1 //===-- APFloat.cpp - Implement APFloat class -----------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Neil Booth and is distributed under the
6 // University of Illinois Open Source 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 //===----------------------------------------------------------------------===//
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/Support/MathExtras.h"
21 #define convolve(lhs, rhs) ((lhs) * 4 + (rhs))
23 /* Assumed in hexadecimal significand parsing. */
24 COMPILE_TIME_ASSERT(integerPartWidth % 4 == 0);
28 /* Represents floating point arithmetic semantics. */
30 /* The largest E such that 2^E is representable; this matches the
31 definition of IEEE 754. */
32 exponent_t maxExponent;
34 /* The smallest E such that 2^E is a normalized number; this
35 matches the definition of IEEE 754. */
36 exponent_t minExponent;
38 /* Number of bits in the significand. This includes the integer
40 unsigned char precision;
42 /* If the target format has an implicit integer bit. */
43 bool implicitIntegerBit;
46 const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, true };
47 const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, true };
48 const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, true };
49 const fltSemantics APFloat::x87DoubleExtended = { 16383, -16382, 64, false };
50 const fltSemantics APFloat::Bogus = { 0, 0, 0, false };
53 /* Put a bunch of private, handy routines in an anonymous namespace. */
57 partCountForBits(unsigned int bits)
59 return ((bits) + integerPartWidth - 1) / integerPartWidth;
63 digitValue(unsigned int c)
75 hexDigitValue (unsigned int c)
94 /* This is ugly and needs cleaning up, but I don't immediately see
95 how whilst remaining safe. */
97 totalExponent(const char *p, int exponentAdjustment)
99 integerPart unsignedExponent;
100 bool negative, overflow;
103 /* Move past the exponent letter and sign to the digits. */
105 negative = *p == '-';
106 if(*p == '-' || *p == '+')
109 unsignedExponent = 0;
114 value = digitValue(*p);
119 unsignedExponent = unsignedExponent * 10 + value;
120 if(unsignedExponent > 65535)
124 if(exponentAdjustment > 65535 || exponentAdjustment < -65536)
128 exponent = unsignedExponent;
130 exponent = -exponent;
131 exponent += exponentAdjustment;
132 if(exponent > 65535 || exponent < -65536)
137 exponent = negative ? -65536: 65535;
143 skipLeadingZeroesAndAnyDot(const char *p, const char **dot)
158 /* Return the trailing fraction of a hexadecimal number.
159 DIGITVALUE is the first hex digit of the fraction, P points to
162 trailingHexadecimalFraction(const char *p, unsigned int digitValue)
164 unsigned int hexDigit;
166 /* If the first trailing digit isn't 0 or 8 we can work out the
167 fraction immediately. */
169 return lfMoreThanHalf;
170 else if(digitValue < 8 && digitValue > 0)
171 return lfLessThanHalf;
173 /* Otherwise we need to find the first non-zero digit. */
177 hexDigit = hexDigitValue(*p);
179 /* If we ran off the end it is exactly zero or one-half, otherwise
182 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf;
184 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf;
187 /* Return the fraction lost were a bignum truncated. */
189 lostFractionThroughTruncation(integerPart *parts,
190 unsigned int partCount,
195 lsb = APInt::tcLSB(parts, partCount);
197 /* Note this is guaranteed true if bits == 0, or LSB == -1U. */
199 return lfExactlyZero;
201 return lfExactlyHalf;
202 if(bits <= partCount * integerPartWidth
203 && APInt::tcExtractBit(parts, bits - 1))
204 return lfMoreThanHalf;
206 return lfLessThanHalf;
209 /* Shift DST right BITS bits noting lost fraction. */
211 shiftRight(integerPart *dst, unsigned int parts, unsigned int bits)
213 lostFraction lost_fraction;
215 lost_fraction = lostFractionThroughTruncation(dst, parts, bits);
