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/FoldingSet.h"
17 #include "llvm/Support/MathExtras.h"
22 #define convolve(lhs, rhs) ((lhs) * 4 + (rhs))
24 /* Assumed in hexadecimal significand parsing, and conversion to
25 hexadecimal strings. */
26 #define COMPILE_TIME_ASSERT(cond) extern int CTAssert[(cond) ? 1 : -1]
27 COMPILE_TIME_ASSERT(integerPartWidth % 4 == 0);
31 /* Represents floating point arithmetic semantics. */
33 /* The largest E such that 2^E is representable; this matches the
34 definition of IEEE 754. */
35 exponent_t maxExponent;
37 /* The smallest E such that 2^E is a normalized number; this
38 matches the definition of IEEE 754. */
39 exponent_t minExponent;
41 /* Number of bits in the significand. This includes the integer
43 unsigned int precision;
45 /* True if arithmetic is supported. */
46 unsigned int arithmeticOK;
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 };
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 };
60 /* A tight upper bound on number of parts required to hold the value
63 power * 815 / (351 * integerPartWidth) + 1
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));
78 /* Put a bunch of private, handy routines in an anonymous namespace. */
81 static inline unsigned int
82 partCountForBits(unsigned int bits)
84 return ((bits) + integerPartWidth - 1) / integerPartWidth;
87 /* Returns 0U-9U. Return values >= 10U are not digits. */
88 static inline unsigned int
89 decDigitValue(unsigned int c)
95 hexDigitValue(unsigned int c)
115 assertArithmeticOK(const llvm::fltSemantics &semantics) {
116 assert(semantics.arithmeticOK
117 && "Compile-time arithmetic does not support these semantics");
120 /* Return the value of a decimal exponent of the form
123 If the exponent overflows, returns a large exponent with the
126 readExponent(const char *p)
129 unsigned int absExponent;
130 const unsigned int overlargeExponent = 24000; /* FIXME. */
132 isNegative = (*p == '-');
133 if (*p == '-' || *p == '+')
136 absExponent = decDigitValue(*p++);
137 assert (absExponent < 10U);
142 value = decDigitValue(*p);
147 value += absExponent * 10;
148 if (absExponent >= overlargeExponent) {
149 absExponent = overlargeExponent;
156 return -(int) absExponent;
158 return (int) absExponent;
161 /* This is ugly and needs cleaning up, but I don't immediately see
162 how whilst remaining safe. */
164 totalExponent(const char *p, int exponentAdjustment)
166 int unsignedExponent;
167 bool negative, overflow;
170 /* Move past the exponent letter and sign to the digits. */
172 negative = *p == '-';
173 if(*p == '-' || *p == '+')
176 unsignedExponent = 0;
181 value = decDigitValue(*p);
186 unsignedExponent = unsignedExponent * 10 + value;
187 if(unsignedExponent > 65535)
191 if(exponentAdjustment > 65535 || exponentAdjustment < -65536)
195 exponent = unsignedExponent;
197 exponent = -exponent;
198 exponent += exponentAdjustment;
199 if(exponent > 65535 || exponent < -65536)
204 exponent = negative ? -65536: 65535;
210 skipLeadingZeroesAndAnyDot(const char *p, const char **dot)
225 /* Given a normal decimal floating point number of the form
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
235 If the value is zero, V->firstSigDigit points to a non-digit, and
236 the return exponent is zero.
239 const char *firstSigDigit;
240 const char *lastSigDigit;
242 int normalizedExponent;
246 interpretDecimal(const char *p, decimalInfo *D)
250 p = skipLeadingZeroesAndAnyDot (p, &dot);
252 D->firstSigDigit = p;
254 D->normalizedExponent = 0;
261 if (decDigitValue(*p) >= 10U)
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);
271 /* Implied decimal point? */
275 /* Drop insignificant trailing zeroes. */
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)));
292 /* Return the trailing fraction of a hexadecimal number.
293 DIGITVALUE is the first hex digit of the fraction, P points to
296 trailingHexadecimalFraction(const char *p, unsigned int digitValue)
298 unsigned int hexDigit;
300 /* If the first trailing digit isn't 0 or 8 we can work out the
301 fraction immediately. */
303 return lfMoreThanHalf;
304 else if(digitValue < 8 && digitValue > 0)
305 return lfLessThanHalf;
307 /* Otherwise we need to find the first non-zero digit. */
311 hexDigit = hexDigitValue(*p);
313 /* If we ran off the end it is exactly zero or one-half, otherwise
316 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf;
318 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf;
321 /* Return the fraction lost were a bignum truncated losing the least
322 significant BITS bits. */
324 lostFractionThroughTruncation(const integerPart *parts,
325 unsigned int partCount,
330 lsb = APInt::tcLSB(parts, partCount);
332 /* Note this is guaranteed true if bits == 0, or LSB == -1U. */
334 return lfExactlyZero;
336 return lfExactlyHalf;
337 if(bits <= partCount * integerPartWidth
338 && APInt::tcExtractBit(parts, bits - 1))
339 return lfMoreThanHalf;
341 return lfLessThanHalf;
344 /* Shift DST right BITS bits noting lost fraction. */
346 shiftRight(integerPart *dst, unsigned int parts, unsigned int bits)
348 lostFraction lost_fraction;
350 lost_fraction = lostFractionThroughTruncation(dst, parts, bits);
352 APInt::tcShiftRight(dst, parts, bits);
354 return lost_fraction;
357 /* Combine the effect of two lost fractions. */
359 combineLostFractions(lostFraction moreSignificant,
360 lostFraction lessSignificant)
362 if(lessSignificant != lfExactlyZero) {
363 if(moreSignificant == lfExactlyZero)
364 moreSignificant = lfLessThanHalf;
365 else if(moreSignificant == lfExactlyHalf)
366 moreSignificant = lfMoreThanHalf;
369 return moreSignificant;
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.
377 See "How to Read Floating Point Numbers Accurately" by William D
380 HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2)
382 assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8));
384 if (HUerr1 + HUerr2 == 0)
385 return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */
387 return inexactMultiply + 2 * (HUerr1 + HUerr2);
390 /* The number of ulps from the boundary (zero, or half if ISNEAREST)
391 when the least significant BITS are truncated. BITS cannot be
394 ulpsFromBoundary(const integerPart *parts, unsigned int bits, bool isNearest)
396 unsigned int count, partBits;
397 integerPart part, boundary;
402 count = bits / integerPartWidth;
403 partBits = bits % integerPartWidth + 1;
405 part = parts[count] & (~(integerPart) 0 >> (integerPartWidth - partBits));
408 boundary = (integerPart) 1 << (partBits - 1);
413 if (part - boundary <= boundary - part)
414 return part - boundary;
416 return boundary - part;
419 if (part == boundary) {
422 return ~(integerPart) 0; /* A lot. */
425 } else if (part == boundary - 1) {
428 return ~(integerPart) 0; /* A lot. */
433 return ~(integerPart) 0; /* A lot. */
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. */
439 powerOf5(integerPart *dst, unsigned int power)
441 static const integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125,
443 static integerPart pow5s[maxPowerOfFiveParts * 2 + 5] = { 78125 * 5 };
444 static unsigned int partsCount[16] = { 1 };
446 integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5;
449 assert(power <= maxExponent);
454 *p1 = firstEightPowers[power & 7];
460 for (unsigned int n = 0; power; power >>= 1, n++) {
465 /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */
467 pc = partsCount[n - 1];
468 APInt::tcFullMultiply(pow5, pow5 - pc, pow5 - pc, pc, pc);
470 if (pow5[pc - 1] == 0)
478 APInt::tcFullMultiply(p2, p1, pow5, result, pc);
480 if (p2[result - 1] == 0)
483 /* Now result is in p1 with partsCount parts and p2 is scratch
485 tmp = p1, p1 = p2, p2 = tmp;
492 APInt::tcAssign(dst, p1, result);
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";
506 /* Write out an integerPart in hexadecimal, starting with the most
507 significant nibble. Write out exactly COUNT hexdigits, return
510 partAsHex (char *dst, integerPart part, unsigned int count,
511 const char *hexDigitChars)
513 unsigned int result = count;
515 assert (count != 0 && count <= integerPartWidth / 4);
517 part >>= (integerPartWidth - 4 * count);
519 dst[count] = hexDigitChars[part & 0xf];
526 /* Write out an unsigned decimal integer. */
528 writeUnsignedDecimal (char *dst, unsigned int n)
544 /* Write out a signed decimal integer. */
546 writeSignedDecimal (char *dst, int value)
550 dst = writeUnsignedDecimal(dst, -(unsigned) value);
552 dst = writeUnsignedDecimal(dst, value);
560 APFloat::initialize(const fltSemantics *ourSemantics)
564 semantics = ourSemantics;
567 significand.parts = new integerPart[count];
571 APFloat::freeSignificand()
574 delete [] significand.parts;
578 APFloat::assign(const APFloat &rhs)
580 assert(semantics == rhs.semantics);
583 category = rhs.category;
584 exponent = rhs.exponent;
586 exponent2 = rhs.exponent2;
587 if(category == fcNormal || category == fcNaN)
588 copySignificand(rhs);
592 APFloat::copySignificand(const APFloat &rhs)
594 assert(category == fcNormal || category == fcNaN);
595 assert(rhs.partCount() >= partCount());
597 APInt::tcAssign(significandParts(), rhs.significandParts(),
601 /* Make this number a NaN, with an arbitrary but deterministic value
602 for the significand. */
604 APFloat::makeNaN(void)
607 APInt::tcSet(significandParts(), ~0U, partCount());
611 APFloat::operator=(const APFloat &rhs)
614 if(semantics != rhs.semantics) {
616 initialize(rhs.semantics);
625 APFloat::bitwiseIsEqual(const APFloat &rhs) const {
628 if (semantics != rhs.semantics ||
629 category != rhs.category ||
632 if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble &&
635 if (category==fcZero || category==fcInfinity)
637 else if (category==fcNormal && exponent!=rhs.exponent)
639 else if (semantics==(const llvm::fltSemantics*)&PPCDoubleDouble &&
640 exponent2!=rhs.exponent2)
644 const integerPart* p=significandParts();
645 const integerPart* q=rhs.significandParts();
646 for (; i>0; i--, p++, q++) {
654 APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value)
656 assertArithmeticOK(ourSemantics);
657 initialize(&ourSemantics);
660 exponent = ourSemantics.precision - 1;
661 significandParts()[0] = value;
662 normalize(rmNearestTiesToEven, lfExactlyZero);
665 APFloat::APFloat(const fltSemantics &ourSemantics,
666 fltCategory ourCategory, bool negative)
668 assertArithmeticOK(ourSemantics);
669 initialize(&ourSemantics);
670 category = ourCategory;
672 if(category == fcNormal)
674 else if (ourCategory == fcNaN)
678 APFloat::APFloat(const fltSemantics &ourSemantics, const char *text)
680 assertArithmeticOK(ourSemantics);
681 initialize(&ourSemantics);
682 convertFromString(text, rmNearestTiesToEven);
685 APFloat::APFloat(const APFloat &rhs)
687 initialize(rhs.semantics);
696 // Profile - This method 'profiles' an APFloat for use with FoldingSet.
