::PPCDoubleDouble);
3008 assert(partCount()==2);
3010 sign = static_cast<unsigned int>(i1>>63);
3011 sign2 = static_cast<unsigned int>(i2>>63);
3012 if (myexponent==0 && mysignificand==0) {
3013 // exponent, significand meaningless
3014 // exponent2 and significand2 are required to be 0; we don't check
3016 } else if (myexponent==0x7ff && mysignificand==0) {
3017 // exponent, significand meaningless
3018 // exponent2 and significand2 are required to be 0; we don't check
3019 category = fcInfinity;
3020 } else if (myexponent==0x7ff && mysignificand!=0) {
3021 // exponent meaningless. So is the whole second word, but keep it
3024 exponent2 = myexponent2;
3025 significandParts()[0] = mysignificand;
3026 significandParts()[1] = mysignificand2;
3028 category = fcNormal;
3029 // Note there is no category2; the second word is treated as if it is
3030 // fcNormal, although it might be something else considered by itself.
3031 exponent = myexponent - 1023;
3032 exponent2 = myexponent2 - 1023;
3033 significandParts()[0] = mysignificand;
3034 significandParts()[1] = mysignificand2;
3035 if (myexponent==0) // denormal
3038 significandParts()[0] |= 0x10000000000000LL; // integer bit
3042 significandParts()[1] |= 0x10000000000000LL; // integer bit
3047 APFloat::initFromQuadrupleAPInt(const APInt &api)
3049 assert(api.getBitWidth()==128);
3050 uint64_t i1 = api.getRawData()[0];
3051 uint64_t i2 = api.getRawData()[1];
3052 uint64_t myexponent = (i2 >> 48) & 0x7fff;
3053 uint64_t mysignificand = i1;
3054 uint64_t mysignificand2 = i2 & 0xffffffffffffLL;
3056 initialize(&APFloat::IEEEquad);
3057 assert(partCount()==2);
3059 sign = static_cast<unsigned int>(i2>>63);
3060 if (myexponent==0 &&
3061 (mysignificand==0 && mysignificand2==0)) {
3062 // exponent, significand meaningless
3064 } else if (myexponent==0x7fff &&
3065 (mysignificand==0 && mysignificand2==0)) {
3066 // exponent, significand meaningless
3067 category = fcInfinity;
3068 } else if (myexponent==0x7fff &&
3069 (mysignificand!=0 || mysignificand2 !=0)) {
3070 // exponent meaningless
3072 significandParts()[0] = mysignificand;
3073 significandParts()[1] = mysignificand2;
3075 category = fcNormal;
3076 exponent = myexponent - 16383;
3077 significandParts()[0] = mysignificand;
3078 significandParts()[1] = mysignificand2;
3079 if (myexponent==0) // denormal
3082 significandParts()[1] |= 0x1000000000000LL; // integer bit
3087 APFloat::initFromDoubleAPInt(const APInt &api)
3089 assert(api.getBitWidth()==64);
3090 uint64_t i = *api.getRawData();
3091 uint64_t myexponent = (i >> 52) & 0x7ff;
3092 uint64_t mysignificand = i & 0xfffffffffffffLL;
3094 initialize(&APFloat::IEEEdouble);
3095 assert(partCount()==1);
3097 sign = static_cast<unsigned int>(i>>63);
3098 if (myexponent==0 && mysignificand==0) {
3099 // exponent, significand meaningless
3101 } else if (myexponent==0x7ff && mysignificand==0) {
3102 // exponent, significand meaningless
3103 category = fcInfinity;
3104 } else if (myexponent==0x7ff && mysignificand!=0) {
3105 // exponent meaningless
3107 *significandParts() = mysignificand;
3109 category = fcNormal;
3110 exponent = myexponent - 1023;
3111 *significandParts() = mysignificand;
3112 if (myexponent==0) // denormal
3115 *significandParts() |= 0x10000000000000LL; // integer bit
3120 APFloat::initFromFloatAPInt(const APInt & api)
3122 assert(api.getBitWidth()==32);
3123 uint32_t i = (uint32_t)*api.getRawData();
3124 uint32_t myexponent = (i >> 23) & 0xff;
3125 uint32_t mysignificand = i & 0x7fffff;
3127 initialize(&APFloat::IEEEsingle);
3128 assert(partCount()==1);
3131 if (myexponent==0 && mysignificand==0) {
3132 // exponent, significand meaningless
3134 } else if (myexponent==0xff && mysignificand==0) {
3135 // exponent, significand meaningless
3136 category = fcInfinity;
3137 } else if (myexponent==0xff && mysignificand!=0) {
3138 // sign, exponent, significand meaningless
3140 *significandParts() = mysignificand;
3142 category = fcNormal;
3143 exponent = myexponent - 127; //bias
3144 *significandParts() = mysignificand;