217 APInt::tcShiftRight(dst, parts, bits);
219 return lost_fraction;
225 APFloat::initialize(const fltSemantics *ourSemantics)
229 semantics = ourSemantics;
232 significand.parts = new integerPart[count];
236 APFloat::freeSignificand()
239 delete [] significand.parts;
243 APFloat::assign(const APFloat &rhs)
245 assert(semantics == rhs.semantics);
248 category = rhs.category;
249 exponent = rhs.exponent;
250 if(category == fcNormal || category == fcNaN)
251 copySignificand(rhs);
255 APFloat::copySignificand(const APFloat &rhs)
257 assert(category == fcNormal || category == fcNaN);
258 assert(rhs.partCount() >= partCount());
260 APInt::tcAssign(significandParts(), rhs.significandParts(),
265 APFloat::operator=(const APFloat &rhs)
268 if(semantics != rhs.semantics) {
270 initialize(rhs.semantics);
279 APFloat::bitwiseIsEqual(const APFloat &rhs) const {
282 if (semantics != rhs.semantics ||
283 category != rhs.category ||
286 if (category==fcZero || category==fcInfinity)
288 else if (category==fcNormal && exponent!=rhs.exponent)
292 const integerPart* p=significandParts();
293 const integerPart* q=rhs.significandParts();
294 for (; i>0; i--, p++, q++) {
302 APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value)
304 initialize(&ourSemantics);
307 exponent = ourSemantics.precision - 1;
308 significandParts()[0] = value;
309 normalize(rmNearestTiesToEven, lfExactlyZero);
312 APFloat::APFloat(const fltSemantics &ourSemantics,
313 fltCategory ourCategory, bool negative)
315 initialize(&ourSemantics);
316 category = ourCategory;
318 if(category == fcNormal)
322 APFloat::APFloat(const fltSemantics &ourSemantics, const char *text)
324 initialize(&ourSemantics);
325 convertFromString(text, rmNearestTiesToEven);
328 APFloat::APFloat(const APFloat &rhs)
330 initialize(rhs.semantics);
340 APFloat::partCount() const
342 return partCountForBits(semantics->precision + 1);
346 APFloat::semanticsPrecision(const fltSemantics &semantics)
348 return semantics.precision;
352 APFloat::significandParts() const
354 return const_cast<APFloat *>(this)->significandParts();
358 APFloat::significandParts()
360 assert(category == fcNormal || category == fcNaN);
363 return significand.parts;
365 return &significand.part;
368 /* Combine the effect of two lost fractions. */
370 APFloat::combineLostFractions(lostFraction moreSignificant,
371 lostFraction lessSignificant)
373 if(lessSignificant != lfExactlyZero) {
374 if(moreSignificant == lfExactlyZero)
375 moreSignificant = lfLessThanHalf;
376 else if(moreSignificant == lfExactlyHalf)
377 moreSignificant = lfMoreThanHalf;
380 return moreSignificant;
384 APFloat::zeroSignificand()
387 APInt::tcSet(significandParts(), 0, partCount());
390 /* Increment an fcNormal floating point number's significand. */
392 APFloat::incrementSignificand()
396 carry = APInt::tcIncrement(significandParts(), partCount());
398 /* Our callers should never cause us to overflow. */
402 /* Add the significand of the RHS. Returns the carry flag. */
404 APFloat::addSignificand(const APFloat &rhs)
408 parts = significandParts();
410 assert(semantics == rhs.semantics);
411 assert(exponent == rhs.exponent);
413 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount());
416 /* Subtract the significand of the RHS with a borrow flag. Returns
419 APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow)
423 parts = significandParts();
425 assert(semantics == rhs.semantics);
426 assert(exponent == rhs.exponent);
428 return APInt::tcSubtract(parts, rhs.significandParts(), borrow,
432 /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it
433 on to the full-precision result of the multiplication. Returns the
436 APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend)
438 unsigned int omsb; // One, not zero, based MSB.