697 void APFloat::Profile(FoldingSetNodeID& ID) const {
698 ID.Add(bitcastToAPInt());
702 APFloat::partCount() const
704 return partCountForBits(semantics->precision + 1);
708 APFloat::semanticsPrecision(const fltSemantics &semantics)
710 return semantics.precision;
714 APFloat::significandParts() const
716 return const_cast<APFloat *>(this)->significandParts();
720 APFloat::significandParts()
722 assert(category == fcNormal || category == fcNaN);
725 return significand.parts;
727 return &significand.part;
731 APFloat::zeroSignificand()
734 APInt::tcSet(significandParts(), 0, partCount());
737 /* Increment an fcNormal floating point number's significand. */
739 APFloat::incrementSignificand()
743 carry = APInt::tcIncrement(significandParts(), partCount());
745 /* Our callers should never cause us to overflow. */
749 /* Add the significand of the RHS. Returns the carry flag. */
751 APFloat::addSignificand(const APFloat &rhs)
755 parts = significandParts();
757 assert(semantics == rhs.semantics);
758 assert(exponent == rhs.exponent);
760 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount());
763 /* Subtract the significand of the RHS with a borrow flag. Returns
766 APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow)
770 parts = significandParts();
772 assert(semantics == rhs.semantics);
773 assert(exponent == rhs.exponent);
775 return APInt::tcSubtract(parts, rhs.significandParts(), borrow,
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
783 APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend)
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;
793 assert(semantics == rhs.semantics);
795 precision = semantics->precision;
796 newPartsCount = partCountForBits(precision * 2);
798 if(newPartsCount > 4)
799 fullSignificand = new integerPart[newPartsCount];
801 fullSignificand = scratch;
803 lhsSignificand = significandParts();
804 partsCount = partCount();
806 APInt::tcFullMultiply(fullSignificand, lhsSignificand,
807 rhs.significandParts(), partsCount, partsCount);
809 lost_fraction = lfExactlyZero;
810 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
811 exponent += rhs.exponent;
814 Significand savedSignificand = significand;
815 const fltSemantics *savedSemantics = semantics;
816 fltSemantics extendedSemantics;
818 unsigned int extendedPrecision;
820 /* Normalize our MSB. */
821 extendedPrecision = precision + precision - 1;
822 if(omsb != extendedPrecision)
824 APInt::tcShiftLeft(fullSignificand, newPartsCount,
825 extendedPrecision - omsb);
826 exponent -= extendedPrecision - omsb;
829 /* Create new semantics. */
830 extendedSemantics = *semantics;
831 extendedSemantics.precision = extendedPrecision;
833 if(newPartsCount == 1)
834 significand.part = fullSignificand[0];
836 significand.parts = fullSignificand;
837 semantics = &extendedSemantics;
839 APFloat extendedAddend(*addend);
840 status = extendedAddend.convert(extendedSemantics, rmTowardZero, &ignored);
841 assert(status == opOK);
842 lost_fraction = addOrSubtractSignificand(extendedAddend, false);
844 /* Restore our state. */
845 if(newPartsCount == 1)
846 fullSignificand[0] = significand.part;
847 significand = savedSignificand;
848 semantics = savedSemantics;
850 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
853 exponent -= (precision - 1);
855 if(omsb > precision) {
856 unsigned int bits, significantParts;
859 bits = omsb - precision;
860 significantParts = partCountForBits(omsb);
861 lf = shiftRight(fullSignificand, significantParts, bits);
862 lost_fraction = combineLostFractions(lf, lost_fraction);
866 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount);
868 if(newPartsCount > 4)
869 delete [] fullSignificand;
871 return lost_fraction;
874 /* Multiply the significands of LHS and RHS to DST. */
876 APFloat::divideSignificand(const APFloat &rhs)
878 unsigned int bit, i, partsCount;
879 const integerPart *rhsSignificand;
880 integerPart *lhsSignificand, *dividend, *divisor;
881 integerPart scratch[4];
882 lostFraction lost_fraction;
884 assert(semantics == rhs.semantics);
886 lhsSignificand = significandParts();
887 rhsSignificand = rhs.significandParts();
888 partsCount = partCount();
891 dividend = new integerPart[partsCount * 2];
895 divisor = dividend + partsCount;
897 /* Copy the dividend and divisor as they will be modified in-place. */
898 for(i = 0; i < partsCount; i++) {
899 dividend[i] = lhsSignificand[i];
900 divisor[i] = rhsSignificand[i];
901 lhsSignificand[i] = 0;
904 exponent -= rhs.exponent;
906 unsigned int precision = semantics->precision;
908 /* Normalize the divisor. */
909 bit = precision - APInt::tcMSB(divisor, partsCount) - 1;
912 APInt::tcShiftLeft(divisor, partsCount, bit);
915 /* Normalize the dividend. */
916 bit = precision - APInt::tcMSB(dividend, partsCount) - 1;
919 APInt::tcShiftLeft(dividend, partsCount, bit);
922 /* Ensure the dividend >= divisor initially for the loop below.