3145 if (myexponent==0) // denormal
3148 *significandParts() |= 0x800000; // integer bit
3153 APFloat::initFromHalfAPInt(const APInt & api)
3155 assert(api.getBitWidth()==16);
3156 uint32_t i = (uint32_t)*api.getRawData();
3157 uint32_t myexponent = (i >> 10) & 0x1f;
3158 uint32_t mysignificand = i & 0x3ff;
3160 initialize(&APFloat::IEEEhalf);
3161 assert(partCount()==1);
3164 if (myexponent==0 && mysignificand==0) {
3165 // exponent, significand meaningless
3167 } else if (myexponent==0x1f && mysignificand==0) {
3168 // exponent, significand meaningless
3169 category = fcInfinity;
3170 } else if (myexponent==0x1f && mysignificand!=0) {
3171 // sign, exponent, significand meaningless
3173 *significandParts() = mysignificand;
3175 category = fcNormal;
3176 exponent = myexponent - 15; //bias
3177 *significandParts() = mysignificand;
3178 if (myexponent==0) // denormal
3181 *significandParts() |= 0x400; // integer bit
3185 /// Treat api as containing the bits of a floating point number. Currently
3186 /// we infer the floating point type from the size of the APInt. The
3187 /// isIEEE argument distinguishes between PPC128 and IEEE128 (not meaningful
3188 /// when the size is anything else).
3190 APFloat::initFromAPInt(const APInt& api, bool isIEEE)
3192 if (api.getBitWidth() == 16)
3193 return initFromHalfAPInt(api);
3194 else if (api.getBitWidth() == 32)
3195 return initFromFloatAPInt(api);
3196 else if (api.getBitWidth()==64)
3197 return initFromDoubleAPInt(api);
3198 else if (api.getBitWidth()==80)
3199 return initFromF80LongDoubleAPInt(api);
3200 else if (api.getBitWidth()==128)
3202 initFromQuadrupleAPInt(api) : initFromPPCDoubleDoubleAPInt(api));
3204 llvm_unreachable(0);
3208 APFloat::getAllOnesValue(unsigned BitWidth, bool isIEEE)
3210 return APFloat(APInt::getAllOnesValue(BitWidth), isIEEE);
3213 APFloat APFloat::getLargest(const fltSemantics &Sem, bool Negative) {
3214 APFloat Val(Sem, fcNormal, Negative);
3216 // We want (in interchange format):
3217 // sign = {Negative}
3219 // significand = 1..1
3221 Val.exponent = Sem.maxExponent; // unbiased
3223 // 1-initialize all bits....
3224 Val.zeroSignificand();
3225 integerPart *significand = Val.significandParts();
3226 unsigned N = partCountForBits(Sem.precision);
3227 for (unsigned i = 0; i != N; ++i)
3228 significand[i] = ~((integerPart) 0);
3230 // ...and then clear the top bits for internal consistency.
3231 if (Sem.precision % integerPartWidth != 0)
3233 (((integerPart) 1) << (Sem.precision % integerPartWidth)) - 1;
3238 APFloat APFloat::getSmallest(const fltSemantics &Sem, bool Negative) {
3239 APFloat Val(Sem, fcNormal, Negative);
3241 // We want (in interchange format):
3242 // sign = {Negative}
3244 // significand = 0..01
3246 Val.exponent = Sem.minExponent; // unbiased
3247 Val.zeroSignificand();
3248 Val.significandParts()[0] = 1;
3252 APFloat APFloat::getSmallestNormalized(const fltSemantics &Sem, bool Negative) {
3253 APFloat Val(Sem, fcNormal, Negative);
3255 // We want (in interchange format):
3256 // sign = {Negative}
3258 // significand = 10..0
3260 Val.exponent = Sem.minExponent;
3261 Val.zeroSignificand();
3262 Val.significandParts()[partCountForBits(Sem.precision)-1] |=
3263 (((integerPart) 1) << ((Sem.precision - 1) % integerPartWidth));
3268 APFloat::APFloat(const APInt& api, bool isIEEE) : exponent2(0), sign2(0) {
3269 initFromAPInt(api, isIEEE);
3272 APFloat::APFloat(float f) : exponent2(0), sign2(0) {
3273 initFromAPInt(APInt::floatToBits(f));
3276 APFloat::APFloat(double d) : exponent2(0), sign2(0) {
3277 initFromAPInt(APInt::doubleToBits(d));
3281 static void append(SmallVectorImpl<char> &Buffer,
3282 unsigned N, const char *Str) {
3283 unsigned Start = Buffer.size();
3284 Buffer.set_size(Start + N);
3285 memcpy(&Buffer[Start], Str, N);
3288 template <unsigned N>
3289 void append(SmallVectorImpl<char> &Buffer, const char (&Str)[N]) {
3290 append(Buffer, N, Str);
3293 /// Removes data from the given significand until it is no more
3294 /// precise than is required for the desired precision.