439 unsigned int partsCount, newPartsCount, precision;
440 integerPart *lhsSignificand;
441 integerPart scratch[4];
442 integerPart *fullSignificand;
443 lostFraction lost_fraction;
445 assert(semantics == rhs.semantics);
447 precision = semantics->precision;
448 newPartsCount = partCountForBits(precision * 2);
450 if(newPartsCount > 4)
451 fullSignificand = new integerPart[newPartsCount];
453 fullSignificand = scratch;
455 lhsSignificand = significandParts();
456 partsCount = partCount();
458 APInt::tcFullMultiply(fullSignificand, lhsSignificand,
459 rhs.significandParts(), partsCount);
461 lost_fraction = lfExactlyZero;
462 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
463 exponent += rhs.exponent;
466 Significand savedSignificand = significand;
467 const fltSemantics *savedSemantics = semantics;
468 fltSemantics extendedSemantics;
470 unsigned int extendedPrecision;
472 /* Normalize our MSB. */
473 extendedPrecision = precision + precision - 1;
474 if(omsb != extendedPrecision)
476 APInt::tcShiftLeft(fullSignificand, newPartsCount,
477 extendedPrecision - omsb);
478 exponent -= extendedPrecision - omsb;
481 /* Create new semantics. */
482 extendedSemantics = *semantics;
483 extendedSemantics.precision = extendedPrecision;
485 if(newPartsCount == 1)
486 significand.part = fullSignificand[0];
488 significand.parts = fullSignificand;
489 semantics = &extendedSemantics;
491 APFloat extendedAddend(*addend);
492 status = extendedAddend.convert(extendedSemantics, rmTowardZero);
493 assert(status == opOK);
494 lost_fraction = addOrSubtractSignificand(extendedAddend, false);
496 /* Restore our state. */
497 if(newPartsCount == 1)
498 fullSignificand[0] = significand.part;
499 significand = savedSignificand;
500 semantics = savedSemantics;
502 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
505 exponent -= (precision - 1);
507 if(omsb > precision) {
508 unsigned int bits, significantParts;
511 bits = omsb - precision;
512 significantParts = partCountForBits(omsb);
513 lf = shiftRight(fullSignificand, significantParts, bits);
514 lost_fraction = combineLostFractions(lf, lost_fraction);
518 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount);
520 if(newPartsCount > 4)
521 delete [] fullSignificand;
523 return lost_fraction;
526 /* Multiply the significands of LHS and RHS to DST. */
528 APFloat::divideSignificand(const APFloat &rhs)
530 unsigned int bit, i, partsCount;
531 const integerPart *rhsSignificand;
532 integerPart *lhsSignificand, *dividend, *divisor;
533 integerPart scratch[4];
534 lostFraction lost_fraction;
536 assert(semantics == rhs.semantics);
538 lhsSignificand = significandParts();
539 rhsSignificand = rhs.significandParts();
540 partsCount = partCount();
543 dividend = new integerPart[partsCount * 2];
547 divisor = dividend + partsCount;
549 /* Copy the dividend and divisor as they will be modified in-place. */
550 for(i = 0; i < partsCount; i++) {
551 dividend[i] = lhsSignificand[i];
552 divisor[i] = rhsSignificand[i];
553 lhsSignificand[i] = 0;
556 exponent -= rhs.exponent;
558 unsigned int precision = semantics->precision;
560 /* Normalize the divisor. */
561 bit = precision - APInt::tcMSB(divisor, partsCount) - 1;
564 APInt::tcShiftLeft(divisor, partsCount, bit);
567 /* Normalize the dividend. */
568 bit = precision - APInt::tcMSB(dividend, partsCount) - 1;
571 APInt::tcShiftLeft(dividend, partsCount, bit);
574 if(APInt::tcCompare(dividend, divisor, partsCount) < 0) {
576 APInt::tcShiftLeft(dividend, partsCount, 1);
577 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0);
581 for(bit = precision; bit; bit -= 1) {
582 if(APInt::tcCompare(dividend, divisor, partsCount) >= 0) {
583 APInt::tcSubtract(dividend, divisor, 0, partsCount);
584 APInt::tcSetBit(lhsSignificand, bit - 1);
587 APInt::tcShiftLeft(dividend, partsCount, 1);
590 /* Figure out the lost fraction. */
591 int cmp = APInt::tcCompare(dividend, divisor, partsCount);
594 lost_fraction = lfMoreThanHalf;
596 lost_fraction = lfExactlyHalf;
597 else if(APInt::tcIsZero(dividend, partsCount))