923 Incidentally, this means that the division loop below is
924 guaranteed to set the integer bit to one. */
925 if(APInt::tcCompare(dividend, divisor, partsCount) < 0) {
927 APInt::tcShiftLeft(dividend, partsCount, 1);
928 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0);
932 for(bit = precision; bit; bit -= 1) {
933 if(APInt::tcCompare(dividend, divisor, partsCount) >= 0) {
934 APInt::tcSubtract(dividend, divisor, 0, partsCount);
935 APInt::tcSetBit(lhsSignificand, bit - 1);
938 APInt::tcShiftLeft(dividend, partsCount, 1);
941 /* Figure out the lost fraction. */
942 int cmp = APInt::tcCompare(dividend, divisor, partsCount);
945 lost_fraction = lfMoreThanHalf;
947 lost_fraction = lfExactlyHalf;
948 else if(APInt::tcIsZero(dividend, partsCount))
949 lost_fraction = lfExactlyZero;
951 lost_fraction = lfLessThanHalf;
956 return lost_fraction;
960 APFloat::significandMSB() const
962 return APInt::tcMSB(significandParts(), partCount());
966 APFloat::significandLSB() const
968 return APInt::tcLSB(significandParts(), partCount());
971 /* Note that a zero result is NOT normalized to fcZero. */
973 APFloat::shiftSignificandRight(unsigned int bits)
975 /* Our exponent should not overflow. */
976 assert((exponent_t) (exponent + bits) >= exponent);
980 return shiftRight(significandParts(), partCount(), bits);
983 /* Shift the significand left BITS bits, subtract BITS from its exponent. */
985 APFloat::shiftSignificandLeft(unsigned int bits)
987 assert(bits < semantics->precision);
990 unsigned int partsCount = partCount();
992 APInt::tcShiftLeft(significandParts(), partsCount, bits);
995 assert(!APInt::tcIsZero(significandParts(), partsCount));
1000 APFloat::compareAbsoluteValue(const APFloat &rhs) const
1004 assert(semantics == rhs.semantics);
1005 assert(category == fcNormal);
1006 assert(rhs.category == fcNormal);
1008 compare = exponent - rhs.exponent;
1010 /* If exponents are equal, do an unsigned bignum comparison of the
1013 compare = APInt::tcCompare(significandParts(), rhs.significandParts(),
1017 return cmpGreaterThan;
1018 else if(compare < 0)
1024 /* Handle overflow. Sign is preserved. We either become infinity or
1025 the largest finite number. */
1027 APFloat::handleOverflow(roundingMode rounding_mode)
1030 if(rounding_mode == rmNearestTiesToEven
1031 || rounding_mode == rmNearestTiesToAway
1032 || (rounding_mode == rmTowardPositive && !sign)
1033 || (rounding_mode == rmTowardNegative && sign))
1035 category = fcInfinity;
1036 return (opStatus) (opOverflow | opInexact);
1039 /* Otherwise we become the largest finite number. */
1040 category = fcNormal;
1041 exponent = semantics->maxExponent;
1042 APInt::tcSetLeastSignificantBits(significandParts(), partCount(),
1043 semantics->precision);
1048 /* Returns TRUE if, when truncating the current number, with BIT the
1049 new LSB, with the given lost fraction and rounding mode, the result
1050 would need to be rounded away from zero (i.e., by increasing the
1051 signficand). This routine must work for fcZero of both signs, and
1052 fcNormal numbers. */
1054 APFloat::roundAwayFromZero(roundingMode rounding_mode,
1055 lostFraction lost_fraction,
1056 unsigned int bit) const
1058 /* NaNs and infinities should not have lost fractions. */
1059 assert(category == fcNormal || category == fcZero);
1061 /* Current callers never pass this so we don't handle it. */
1062 assert(lost_fraction != lfExactlyZero);
1064 switch(rounding_mode) {
1068 case rmNearestTiesToAway:
1069 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf;
1071 case rmNearestTiesToEven:
1072 if(lost_fraction == lfMoreThanHalf)
1075 /* Our zeroes don't have a significand to test. */
1076 if(lost_fraction == lfExactlyHalf && category != fcZero)
1077 return APInt::tcExtractBit(significandParts(), bit);
1084 case rmTowardPositive:
1085 return sign == false;
1087 case rmTowardNegative:
1088 return sign == true;
1093 APFloat::normalize(roundingMode rounding_mode,
1094 lostFraction lost_fraction)
1096 unsigned int omsb; /* One, not zero, based MSB. */
1099 if(category != fcNormal)
1102 /* Before rounding normalize the exponent of fcNormal numbers. */
1103 omsb = significandMSB() + 1;
1106 /* OMSB is numbered from 1. We want to place it in the integer
1107 bit numbered PRECISON if possible, with a compensating change in
1109 exponentChange = omsb - semantics->precision;
1111 /* If the resulting exponent is too high, overflow according to
1112 the rounding mode. */
1113 if(exponent + exponentChange > semantics->maxExponent)
1114 return handleOverflow(rounding_mode);
1116 /* Subnormal numbers have exponent minExponent, and their MSB
1117 is forced based on that. */
1118 if(exponent + exponentChange < semantics->minExponent)
1119 exponentChange = semantics->minExponent - exponent;
1121 /* Shifting left is easy as we don't lose precision. */
1122 if(exponentChange < 0) {
1123 assert(lost_fraction == lfExactlyZero);
1125 shiftSignificandLeft(-exponentChange);
1130 if(exponentChange > 0) {
1133 /* Shift right and capture any new lost fraction. */
1134 lf = shiftSignificandRight(exponentChange);
1136 lost_fraction = combineLostFractions(lf, lost_fraction);
1138 /* Keep OMSB up-to-date. */
1139 if(omsb > (unsigned) exponentChange)
1140 omsb -= exponentChange;
1146 /* Now round the number according to rounding_mode given the lost
1149 /* As specified in IEEE 754, since we do not trap we do not report
1150 underflow for exact results. */
1151 if(lost_fraction == lfExactlyZero) {
1152 /* Canonicalize zeroes. */
1159 /* Increment the significand if we're rounding away from zero. */
1160 if(roundAwayFromZero(rounding_mode, lost_fraction, 0)) {
1162 exponent = semantics->minExponent;
1164 incrementSignificand();
1165 omsb = significandMSB() + 1;
1167 /* Did the significand increment overflow? */
1168 if(omsb == (unsigned) semantics->precision + 1) {
1169 /* Renormalize by incrementing the exponent and shifting our
1170 significand right one. However if we already have the
1171 maximum exponent we overflow to infinity. */
1172 if(exponent == semantics->maxExponent) {
1173 category = fcInfinity;
1175 return (opStatus) (opOverflow | opInexact);
1178 shiftSignificandRight(1);
1184 /* The normal case - we were and are not denormal, and any
1185 significand increment above didn't overflow. */
1186 if(omsb == semantics->precision)
1189 /* We have a non-zero denormal. */
1190 assert(omsb < semantics->precision);
1192 /* Canonicalize zeroes. */
1196 /* The fcZero case is a denormal that underflowed to zero. */
1197 return (opStatus) (opUnderflow | opInexact);
1201 APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract)
1203 switch(convolve(category, rhs.category)) {
1207 case convolve(fcNaN, fcZero):
1208 case convolve(fcNaN, fcNormal):
1209 case convolve(fcNaN, fcInfinity):
1210 case convolve(fcNaN, fcNaN):
1211 case convolve(fcNormal, fcZero):
1212 case convolve(fcInfinity, fcNormal):
1213 case convolve(fcInfinity, fcZero):
1216 case convolve(fcZero, fcNaN):
1217 case convolve(fcNormal, fcNaN):
1218 case convolve(fcInfinity, fcNaN):
1220 copySignificand(rhs);
1223 case convolve(fcNormal, fcInfinity):
1224 case convolve(fcZero, fcInfinity):
1225 category = fcInfinity;
1226 sign = rhs.sign ^ subtract;
1229 case convolve(fcZero, fcNormal):
1231 sign = rhs.sign ^ subtract;
1234 case convolve(fcZero, fcZero):
1235 /* Sign depends on rounding mode; handled by caller. */
1238 case convolve(fcInfinity, fcInfinity):
1239 /* Differently signed infinities can only be validly
1241 if((sign ^ rhs.sign) != subtract) {
1248 case convolve(fcNormal, fcNormal):
1253 /* Add or subtract two normal numbers. */
1255 APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract)
1258 lostFraction lost_fraction;
1261 /* Determine if the operation on the absolute values is effectively
1262 an addition or subtraction. */
1263 subtract ^= (sign ^ rhs.sign) ? true : false;
1265 /* Are we bigger exponent-wise than the RHS? */
1266 bits = exponent - rhs.exponent;
1268 /* Subtraction is more subtle than one might naively expect. */
1270 APFloat temp_rhs(rhs);
1274 reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan;
1275 lost_fraction = lfExactlyZero;
1276 } else if (bits > 0) {
1277 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1);
1278 shiftSignificandLeft(1);
1281 lost_fraction = shiftSignificandRight(-bits - 1);
1282 temp_rhs.shiftSignificandLeft(1);
1287 carry = temp_rhs.subtractSignificand
1288 (*this, lost_fraction != lfExactlyZero);
1289 copySignificand(temp_rhs);
1292 carry = subtractSignificand
1293 (temp_rhs, lost_fraction != lfExactlyZero);
1296 /* Invert the lost fraction - it was on the RHS and
1298 if(lost_fraction == lfLessThanHalf)
1299 lost_fraction = lfMoreThanHalf;
1300 else if(lost_fraction == lfMoreThanHalf)