3295 void AdjustToPrecision(APInt &significand,
3296 int &exp, unsigned FormatPrecision) {
3297 unsigned bits = significand.getActiveBits();
3299 // 196/59 is a very slight overestimate of lg_2(10).
3300 unsigned bitsRequired = (FormatPrecision * 196 + 58) / 59;
3302 if (bits <= bitsRequired) return;
3304 unsigned tensRemovable = (bits - bitsRequired) * 59 / 196;
3305 if (!tensRemovable) return;
3307 exp += tensRemovable;
3309 APInt divisor(significand.getBitWidth(), 1);
3310 APInt powten(significand.getBitWidth(), 10);
3312 if (tensRemovable & 1)
3314 tensRemovable >>= 1;
3315 if (!tensRemovable) break;
3319 significand = significand.udiv(divisor);
3321 // Truncate the significand down to its active bit count, but
3322 // don't try to drop below 32.
3323 unsigned newPrecision = std::max(32U, significand.getActiveBits());
3324 significand = significand.trunc(newPrecision);
3328 void AdjustToPrecision(SmallVectorImpl<char> &buffer,
3329 int &exp, unsigned FormatPrecision) {
3330 unsigned N = buffer.size();
3331 if (N <= FormatPrecision) return;
3333 // The most significant figures are the last ones in the buffer.
3334 unsigned FirstSignificant = N - FormatPrecision;
3337 // FIXME: this probably shouldn't use 'round half up'.
3339 // Rounding down is just a truncation, except we also want to drop
3340 // trailing zeros from the new result.
3341 if (buffer[FirstSignificant - 1] < '5') {
3342 while (buffer[FirstSignificant] == '0')
3345 exp += FirstSignificant;
3346 buffer.erase(&buffer[0], &buffer[FirstSignificant]);
3350 // Rounding up requires a decimal add-with-carry. If we continue
3351 // the carry, the newly-introduced zeros will just be truncated.
3352 for (unsigned I = FirstSignificant; I != N; ++I) {
3353 if (buffer[I] == '9') {
3361 // If we carried through, we have exactly one digit of precision.
3362 if (FirstSignificant == N) {
3363 exp += FirstSignificant;
3365 buffer.push_back('1');
3369 exp += FirstSignificant;
3370 buffer.erase(&buffer[0], &buffer[FirstSignificant]);
3374 void APFloat::toString(SmallVectorImpl<char> &Str,
3375 unsigned FormatPrecision,
3376 unsigned FormatMaxPadding) const {
3380 return append(Str, "-Inf");
3382 return append(Str, "+Inf");
3384 case fcNaN: return append(Str, "NaN");
3390 if (!FormatMaxPadding)
3391 append(Str, "0.0E+0");
3403 // Decompose the number into an APInt and an exponent.
3404 int exp = exponent - ((int) semantics->precision - 1);
3405 APInt significand(semantics->precision,
3406 makeArrayRef(significandParts(),
3407 partCountForBits(semantics->precision)));
3409 // Set FormatPrecision if zero. We want to do this before we
3410 // truncate trailing zeros, as those are part of the precision.
3411 if (!FormatPrecision) {
3412 // It's an interesting question whether to use the nominal
3413 // precision or the active precision here for denormals.
3415 // FormatPrecision = ceil(significandBits / lg_2(10))
3416 FormatPrecision = (semantics->precision * 59 + 195) / 196;
3419 // Ignore trailing binary zeros.
3420 int trailingZeros = significand.countTrailingZeros();
3421 exp += trailingZeros;
3422 significand = significand.lshr(trailingZeros);
3424 // Change the exponent from 2^e to 10^e.
3427 } else if (exp > 0) {
3429 significand = significand.zext(semantics->precision + exp);
3430 significand <<= exp;
3432 } else { /* exp < 0 */
3435 // We transform this using the identity:
3436 // (N)(2^-e) == (N)(5^e)(10^-e)
3437 // This means we have to multiply N (the significand) by 5^e.
3438 // To avoid overflow, we have to operate on numbers large
3439 // enough to store N * 5^e:
3440 // log2(N * 5^e) == log2(N) + e * log2(5)
3441 // <= semantics->precision + e * 137 / 59
3442 // (log_2(5) ~ 2.321928 < 2.322034 ~ 137/59)
3444 unsigned precision = semantics->precision + (137 * texp + 136) / 59;
3446 // Multiply significand by 5^e.