598 lost_fraction = lfExactlyZero;
600 lost_fraction = lfLessThanHalf;
605 return lost_fraction;
609 APFloat::significandMSB() const
611 return APInt::tcMSB(significandParts(), partCount());
615 APFloat::significandLSB() const
617 return APInt::tcLSB(significandParts(), partCount());
620 /* Note that a zero result is NOT normalized to fcZero. */
622 APFloat::shiftSignificandRight(unsigned int bits)
624 /* Our exponent should not overflow. */
625 assert((exponent_t) (exponent + bits) >= exponent);
629 return shiftRight(significandParts(), partCount(), bits);
632 /* Shift the significand left BITS bits, subtract BITS from its exponent. */
634 APFloat::shiftSignificandLeft(unsigned int bits)
636 assert(bits < semantics->precision);
639 unsigned int partsCount = partCount();
641 APInt::tcShiftLeft(significandParts(), partsCount, bits);
644 assert(!APInt::tcIsZero(significandParts(), partsCount));
649 APFloat::compareAbsoluteValue(const APFloat &rhs) const
653 assert(semantics == rhs.semantics);
654 assert(category == fcNormal);
655 assert(rhs.category == fcNormal);
657 compare = exponent - rhs.exponent;
659 /* If exponents are equal, do an unsigned bignum comparison of the
662 compare = APInt::tcCompare(significandParts(), rhs.significandParts(),
666 return cmpGreaterThan;
673 /* Handle overflow. Sign is preserved. We either become infinity or
674 the largest finite number. */
676 APFloat::handleOverflow(roundingMode rounding_mode)
679 if(rounding_mode == rmNearestTiesToEven
680 || rounding_mode == rmNearestTiesToAway
681 || (rounding_mode == rmTowardPositive && !sign)
682 || (rounding_mode == rmTowardNegative && sign))
684 category = fcInfinity;
685 return (opStatus) (opOverflow | opInexact);
688 /* Otherwise we become the largest finite number. */
690 exponent = semantics->maxExponent;
691 APInt::tcSetLeastSignificantBits(significandParts(), partCount(),
692 semantics->precision);
697 /* This routine must work for fcZero of both signs, and fcNormal
700 APFloat::roundAwayFromZero(roundingMode rounding_mode,
701 lostFraction lost_fraction)
703 /* NaNs and infinities should not have lost fractions. */
704 assert(category == fcNormal || category == fcZero);
706 /* Our caller has already handled this case. */
707 assert(lost_fraction != lfExactlyZero);
709 switch(rounding_mode) {
713 case rmNearestTiesToAway:
714 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf;
716 case rmNearestTiesToEven:
717 if(lost_fraction == lfMoreThanHalf)
720 /* Our zeroes don't have a significand to test. */
721 if(lost_fraction == lfExactlyHalf && category != fcZero)
722 return significandParts()[0] & 1;
729 case rmTowardPositive:
730 return sign == false;
732 case rmTowardNegative:
738 APFloat::normalize(roundingMode rounding_mode,
739 lostFraction lost_fraction)
741 unsigned int omsb; /* One, not zero, based MSB. */
744 if(category != fcNormal)
747 /* Before rounding normalize the exponent of fcNormal numbers. */
748 omsb = significandMSB() + 1;
751 /* OMSB is numbered from 1. We want to place it in the integer
752 bit numbered PRECISON if possible, with a compensating change in
754 exponentChange = omsb - semantics->precision;
756 /* If the resulting exponent is too high, overflow according to
757 the rounding mode. */
758 if(exponent + exponentChange > semantics->maxExponent)
759 return handleOverflow(rounding_mode);
761 /* Subnormal numbers have exponent minExponent, and their MSB
762 is forced based on that. */
763 if(exponent + exponentChange < semantics->minExponent)
764 exponentChange = semantics->minExponent - exponent;
766 /* Shifting left is easy as we don't lose precision. */
767 if(exponentChange < 0) {
768 assert(lost_fraction == lfExactlyZero);
770 shiftSignificandLeft(-exponentChange);
775 if(exponentChange > 0) {
778 /* Shift right and capture any new lost fraction. */
779 lf = shiftSignificandRight(exponentChange);
781 lost_fraction = combineLostFractions(lf, lost_fraction);
783 /* Keep OMSB up-to-date. */
784 if(omsb > (unsigned) exponentChange)
785 omsb -= (unsigned) exponentChange;