1301 lost_fraction = lfLessThanHalf;
1303 /* The code above is intended to ensure that no borrow is
1308 APFloat temp_rhs(rhs);
1310 lost_fraction = temp_rhs.shiftSignificandRight(bits);
1311 carry = addSignificand(temp_rhs);
1313 lost_fraction = shiftSignificandRight(-bits);
1314 carry = addSignificand(rhs);
1317 /* We have a guard bit; generating a carry cannot happen. */
1321 return lost_fraction;
1325 APFloat::multiplySpecials(const APFloat &rhs)
1327 switch(convolve(category, rhs.category)) {
1331 case convolve(fcNaN, fcZero):
1332 case convolve(fcNaN, fcNormal):
1333 case convolve(fcNaN, fcInfinity):
1334 case convolve(fcNaN, fcNaN):
1337 case convolve(fcZero, fcNaN):
1338 case convolve(fcNormal, fcNaN):
1339 case convolve(fcInfinity, fcNaN):
1341 copySignificand(rhs);
1344 case convolve(fcNormal, fcInfinity):
1345 case convolve(fcInfinity, fcNormal):
1346 case convolve(fcInfinity, fcInfinity):
1347 category = fcInfinity;
1350 case convolve(fcZero, fcNormal):
1351 case convolve(fcNormal, fcZero):
1352 case convolve(fcZero, fcZero):
1356 case convolve(fcZero, fcInfinity):
1357 case convolve(fcInfinity, fcZero):
1361 case convolve(fcNormal, fcNormal):
1367 APFloat::divideSpecials(const APFloat &rhs)
1369 switch(convolve(category, rhs.category)) {
1373 case convolve(fcNaN, fcZero):
1374 case convolve(fcNaN, fcNormal):
1375 case convolve(fcNaN, fcInfinity):
1376 case convolve(fcNaN, fcNaN):
1377 case convolve(fcInfinity, fcZero):
1378 case convolve(fcInfinity, fcNormal):
1379 case convolve(fcZero, fcInfinity):
1380 case convolve(fcZero, fcNormal):
1383 case convolve(fcZero, fcNaN):
1384 case convolve(fcNormal, fcNaN):
1385 case convolve(fcInfinity, fcNaN):
1387 copySignificand(rhs);
1390 case convolve(fcNormal, fcInfinity):
1394 case convolve(fcNormal, fcZero):
1395 category = fcInfinity;
1398 case convolve(fcInfinity, fcInfinity):
1399 case convolve(fcZero, fcZero):
1403 case convolve(fcNormal, fcNormal):
1410 APFloat::changeSign()
1412 /* Look mummy, this one's easy. */
1417 APFloat::clearSign()
1419 /* So is this one. */
1424 APFloat::copySign(const APFloat &rhs)
1430 /* Normalized addition or subtraction. */
1432 APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode,
1437 assertArithmeticOK(*semantics);
1439 fs = addOrSubtractSpecials(rhs, subtract);
1441 /* This return code means it was not a simple case. */
1442 if(fs == opDivByZero) {
1443 lostFraction lost_fraction;
1445 lost_fraction = addOrSubtractSignificand(rhs, subtract);
1446 fs = normalize(rounding_mode, lost_fraction);
1448 /* Can only be zero if we lost no fraction. */
1449 assert(category != fcZero || lost_fraction == lfExactlyZero);
1452 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1453 positive zero unless rounding to minus infinity, except that
1454 adding two like-signed zeroes gives that zero. */
1455 if(category == fcZero) {
1456 if(rhs.category != fcZero || (sign == rhs.sign) == subtract)
1457 sign = (rounding_mode == rmTowardNegative);
1463 /* Normalized addition. */
1465 APFloat::add(const APFloat &rhs, roundingMode rounding_mode)
1467 return addOrSubtract(rhs, rounding_mode, false);
1470 /* Normalized subtraction. */
1472 APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode)
1474 return addOrSubtract(rhs, rounding_mode, true);
1477 /* Normalized multiply. */
1479 APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode)
1483 assertArithmeticOK(*semantics);
1485 fs = multiplySpecials(rhs);
1487 if(category == fcNormal) {
1488 lostFraction lost_fraction = multiplySignificand(rhs, 0);
1489 fs = normalize(rounding_mode, lost_fraction);
1490 if(lost_fraction != lfExactlyZero)
1491 fs = (opStatus) (fs | opInexact);
1497 /* Normalized divide. */
1499 APFloat::divide(const APFloat &rhs, roundingMode rounding_mode)
1503 assertArithmeticOK(*semantics);
1505 fs = divideSpecials(rhs);
1507 if(category == fcNormal) {
1508 lostFraction lost_fraction = divideSignificand(rhs);
1509 fs = normalize(rounding_mode, lost_fraction);
1510 if(lost_fraction != lfExactlyZero)
1511 fs = (opStatus) (fs | opInexact);
1517 /* Normalized remainder. This is not currently doing TRT. */
1519 APFloat::mod(const APFloat &rhs, roundingMode rounding_mode)
1523 unsigned int origSign = sign;
1525 assertArithmeticOK(*semantics);
1526 fs = V.divide(rhs, rmNearestTiesToEven);
1527 if (fs == opDivByZero)
1530 int parts = partCount();
1531 integerPart *x = new integerPart[parts];
1533 fs = V.convertToInteger(x, parts * integerPartWidth, true,
1534 rmTowardZero, &ignored);
1535 if (fs==opInvalidOp)
1538 fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true,
1539 rmNearestTiesToEven);
1540 assert(fs==opOK); // should always work
1542 fs = V.multiply(rhs, rounding_mode);
1543 assert(fs==opOK || fs==opInexact); // should not overflow or underflow
1545 fs = subtract(V, rounding_mode);
1546 assert(fs==opOK || fs==opInexact); // likewise
1549 sign = origSign; // IEEE754 requires this
1554 /* Normalized fused-multiply-add. */
1556 APFloat::fusedMultiplyAdd(const APFloat &multiplicand,
1557 const APFloat &addend,
1558 roundingMode rounding_mode)
1562 assertArithmeticOK(*semantics);
1564 /* Post-multiplication sign, before addition. */
1565 sign ^= multiplicand.sign;
1567 /* If and only if all arguments are normal do we need to do an
1568 extended-precision calculation. */
1569 if(category == fcNormal
1570 && multiplicand.category == fcNormal
1571 && addend.category == fcNormal) {
1572 lostFraction lost_fraction;
1574 lost_fraction = multiplySignificand(multiplicand, &addend);
1575 fs = normalize(rounding_mode, lost_fraction);
1576 if(lost_fraction != lfExactlyZero)
1577 fs = (opStatus) (fs | opInexact);
1579 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1580 positive zero unless rounding to minus infinity, except that
1581 adding two like-signed zeroes gives that zero. */
1582 if(category == fcZero && sign != addend.sign)
1583 sign = (rounding_mode == rmTowardNegative);
1585 fs = multiplySpecials(multiplicand);
1587 /* FS can only be opOK or opInvalidOp. There is no more work
1588 to do in the latter case. The IEEE-754R standard says it is
1589 implementation-defined in this case whether, if ADDEND is a
1590 quiet NaN, we raise invalid op; this implementation does so.
1592 If we need to do the addition we can do so with normal
1595 fs = addOrSubtract(addend, rounding_mode, false);
1601 /* Comparison requires normalized numbers. */
1603 APFloat::compare(const APFloat &rhs) const
1607 assertArithmeticOK(*semantics);
1608 assert(semantics == rhs.semantics);
1610 switch(convolve(category, rhs.category)) {
1614 case convolve(fcNaN, fcZero):
1615 case convolve(fcNaN, fcNormal):
1616 case convolve(fcNaN, fcInfinity):
1617 case convolve(fcNaN, fcNaN):
1618 case convolve(fcZero, fcNaN):
1619 case convolve(fcNormal, fcNaN):
1620 case convolve(fcInfinity, fcNaN):
1621 return cmpUnordered;
1623 case convolve(fcInfinity, fcNormal):
1624 case convolve(fcInfinity, fcZero):
1625 case convolve(fcNormal, fcZero):
1629 return cmpGreaterThan;
1631 case convolve(fcNormal, fcInfinity):
1632 case convolve(fcZero, fcInfinity):
1633 case convolve(fcZero, fcNormal):
1635 return cmpGreaterThan;
1639 case convolve(fcInfinity, fcInfinity):
1640 if(sign == rhs.sign)
1645 return cmpGreaterThan;
1647 case convolve(fcZero, fcZero):
1650 case convolve(fcNormal, fcNormal):
1654 /* Two normal numbers. Do they have the same sign? */
1655 if(sign != rhs.sign) {
1657 result = cmpLessThan;
1659 result = cmpGreaterThan;
1661 /* Compare absolute values; invert result if negative. */
1662 result = compareAbsoluteValue(rhs);
1665 if(result == cmpLessThan)
1666 result = cmpGreaterThan;
1667 else if(result == cmpGreaterThan)
1668 result = cmpLessThan;
1675 /// APFloat::convert - convert a value of one floating point type to another.
1676 /// The return value corresponds to the IEEE754 exceptions. *losesInfo
1677 /// records whether the transformation lost information, i.e. whether
1678 /// converting the result back to the original type will produce the
1679 /// original value (this is almost the same as return value==fsOK, but there
1680 /// are edge cases where this is not so).
1683 APFloat::convert(const fltSemantics &toSemantics,
1684 roundingMode rounding_mode, bool *losesInfo)
1686 lostFraction lostFraction;
1687 unsigned int newPartCount, oldPartCount;
1690 assertArithmeticOK(*semantics);
1691 assertArithmeticOK(toSemantics);
1692 lostFraction = lfExactlyZero;
1693 newPartCount = partCountForBits(toSemantics.precision + 1);
1694 oldPartCount = partCount();
1696 /* Handle storage complications. If our new form is wider,
1697 re-allocate our bit pattern into wider storage. If it is
1698 narrower, we ignore the excess parts, but if narrowing to a
1699 single part we need to free the old storage.