3447 // N * 5^0101 == N * 5^(1*1) * 5^(0*2) * 5^(1*4) * 5^(0*8)
3448 significand = significand.zext(precision);
3449 APInt five_to_the_i(precision, 5);
3451 if (texp & 1) significand *= five_to_the_i;
3455 five_to_the_i *= five_to_the_i;
3459 AdjustToPrecision(significand, exp, FormatPrecision);
3461 llvm::SmallVector<char, 256> buffer;
3464 unsigned precision = significand.getBitWidth();
3465 APInt ten(precision, 10);
3466 APInt digit(precision, 0);
3468 bool inTrail = true;
3469 while (significand != 0) {
3470 // digit <- significand % 10
3471 // significand <- significand / 10
3472 APInt::udivrem(significand, ten, significand, digit);
3474 unsigned d = digit.getZExtValue();
3476 // Drop trailing zeros.
3477 if (inTrail && !d) exp++;
3479 buffer.push_back((char) ('0' + d));
3484 assert(!buffer.empty() && "no characters in buffer!");
3486 // Drop down to FormatPrecision.
3487 // TODO: don't do more precise calculations above than are required.
3488 AdjustToPrecision(buffer, exp, FormatPrecision);
3490 unsigned NDigits = buffer.size();
3492 // Check whether we should use scientific notation.
3493 bool FormatScientific;
3494 if (!FormatMaxPadding)
3495 FormatScientific = true;
3500 // But we shouldn't make the number look more precise than it is.
3501 FormatScientific = ((unsigned) exp > FormatMaxPadding ||
3502 NDigits + (unsigned) exp > FormatPrecision);
3504 // Power of the most significant digit.
3505 int MSD = exp + (int) (NDigits - 1);
3508 FormatScientific = false;
3510 // 765e-5 == 0.00765
3512 FormatScientific = ((unsigned) -MSD) > FormatMaxPadding;
3517 // Scientific formatting is pretty straightforward.
3518 if (FormatScientific) {
3519 exp += (NDigits - 1);
3521 Str.push_back(buffer[NDigits-1]);
3526 for (unsigned I = 1; I != NDigits; ++I)
3527 Str.push_back(buffer[NDigits-1-I]);
3530 Str.push_back(exp >= 0 ? '+' : '-');
3531 if (exp < 0) exp = -exp;
3532 SmallVector<char, 6> expbuf;
3534 expbuf.push_back((char) ('0' + (exp % 10)));
3537 for (unsigned I = 0, E = expbuf.size(); I != E; ++I)
3538 Str.push_back(expbuf[E-1-I]);
3542 // Non-scientific, positive exponents.
3544 for (unsigned I = 0; I != NDigits; ++I)
3545 Str.push_back(buffer[NDigits-1-I]);
3546 for (unsigned I = 0; I != (unsigned) exp; ++I)
3551 // Non-scientific, negative exponents.
3553 // The number of digits to the left of the decimal point.
3554 int NWholeDigits = exp + (int) NDigits;
3557 if (NWholeDigits > 0) {
3558 for (; I != (unsigned) NWholeDigits; ++I)
3559 Str.push_back(buffer[NDigits-I-1]);
3562 unsigned NZeros = 1 + (unsigned) -NWholeDigits;
3566 for (unsigned Z = 1; Z != NZeros; ++Z)
3570 for (; I != NDigits; ++I)
3571 Str.push_back(buffer[NDigits-I-1]);
3574 bool APFloat::getExactInverse(APFloat *inv) const {
3575 // We can only guarantee the existence of an exact inverse for IEEE floats.
3576 if (semantics != &IEEEhalf && semantics != &IEEEsingle &&
3577 semantics != &IEEEdouble && semantics != &IEEEquad)
3580 // Special floats and denormals have no exact inverse.
3581 if (category != fcNormal)
3584 // Check that the number is a power of two by making sure that only the
3585 // integer bit is set in the significand.
3586 if (significandLSB() != semantics->precision - 1)
3590 APFloat reciprocal(*semantics, 1ULL);
3591 if (reciprocal.divide(*this, rmNearestTiesToEven) != opOK)
3594 // Avoid multiplication with a denormal, it is not safe on all platforms and
3595 // may be slower than a normal division.
3596 if (reciprocal.significandMSB() + 1 < reciprocal.semantics->precision)
3599 assert(reciprocal.category == fcNormal &&
3600 reciprocal.significandLSB() == reciprocal.semantics->precision - 1);