791 /* Now round the number according to rounding_mode given the lost
794 /* As specified in IEEE 754, since we do not trap we do not report
795 underflow for exact results. */
796 if(lost_fraction == lfExactlyZero) {
797 /* Canonicalize zeroes. */
804 /* Increment the significand if we're rounding away from zero. */
805 if(roundAwayFromZero(rounding_mode, lost_fraction)) {
807 exponent = semantics->minExponent;
809 incrementSignificand();
810 omsb = significandMSB() + 1;
812 /* Did the significand increment overflow? */
813 if(omsb == (unsigned) semantics->precision + 1) {
814 /* Renormalize by incrementing the exponent and shifting our
815 significand right one. However if we already have the
816 maximum exponent we overflow to infinity. */
817 if(exponent == semantics->maxExponent) {
818 category = fcInfinity;
820 return (opStatus) (opOverflow | opInexact);
823 shiftSignificandRight(1);
829 /* The normal case - we were and are not denormal, and any
830 significand increment above didn't overflow. */
831 if(omsb == semantics->precision)
834 /* We have a non-zero denormal. */
835 assert(omsb < semantics->precision);
836 assert(exponent == semantics->minExponent);
838 /* Canonicalize zeroes. */
842 /* The fcZero case is a denormal that underflowed to zero. */
843 return (opStatus) (opUnderflow | opInexact);
847 APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract)
849 switch(convolve(category, rhs.category)) {
853 case convolve(fcNaN, fcZero):
854 case convolve(fcNaN, fcNormal):
855 case convolve(fcNaN, fcInfinity):
856 case convolve(fcNaN, fcNaN):
857 case convolve(fcNormal, fcZero):
858 case convolve(fcInfinity, fcNormal):
859 case convolve(fcInfinity, fcZero):
862 case convolve(fcZero, fcNaN):
863 case convolve(fcNormal, fcNaN):
864 case convolve(fcInfinity, fcNaN):
866 copySignificand(rhs);
869 case convolve(fcNormal, fcInfinity):
870 case convolve(fcZero, fcInfinity):
871 category = fcInfinity;
872 sign = rhs.sign ^ subtract;
875 case convolve(fcZero, fcNormal):
877 sign = rhs.sign ^ subtract;
880 case convolve(fcZero, fcZero):
881 /* Sign depends on rounding mode; handled by caller. */
884 case convolve(fcInfinity, fcInfinity):
885 /* Differently signed infinities can only be validly
887 if(sign ^ rhs.sign != subtract) {
889 // Arbitrary but deterministic value for significand
890 APInt::tcSet(significandParts(), ~0U, partCount());
896 case convolve(fcNormal, fcNormal):
901 /* Add or subtract two normal numbers. */
903 APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract)
906 lostFraction lost_fraction;
909 /* Determine if the operation on the absolute values is effectively
910 an addition or subtraction. */
911 subtract ^= (sign ^ rhs.sign);
913 /* Are we bigger exponent-wise than the RHS? */
914 bits = exponent - rhs.exponent;
916 /* Subtraction is more subtle than one might naively expect. */
918 APFloat temp_rhs(rhs);
922 reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan;
923 lost_fraction = lfExactlyZero;
924 } else if (bits > 0) {
925 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1);
926 shiftSignificandLeft(1);
929 lost_fraction = shiftSignificandRight(-bits - 1);
930 temp_rhs.shiftSignificandLeft(1);
935 carry = temp_rhs.subtractSignificand
936 (*this, lost_fraction != lfExactlyZero);
937 copySignificand(temp_rhs);
940 carry = subtractSignificand
941 (temp_rhs, lost_fraction != lfExactlyZero);
944 /* Invert the lost fraction - it was on the RHS and
946 if(lost_fraction == lfLessThanHalf)
947 lost_fraction = lfMoreThanHalf;
948 else if(lost_fraction == lfMoreThanHalf)
949 lost_fraction = lfLessThanHalf;
951 /* The code above is intended to ensure that no borrow is
956 APFloat temp_rhs(rhs);
958 lost_fraction = temp_rhs.shiftSignificandRight(bits);
959 carry = addSignificand(temp_rhs);
961 lost_fraction = shiftSignificandRight(-bits);
962 carry = addSignificand(rhs);
965 /* We have a guard bit; generating a carry cannot happen. */
969 return lost_fraction;
973 APFloat::multiplySpecials(const APFloat &rhs)
975 switch(convolve(category, rhs.category)) {
979 case convolve(fcNaN, fcZero):