1700 Be careful not to reference significandParts for zeroes
1701 and infinities, since it aborts. */
1702 if (newPartCount > oldPartCount) {
1703 integerPart *newParts;
1704 newParts = new integerPart[newPartCount];
1705 APInt::tcSet(newParts, 0, newPartCount);
1706 if (category==fcNormal || category==fcNaN)
1707 APInt::tcAssign(newParts, significandParts(), oldPartCount);
1709 significand.parts = newParts;
1710 } else if (newPartCount < oldPartCount) {
1711 /* Capture any lost fraction through truncation of parts so we get
1712 correct rounding whilst normalizing. */
1713 if (category==fcNormal)
1714 lostFraction = lostFractionThroughTruncation
1715 (significandParts(), oldPartCount, toSemantics.precision);
1716 if (newPartCount == 1) {
1717 integerPart newPart = 0;
1718 if (category==fcNormal || category==fcNaN)
1719 newPart = significandParts()[0];
1721 significand.part = newPart;
1725 if(category == fcNormal) {
1726 /* Re-interpret our bit-pattern. */
1727 exponent += toSemantics.precision - semantics->precision;
1728 semantics = &toSemantics;
1729 fs = normalize(rounding_mode, lostFraction);
1730 *losesInfo = (fs != opOK);
1731 } else if (category == fcNaN) {
1732 int shift = toSemantics.precision - semantics->precision;
1733 // Do this now so significandParts gets the right answer
1734 const fltSemantics *oldSemantics = semantics;
1735 semantics = &toSemantics;
1737 // No normalization here, just truncate
1739 APInt::tcShiftLeft(significandParts(), newPartCount, shift);
1740 else if (shift < 0) {
1741 unsigned ushift = -shift;
1742 // Figure out if we are losing information. This happens
1743 // if are shifting out something other than 0s, or if the x87 long
1744 // double input did not have its integer bit set (pseudo-NaN), or if the
1745 // x87 long double input did not have its QNan bit set (because the x87
1746 // hardware sets this bit when converting a lower-precision NaN to
1747 // x87 long double).
1748 if (APInt::tcLSB(significandParts(), newPartCount) < ushift)
1750 if (oldSemantics == &APFloat::x87DoubleExtended &&
1751 (!(*significandParts() & 0x8000000000000000ULL) ||
1752 !(*significandParts() & 0x4000000000000000ULL)))
1754 APInt::tcShiftRight(significandParts(), newPartCount, ushift);
1756 // gcc forces the Quiet bit on, which means (float)(double)(float_sNan)
1757 // does not give you back the same bits. This is dubious, and we
1758 // don't currently do it. You're really supposed to get
1759 // an invalid operation signal at runtime, but nobody does that.
1762 semantics = &toSemantics;
1770 /* Convert a floating point number to an integer according to the
1771 rounding mode. If the rounded integer value is out of range this
1772 returns an invalid operation exception and the contents of the
1773 destination parts are unspecified. If the rounded value is in
1774 range but the floating point number is not the exact integer, the C
1775 standard doesn't require an inexact exception to be raised. IEEE
1776 854 does require it so we do that.
1778 Note that for conversions to integer type the C standard requires
1779 round-to-zero to always be used. */
1781 APFloat::convertToSignExtendedInteger(integerPart *parts, unsigned int width,
1783 roundingMode rounding_mode,
1784 bool *isExact) const
1786 lostFraction lost_fraction;
1787 const integerPart *src;
1788 unsigned int dstPartsCount, truncatedBits;
1790 assertArithmeticOK(*semantics);
1794 /* Handle the three special cases first. */
1795 if(category == fcInfinity || category == fcNaN)
1798 dstPartsCount = partCountForBits(width);
1800 if(category == fcZero) {
1801 APInt::tcSet(parts, 0, dstPartsCount);
1802 // Negative zero can't be represented as an int.
1807 src = significandParts();
1809 /* Step 1: place our absolute value, with any fraction truncated, in
1812 /* Our absolute value is less than one; truncate everything. */
1813 APInt::tcSet(parts, 0, dstPartsCount);
1814 /* For exponent -1 the integer bit represents .5, look at that.
1815 For smaller exponents leftmost truncated bit is 0. */
1816 truncatedBits = semantics->precision -1U - exponent;
1818 /* We want the most significant (exponent + 1) bits; the rest are
1820 unsigned int bits = exponent + 1U;
1822 /* Hopelessly large in magnitude? */
1826 if (bits < semantics->precision) {
1827 /* We truncate (semantics->precision - bits) bits. */
1828 truncatedBits = semantics->precision - bits;
1829 APInt::tcExtract(parts, dstPartsCount, src, bits, truncatedBits);
1831 /* We want at least as many bits as are available. */
1832 APInt::tcExtract(parts, dstPartsCount, src, semantics->precision, 0);
1833 APInt::tcShiftLeft(parts, dstPartsCount, bits - semantics->precision);
1838 /* Step 2: work out any lost fraction, and increment the absolute
1839 value if we would round away from zero. */
1840 if (truncatedBits) {
1841 lost_fraction = lostFractionThroughTruncation(src, partCount(),
1843 if (lost_fraction != lfExactlyZero
1844 && roundAwayFromZero(rounding_mode, lost_fraction, truncatedBits)) {
1845 if (APInt::tcIncrement(parts, dstPartsCount))
1846 return opInvalidOp; /* Overflow. */
1849 lost_fraction = lfExactlyZero;
1852 /* Step 3: check if we fit in the destination. */
1853 unsigned int omsb = APInt::tcMSB(parts, dstPartsCount) + 1;
1857 /* Negative numbers cannot be represented as unsigned. */
1861 /* It takes omsb bits to represent the unsigned integer value.
1862 We lose a bit for the sign, but care is needed as the
1863 maximally negative integer is a special case. */
1864 if (omsb == width && APInt::tcLSB(parts, dstPartsCount) + 1 != omsb)
1867 /* This case can happen because of rounding. */
1872 APInt::tcNegate (parts, dstPartsCount);
1874 if (omsb >= width + !isSigned)
1878 if (lost_fraction == lfExactlyZero) {
1885 /* Same as convertToSignExtendedInteger, except we provide
1886 deterministic values in case of an invalid operation exception,
1887 namely zero for NaNs and the minimal or maximal value respectively
1888 for underflow or overflow.
1889 The *isExact output tells whether the result is exact, in the sense
1890 that converting it back to the original floating point type produces
1891 the original value. This is almost equivalent to result==opOK,
1892 except for negative zeroes.
1895 APFloat::convertToInteger(integerPart *parts, unsigned int width,
1897 roundingMode rounding_mode, bool *isExact) const
1901 fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode,
1904 if (fs == opInvalidOp) {
1905 unsigned int bits, dstPartsCount;
1907 dstPartsCount = partCountForBits(width);
1909 if (category == fcNaN)
1914 bits = width - isSigned;
1916 APInt::tcSetLeastSignificantBits(parts, dstPartsCount, bits);
1917 if (sign && isSigned)
1918 APInt::tcShiftLeft(parts, dstPartsCount, width - 1);
1924 /* Convert an unsigned integer SRC to a floating point number,
1925 rounding according to ROUNDING_MODE. The sign of the floating
1926 point number is not modified. */
1928 APFloat::convertFromUnsignedParts(const integerPart *src,
1929 unsigned int srcCount,
1930 roundingMode rounding_mode)
1932 unsigned int omsb, precision, dstCount;
1934 lostFraction lost_fraction;
1936 assertArithmeticOK(*semantics);
1937 category = fcNormal;
1938 omsb = APInt::tcMSB(src, srcCount) + 1;
1939 dst = significandParts();
1940 dstCount = partCount();
1941 precision = semantics->precision;
1943 /* We want the most significant PRECISON bits of SRC. There may not
1944 be that many; extract what we can. */
1945 if (precision <= omsb) {
1946 exponent = omsb - 1;
1947 lost_fraction = lostFractionThroughTruncation(src, srcCount,
1949 APInt::tcExtract(dst, dstCount, src, precision, omsb - precision);
1951 exponent = precision - 1;
1952 lost_fraction = lfExactlyZero;
1953 APInt::tcExtract(dst, dstCount, src, omsb, 0);
1956 return normalize(rounding_mode, lost_fraction);
1960 APFloat::convertFromAPInt(const APInt &Val,
1962 roundingMode rounding_mode)
1964 unsigned int partCount = Val.getNumWords();
1968 if (isSigned && api.isNegative()) {
1973 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
1976 /* Convert a two's complement integer SRC to a floating point number,
1977 rounding according to ROUNDING_MODE. ISSIGNED is true if the
1978 integer is signed, in which case it must be sign-extended. */
1980 APFloat::convertFromSignExtendedInteger(const integerPart *src,
1981 unsigned int srcCount,
1983 roundingMode rounding_mode)
1987 assertArithmeticOK(*semantics);
1989 && APInt::tcExtractBit(src, srcCount * integerPartWidth - 1)) {
1992 /* If we're signed and negative negate a copy. */
1994 copy = new integerPart[srcCount];
1995 APInt::tcAssign(copy, src, srcCount);
1996 APInt::tcNegate(copy, srcCount);
1997 status = convertFromUnsignedParts(copy, srcCount, rounding_mode);
2001 status = convertFromUnsignedParts(src, srcCount, rounding_mode);
2007 /* FIXME: should this just take a const APInt reference? */
2009 APFloat::convertFromZeroExtendedInteger(const integerPart *parts,
2010 unsigned int width, bool isSigned,
2011 roundingMode rounding_mode)
2013 unsigned int partCount = partCountForBits(width);
2014 APInt api = APInt(width, partCount, parts);
2017 if(isSigned && APInt::tcExtractBit(parts, width - 1)) {
2022 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
2026 APFloat::convertFromHexadecimalString(const char *p,
2027 roundingMode rounding_mode)
2029 lostFraction lost_fraction;
2030 integerPart *significand;
2031 unsigned int bitPos, partsCount;
2032 const char *dot, *firstSignificantDigit;
2036 category = fcNormal;
2038 significand = significandParts();
2039 partsCount = partCount();
2040 bitPos = partsCount * integerPartWidth;
2042 /* Skip leading zeroes and any (hexa)decimal point. */
2043 p = skipLeadingZeroesAndAnyDot(p, &dot);
2044 firstSignificantDigit = p;
2047 integerPart hex_value;
2054 hex_value = hexDigitValue(*p);
2055 if(hex_value == -1U) {
2056 lost_fraction = lfExactlyZero;
2062 /* Store the number whilst 4-bit nibbles remain. */
2065 hex_value <<= bitPos % integerPartWidth;
2066 significand[bitPos / integerPartWidth] |= hex_value;
2068 lost_fraction = trailingHexadecimalFraction(p, hex_value);
2069 while(hexDigitValue(*p) != -1U)
2075 /* Hex floats require an exponent but not a hexadecimal point. */
2076 assert(*p == 'p' || *p == 'P');
2078 /* Ignore the exponent if we are zero. */
2079 if(p != firstSignificantDigit) {
2082 /* Implicit hexadecimal point? */
2086 /* Calculate the exponent adjustment implicit in the number of
2087 significant digits. */
2088 expAdjustment = static_cast<int>(dot - firstSignificantDigit);
2089 if(expAdjustment < 0)
2091 expAdjustment = expAdjustment * 4 - 1;
2093 /* Adjust for writing the significand starting at the most
2094 significant nibble. */
2095 expAdjustment += semantics->precision;
2096 expAdjustment -= partsCount * integerPartWidth;
2098 /* Adjust for the given exponent. */
2099 exponent = totalExponent(p, expAdjustment);
2102 return normalize(rounding_mode, lost_fraction);
2106 APFloat::roundSignificandWithExponent(const integerPart *decSigParts,
2107 unsigned sigPartCount, int exp,
2108 roundingMode rounding_mode)
2110 unsigned int parts, pow5PartCount;
2111 fltSemantics calcSemantics = { 32767, -32767, 0, true };
2112 integerPart pow5Parts[maxPowerOfFiveParts];
2115 isNearest = (rounding_mode == rmNearestTiesToEven
2116 || rounding_mode == rmNearestTiesToAway);
2118 parts = partCountForBits(semantics->precision + 11);
2120 /* Calculate pow(5, abs(exp)). */
2121 pow5PartCount = powerOf5(pow5Parts, exp >= 0 ? exp: -exp);
2123 for (;; parts *= 2) {
2124 opStatus sigStatus, powStatus;
2125 unsigned int excessPrecision, truncatedBits;
2127 calcSemantics.precision = parts * integerPartWidth - 1;
2128 excessPrecision = calcSemantics.precision - semantics->precision;
2129 truncatedBits = excessPrecision;
2131 APFloat decSig(calcSemantics, fcZero, sign);
2132 APFloat pow5(calcSemantics, fcZero, false);
2134 sigStatus = decSig.convertFromUnsignedParts(decSigParts, sigPartCount,
2135 rmNearestTiesToEven);
2136 powStatus = pow5.convertFromUnsignedParts(pow5Parts, pow5PartCount,
2137 rmNearestTiesToEven);
2138 /* Add exp, as 10^n = 5^n * 2^n. */
2139 decSig.exponent += exp;
2141 lostFraction calcLostFraction;
2142 integerPart HUerr, HUdistance;
2143 unsigned int powHUerr;
2146 /* multiplySignificand leaves the precision-th bit set to 1. */
2147 calcLostFraction = decSig.multiplySignificand(pow5, NULL);
2148 powHUerr = powStatus != opOK;
2150 calcLostFraction = decSig.divideSignificand(pow5);
2151 /* Denormal numbers have less precision. */
2152 if (decSig.exponent < semantics->minExponent) {
2153 excessPrecision += (semantics->minExponent - decSig.exponent);
2154 truncatedBits = excessPrecision;
2155 if (excessPrecision > calcSemantics.precision)
2156 excessPrecision = calcSemantics.precision;
2158 /* Extra half-ulp lost in reciprocal of exponent. */
2159 powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2;
2162 /* Both multiplySignificand and divideSignificand return the
2163 result with the integer bit set. */
2164 assert (APInt::tcExtractBit
2165 (decSig.significandParts(), calcSemantics.precision - 1) == 1);
2167 HUerr = HUerrBound(calcLostFraction != lfExactlyZero, sigStatus != opOK,
2169 HUdistance = 2 * ulpsFromBoundary(decSig.significandParts(),
2170 excessPrecision, isNearest);
2172 /* Are we guaranteed to round correctly if we truncate? */
2173 if (HUdistance >= HUerr) {
2174 APInt::tcExtract(significandParts(), partCount(), decSig.significandParts(),
2175 calcSemantics.precision - excessPrecision,
2177 /* Take the exponent of decSig. If we tcExtract-ed less bits
2178 above we must adjust our exponent to compensate for the
2179 implicit right shift. */
2180 exponent = (decSig.exponent + semantics->precision
2181 - (calcSemantics.precision - excessPrecision));
2182 calcLostFraction = lostFractionThroughTruncation(decSig.significandParts(),
2185 return normalize(rounding_mode, calcLostFraction);
2191 APFloat::convertFromDecimalString(const char *p, roundingMode rounding_mode)
2196 /* Scan the text. */
2197 interpretDecimal(p, &D);
2199 /* Handle the quick cases. First the case of no significant digits,
2200 i.e. zero, and then exponents that are obviously too large or too
2201 small. Writing L for log 10 / log 2, a number d.ddddd*10^exp
2202 definitely overflows if
2204 (exp - 1) * L >= maxExponent
2206 and definitely underflows to zero where
2208 (exp + 1) * L <= minExponent - precision
2210 With integer arithmetic the tightest bounds for L are
2212 93/28 < L < 196/59 [ numerator <= 256 ]
2213 42039/12655 < L < 28738/8651 [ numerator <= 65536 ]
2216 if (decDigitValue(*D.firstSigDigit) >= 10U) {
2219 } else if ((D.normalizedExponent + 1) * 28738
2220 <= 8651 * (semantics->minExponent - (int) semantics->precision)) {
2221 /* Underflow to zero and round. */
2223 fs = normalize(rounding_mode, lfLessThanHalf);
2224 } else if ((D.normalizedExponent - 1) * 42039
2225 >= 12655 * semantics->maxExponent) {
2226 /* Overflow and round. */
2227 fs = handleOverflow(rounding_mode);
2229 integerPart *decSignificand;
2230 unsigned int partCount;
2232 /* A tight upper bound on number of bits required to hold an
2233 N-digit decimal integer is N * 196 / 59. Allocate enough space
2234 to hold the full significand, and an extra part required by
2236 partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1;
2237 partCount = partCountForBits(1 + 196 * partCount / 59);
2238 decSignificand = new integerPart[partCount + 1];
2241 /* Convert to binary efficiently - we do almost all multiplication
2242 in an integerPart. When this would overflow do we do a single
2243 bignum multiplication, and then revert again to multiplication
2244 in an integerPart. */
2246 integerPart decValue, val, multiplier;
2255 decValue = decDigitValue(*p++);
2257 val = val * 10 + decValue;
2258 /* The maximum number that can be multiplied by ten with any
2259 digit added without overflowing an integerPart. */
2260 } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10);
2262 /* Multiply out the current part. */
2263 APInt::tcMultiplyPart(decSignificand, decSignificand, multiplier, val,
2264 partCount, partCount + 1, false);
2266 /* If we used another part (likely but not guaranteed), increase
2268 if (decSignificand[partCount])
2270 } while (p <= D.lastSigDigit);
2272 category = fcNormal;
2273 fs = roundSignificandWithExponent(decSignificand, partCount,
2274 D.exponent, rounding_mode);
2276 delete [] decSignificand;
2283 APFloat::convertFromString(const char *p, roundingMode rounding_mode)
2285 assertArithmeticOK(*semantics);
2287 /* Handle a leading minus sign. */
2293 if(p[0] == '0' && (p[1] == 'x' || p[1] == 'X'))
2294 return convertFromHexadecimalString(p + 2, rounding_mode);
2296 return convertFromDecimalString(p, rounding_mode);
2299 /* Write out a hexadecimal representation of the floating point value
2300 to DST, which must be of sufficient size, in the C99 form
2301 [-]0xh.hhhhp[+-]d. Return the number of characters written,
2302 excluding the terminating NUL.
2304 If UPPERCASE, the output is in upper case, otherwise in lower case.
2306 HEXDIGITS digits appear altogether, rounding the value if
2307 necessary. If HEXDIGITS is 0, the minimal precision to display the
2308 number precisely is used instead. If nothing would appear after
2309 the decimal point it is suppressed.
2311 The decimal exponent is always printed and has at least one digit.
2312 Zero values display an exponent of zero. Infinities and NaNs
2313 appear as "infinity" or "nan" respectively.
2315 The above rules are as specified by C99. There is ambiguity about
2316 what the leading hexadecimal digit should be. This implementation
2317 uses whatever is necessary so that the exponent is displayed as
2318 stored. This implies the exponent will fall within the IEEE format
2319 range, and the leading hexadecimal digit will be 0 (for denormals),
2320 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with
2321 any other digits zero).
2324 APFloat::convertToHexString(char *dst, unsigned int hexDigits,
2325 bool upperCase, roundingMode rounding_mode) const
2329 assertArithmeticOK(*semantics);
2337 memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1);
2338 dst += sizeof infinityL - 1;
2342 memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1);
2343 dst += sizeof NaNU - 1;
2348 *dst++ = upperCase ? 'X': 'x';
2350 if (hexDigits > 1) {
2352 memset (dst, '0', hexDigits - 1);
2353 dst += hexDigits - 1;
2355 *dst++ = upperCase ? 'P': 'p';
2360 dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode);
2366 return static_cast<unsigned int>(dst - p);
2369 /* Does the hard work of outputting the correctly rounded hexadecimal
2370 form of a normal floating point number with the specified number of
2371 hexadecimal digits. If HEXDIGITS is zero the minimum number of
2372 digits necessary to print the value precisely is output. */
2374 APFloat::convertNormalToHexString(char *dst, unsigned int hexDigits,
2376 roundingMode rounding_mode) const
2378 unsigned int count, valueBits, shift, partsCount, outputDigits;
2379 const char *hexDigitChars;
2380 const integerPart *significand;
2385 *dst++ = upperCase ? 'X': 'x';
2388 hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower;
2390 significand = significandParts();
2391 partsCount = partCount();
2393 /* +3 because the first digit only uses the single integer bit, so
2394 we have 3 virtual zero most-significant-bits. */
2395 valueBits = semantics->precision + 3;
2396 shift = integerPartWidth - valueBits % integerPartWidth;
2398 /* The natural number of digits required ignoring trailing
2399 insignificant zeroes. */
2400 outputDigits = (valueBits - significandLSB () + 3) / 4;
2402 /* hexDigits of zero means use the required number for the
2403 precision. Otherwise, see if we are truncating. If we are,
2404 find out if we need to round away from zero. */
2406 if (hexDigits < outputDigits) {
2407 /* We are dropping non-zero bits, so need to check how to round.
2408 "bits" is the number of dropped bits. */
2410 lostFraction fraction;
2412 bits = valueBits - hexDigits * 4;
2413 fraction = lostFractionThroughTruncation (significand, partsCount, bits);
2414 roundUp = roundAwayFromZero(rounding_mode, fraction, bits);
2416 outputDigits = hexDigits;
2419 /* Write the digits consecutively, and start writing in the location
2420 of the hexadecimal point. We move the most significant digit
2421 left and add the hexadecimal point later. */
2424 count = (valueBits + integerPartWidth - 1) / integerPartWidth;
2426 while (outputDigits && count) {
2429 /* Put the most significant integerPartWidth bits in "part". */
2430 if (--count == partsCount)
2431 part = 0; /* An imaginary higher zero part. */
2433 part = significand[count] << shift;
2436 part |= significand[count - 1] >> (integerPartWidth - shift);
2438 /* Convert as much of "part" to hexdigits as we can. */
2439 unsigned int curDigits = integerPartWidth / 4;
2441 if (curDigits > outputDigits)
2442 curDigits = outputDigits;
2443 dst += partAsHex (dst, part, curDigits, hexDigitChars);
2444 outputDigits -= curDigits;
2450 /* Note that hexDigitChars has a trailing '0'. */
2453 *q = hexDigitChars[hexDigitValue (*q) + 1];
2454 } while (*q == '0');
2457 /* Add trailing zeroes. */
2458 memset (dst, '0', outputDigits);
2459 dst += outputDigits;
2462 /* Move the most significant digit to before the point, and if there
2463 is something after the decimal point add it. This must come
2464 after rounding above. */
2471 /* Finally output the exponent. */
2472 *dst++ = upperCase ? 'P': 'p';
2474 return writeSignedDecimal (dst, exponent);
2477 // For good performance it is desirable for different APFloats
2478 // to produce different integers.
2480 APFloat::getHashValue() const
2482 if (category==fcZero) return sign<<8 | semantics->precision ;
2483 else if (category==fcInfinity) return sign<<9 | semantics->precision;
2484 else if (category==fcNaN) return 1<<10 | semantics->precision;
2486 uint32_t hash = sign<<11 | semantics->precision | exponent<<12;
2487 const integerPart* p = significandParts();
2488 for (int i=partCount(); i>0; i--, p++)
2489 hash ^= ((uint32_t)*p) ^ (uint32_t)((*p)>>32);
2494 // Conversion from APFloat to/from host float/double. It may eventually be
2495 // possible to eliminate these and have everybody deal with APFloats, but that
2496 // will take a while. This approach will not easily extend to long double.
2497 // Current implementation requires integerPartWidth==64, which is correct at
2498 // the moment but could be made more general.
2500 // Denormals have exponent minExponent in APFloat, but minExponent-1 in
2501 // the actual IEEE respresentations. We compensate for that here.
2504 APFloat::convertF80LongDoubleAPFloatToAPInt() const
2506 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended);
2507 assert (partCount()==2);
2509 uint64_t myexponent, mysignificand;
2511 if (category==fcNormal) {
2512 myexponent = exponent+16383; //bias
2513 mysignificand = significandParts()[0];
2514 if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL))
2515 myexponent = 0; // denormal
2516 } else if (category==fcZero) {
2519 } else if (category==fcInfinity) {
2520 myexponent = 0x7fff;
2521 mysignificand = 0x8000000000000000ULL;
2523 assert(category == fcNaN && "Unknown category");
2524 myexponent = 0x7fff;
2525 mysignificand = significandParts()[0];
2529 words[0] = ((uint64_t)(sign & 1) << 63) |
2530 ((myexponent & 0x7fffLL) << 48) |
2531 ((mysignificand >>16) & 0xffffffffffffLL);
2532 words[1] = mysignificand & 0xffff;
2533 return APInt(80, 2, words);
2537 APFloat::convertPPCDoubleDoubleAPFloatToAPInt() const
2539 assert(semantics == (const llvm::fltSemantics*)&PPCDoubleDouble);
2540 assert (partCount()==2);
2542 uint64_t myexponent, mysignificand, myexponent2, mysignificand2;
2544 if (category==fcNormal) {
2545 myexponent = exponent + 1023; //bias
2546 myexponent2 = exponent2 + 1023;
2547 mysignificand = significandParts()[0];
2548 mysignificand2 = significandParts()[1];
2549 if (myexponent==1 && !(mysignificand & 0x10000000000000LL))
2550 myexponent = 0; // denormal
2551 if (myexponent2==1 && !(mysignificand2 & 0x10000000000000LL))
2552 myexponent2 = 0; // denormal
2553 } else if (category==fcZero) {
2558 } else if (category==fcInfinity) {
2564 assert(category == fcNaN && "Unknown category");
2566 mysignificand = significandParts()[0];
2567 myexponent2 = exponent2;
2568 mysignificand2 = significandParts()[1];
2572 words[0] = ((uint64_t)(sign & 1) << 63) |
2573 ((myexponent & 0x7ff) << 52) |
2574 (mysignificand & 0xfffffffffffffLL);
2575 words[1] = ((uint64_t)(sign2 & 1) << 63) |
2576 ((myexponent2 & 0x7ff) << 52) |
2577 (mysignificand2 & 0xfffffffffffffLL);
2578 return APInt(128, 2, words);
2582 APFloat::convertDoubleAPFloatToAPInt() const
2584 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble);
2585 assert (partCount()==1);
2587 uint64_t myexponent, mysignificand;
2589 if (category==fcNormal) {
2590 myexponent = exponent+1023; //bias
2591 mysignificand = *significandParts();
2592 if (myexponent==1 && !(mysignificand & 0x10000000000000LL))
2593 myexponent = 0; // denormal
2594 } else if (category==fcZero) {
2597 } else if (category==fcInfinity) {
2601 assert(category == fcNaN && "Unknown category!");
2603 mysignificand = *significandParts();
2606 return APInt(64, ((((uint64_t)(sign & 1) << 63) |
2607 ((myexponent & 0x7ff) << 52) |
2608 (mysignificand & 0xfffffffffffffLL))));
2612 APFloat::convertFloatAPFloatToAPInt() const
2614 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle);
2615 assert (partCount()==1);
2617 uint32_t myexponent, mysignificand;
2619 if (category==fcNormal) {
2620 myexponent = exponent+127; //bias
2621 mysignificand = (uint32_t)*significandParts();
2622 if (myexponent == 1 && !(mysignificand & 0x800000))
2623 myexponent = 0; // denormal
2624 } else if (category==fcZero) {
2627 } else if (category==fcInfinity) {
2631 assert(category == fcNaN && "Unknown category!");
2633 mysignificand = (uint32_t)*significandParts();
2636 return APInt(32, (((sign&1) << 31) | ((myexponent&0xff) << 23) |
2637 (mysignificand & 0x7fffff)));
2640 // This function creates an APInt that is just a bit map of the floating
2641 // point constant as it would appear in memory. It is not a conversion,
2642 // and treating the result as a normal integer is unlikely to be useful.
2645 APFloat::bitcastToAPInt() const
2647 if (semantics == (const llvm::fltSemantics*)&IEEEsingle)
2648 return convertFloatAPFloatToAPInt();
2650 if (semantics == (const llvm::fltSemantics*)&IEEEdouble)
2651 return convertDoubleAPFloatToAPInt();
2653 if (semantics == (const llvm::fltSemantics*)&PPCDoubleDouble)
2654 return convertPPCDoubleDoubleAPFloatToAPInt();
2656 assert(semantics == (const llvm::fltSemantics*)&x87DoubleExtended &&
2658 return convertF80LongDoubleAPFloatToAPInt();
2662 APFloat::convertToFloat() const
2664 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle);
2665 APInt api = bitcastToAPInt();
2666 return api.bitsToFloat();
2670 APFloat::convertToDouble() const
2672 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble);
2673 APInt api = bitcastToAPInt();
2674 return api.bitsToDouble();
2677 /// Integer bit is explicit in this format. Intel hardware (387 and later)
2678 /// does not support these bit patterns:
2679 /// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity")
2680 /// exponent = all 1's, integer bit 0, significand nonzero ("pseudoNaN")
2681 /// exponent = 0, integer bit 1 ("pseudodenormal")
2682 /// exponent!=0 nor all 1's, integer bit 0 ("unnormal")
2683 /// At the moment, the first two are treated as NaNs, the second two as Normal.
2685 APFloat::initFromF80LongDoubleAPInt(const APInt &api)
2687 assert(api.getBitWidth()==80);
2688 uint64_t i1 = api.getRawData()[0];
2689 uint64_t i2 = api.getRawData()[1];
2690 uint64_t myexponent = (i1 >> 48) & 0x7fff;
2691 uint64_t mysignificand = ((i1 << 16) & 0xffffffffffff0000ULL) |
2694 initialize(&APFloat::x87DoubleExtended);
2695 assert(partCount()==2);
2697 sign = static_cast<unsigned int>(i1>>63);
2698 if (myexponent==0 && mysignificand==0) {
2699 // exponent, significand meaningless
2701 } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) {
2702 // exponent, significand meaningless
2703 category = fcInfinity;
2704 } else if (myexponent==0x7fff && mysignificand!=0x8000000000000000ULL) {
2705 // exponent meaningless
2707 significandParts()[0] = mysignificand;
2708 significandParts()[1] = 0;
2710 category = fcNormal;
2711 exponent = myexponent - 16383;
2712 significandParts()[0] = mysignificand;
2713 significandParts()[1] = 0;
2714 if (myexponent==0) // denormal
2720 APFloat::initFromPPCDoubleDoubleAPInt(const APInt &api)
2722 assert(api.getBitWidth()==128);
2723 uint64_t i1 = api.getRawData()[0];
2724 uint64_t i2 = api.getRawData()[1];
2725 uint64_t myexponent = (i1 >> 52) & 0x7ff;
2726 uint64_t mysignificand = i1 & 0xfffffffffffffLL;
2727 uint64_t myexponent2 = (i2 >> 52) & 0x7ff;
2728 uint64_t mysignificand2 = i2 & 0xfffffffffffffLL;
2730 initialize(&APFloat::PPCDoubleDouble);
2731 assert(partCount()==2);
2733 sign = static_cast<unsigned int>(i1>>63);
2734 sign2 = static_cast<unsigned int>(i2>>63);
2735 if (myexponent==0 && mysignificand==0) {
2736 // exponent, significand meaningless
2737 // exponent2 and significand2 are required to be 0; we don't check
2739 } else if (myexponent==0x7ff && mysignificand==0) {
2740 // exponent, significand meaningless
2741 // exponent2 and significand2 are required to be 0; we don't check
2742 category = fcInfinity;
2743 } else if (myexponent==0x7ff && mysignificand!=0) {
2744 // exponent meaningless. So is the whole second word, but keep it
2747 exponent2 = myexponent2;
2748 significandParts()[0] = mysignificand;
2749 significandParts()[1] = mysignificand2;
2751 category = fcNormal;
2752 // Note there is no category2; the second word is treated as if it is
2753 // fcNormal, although it might be something else considered by itself.
2754 exponent = myexponent - 1023;
2755 exponent2 = myexponent2 - 1023;
2756 significandParts()[0] = mysignificand;
2757 significandParts()[1] = mysignificand2;
2758 if (myexponent==0) // denormal
2761 significandParts()[0] |= 0x10000000000000LL; // integer bit
2765 significandParts()[1] |= 0x10000000000000LL; // integer bit
2770 APFloat::initFromDoubleAPInt(const APInt &api)
2772 assert(api.getBitWidth()==64);
2773 uint64_t i = *api.getRawData();
2774 uint64_t myexponent = (i >> 52) & 0x7ff;
2775 uint64_t mysignificand = i & 0xfffffffffffffLL;
2777 initialize(&APFloat::IEEEdouble);
2778 assert(partCount()==1);
2780 sign = static_cast<unsigned int>(i>>63);
2781 if (myexponent==0 && mysignificand==0) {
2782 // exponent, significand meaningless
2784 } else if (myexponent==0x7ff && mysignificand==0) {
2785 // exponent, significand meaningless
2786 category = fcInfinity;
2787 } else if (myexponent==0x7ff && mysignificand!=0) {
2788 // exponent meaningless
2790 *significandParts() = mysignificand;
2792 category = fcNormal;
2793 exponent = myexponent - 1023;
2794 *significandParts() = mysignificand;
2795 if (myexponent==0) // denormal
2798 *significandParts() |= 0x10000000000000LL; // integer bit
2803 APFloat::initFromFloatAPInt(const APInt & api)
2805 assert(api.getBitWidth()==32);
2806 uint32_t i = (uint32_t)*api.getRawData();
2807 uint32_t myexponent = (i >> 23) & 0xff;
2808 uint32_t mysignificand = i & 0x7fffff;
2810 initialize(&APFloat::IEEEsingle);
2811 assert(partCount()==1);
2814 if (myexponent==0 && mysignificand==0) {
2815 // exponent, significand meaningless
2817 } else if (myexponent==0xff && mysignificand==0) {
2818 // exponent, significand meaningless
2819 category = fcInfinity;
2820 } else if (myexponent==0xff && mysignificand!=0) {
2821 // sign, exponent, significand meaningless
2823 *significandParts() = mysignificand;
2825 category = fcNormal;
2826 exponent = myexponent - 127; //bias
2827 *significandParts() = mysignificand;
2828 if (myexponent==0) // denormal
2831 *significandParts() |= 0x800000; // integer bit
2835 /// Treat api as containing the bits of a floating point number. Currently
2836 /// we infer the floating point type from the size of the APInt. The
2837 /// isIEEE argument distinguishes between PPC128 and IEEE128 (not meaningful
2838 /// when the size is anything else).
2840 APFloat::initFromAPInt(const APInt& api, bool isIEEE)
2842 if (api.getBitWidth() == 32)
2843 return initFromFloatAPInt(api);
2844 else if (api.getBitWidth()==64)
2845 return initFromDoubleAPInt(api);
2846 else if (api.getBitWidth()==80)
2847 return initFromF80LongDoubleAPInt(api);
2848 else if (api.getBitWidth()==128 && !isIEEE)
2849 return initFromPPCDoubleDoubleAPInt(api);
2854 APFloat::APFloat(const APInt& api, bool isIEEE)
2856 initFromAPInt(api, isIEEE);
2859 APFloat::APFloat(float f)
2861 APInt api = APInt(32, 0);
2862 initFromAPInt(api.floatToBits(f));
2865 APFloat::APFloat(double d)
2867 APInt api = APInt(64, 0);
2868 initFromAPInt(api.doubleToBits(d));