980 case convolve(fcNaN, fcNormal):
981 case convolve(fcNaN, fcInfinity):
982 case convolve(fcNaN, fcNaN):
985 case convolve(fcZero, fcNaN):
986 case convolve(fcNormal, fcNaN):
987 case convolve(fcInfinity, fcNaN):
989 copySignificand(rhs);
992 case convolve(fcNormal, fcInfinity):
993 case convolve(fcInfinity, fcNormal):
994 case convolve(fcInfinity, fcInfinity):
995 category = fcInfinity;
998 case convolve(fcZero, fcNormal):
999 case convolve(fcNormal, fcZero):
1000 case convolve(fcZero, fcZero):
1004 case convolve(fcZero, fcInfinity):
1005 case convolve(fcInfinity, fcZero):
1007 // Arbitrary but deterministic value for significand
1008 APInt::tcSet(significandParts(), ~0U, partCount());
1011 case convolve(fcNormal, fcNormal):
1017 APFloat::divideSpecials(const APFloat &rhs)
1019 switch(convolve(category, rhs.category)) {
1023 case convolve(fcNaN, fcZero):
1024 case convolve(fcNaN, fcNormal):
1025 case convolve(fcNaN, fcInfinity):
1026 case convolve(fcNaN, fcNaN):
1027 case convolve(fcInfinity, fcZero):
1028 case convolve(fcInfinity, fcNormal):
1029 case convolve(fcZero, fcInfinity):
1030 case convolve(fcZero, fcNormal):
1033 case convolve(fcZero, fcNaN):
1034 case convolve(fcNormal, fcNaN):
1035 case convolve(fcInfinity, fcNaN):
1037 copySignificand(rhs);
1040 case convolve(fcNormal, fcInfinity):
1044 case convolve(fcNormal, fcZero):
1045 category = fcInfinity;
1048 case convolve(fcInfinity, fcInfinity):
1049 case convolve(fcZero, fcZero):
1051 // Arbitrary but deterministic value for significand
1052 APInt::tcSet(significandParts(), ~0U, partCount());
1055 case convolve(fcNormal, fcNormal):
1062 APFloat::changeSign()
1064 /* Look mummy, this one's easy. */
1068 /* Normalized addition or subtraction. */
1070 APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode,
1075 fs = addOrSubtractSpecials(rhs, subtract);
1077 /* This return code means it was not a simple case. */
1078 if(fs == opDivByZero) {
1079 lostFraction lost_fraction;
1081 lost_fraction = addOrSubtractSignificand(rhs, subtract);
1082 fs = normalize(rounding_mode, lost_fraction);
1084 /* Can only be zero if we lost no fraction. */
1085 assert(category != fcZero || lost_fraction == lfExactlyZero);
1088 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1089 positive zero unless rounding to minus infinity, except that
1090 adding two like-signed zeroes gives that zero. */
1091 if(category == fcZero) {
1092 if(rhs.category != fcZero || (sign == rhs.sign) == subtract)
1093 sign = (rounding_mode == rmTowardNegative);
1099 /* Normalized addition. */
1101 APFloat::add(const APFloat &rhs, roundingMode rounding_mode)
1103 return addOrSubtract(rhs, rounding_mode, false);
1106 /* Normalized subtraction. */
1108 APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode)
1110 return addOrSubtract(rhs, rounding_mode, true);
1113 /* Normalized multiply. */
1115 APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode)
1120 fs = multiplySpecials(rhs);
1122 if(category == fcNormal) {
1123 lostFraction lost_fraction = multiplySignificand(rhs, 0);
1124 fs = normalize(rounding_mode, lost_fraction);
1125 if(lost_fraction != lfExactlyZero)
1126 fs = (opStatus) (fs | opInexact);
1132 /* Normalized divide. */
1134 APFloat::divide(const APFloat &rhs, roundingMode rounding_mode)
1139 fs = divideSpecials(rhs);
1141 if(category == fcNormal) {
1142 lostFraction lost_fraction = divideSignificand(rhs);
1143 fs = normalize(rounding_mode, lost_fraction);
1144 if(lost_fraction != lfExactlyZero)
1145 fs = (opStatus) (fs | opInexact);
1151 /* Normalized fused-multiply-add. */
1153 APFloat::fusedMultiplyAdd(const APFloat &multiplicand,
1154 const APFloat &addend,
1155 roundingMode rounding_mode)
1159 /* Post-multiplication sign, before addition. */
1160 sign ^= multiplicand.sign;
1162 /* If and only if all arguments are normal do we need to do an
1163 extended-precision calculation. */
1164 if(category == fcNormal
1165 && multiplicand.category == fcNormal
1166 && addend.category == fcNormal) {
1167 lostFraction lost_fraction;
1169 lost_fraction = multiplySignificand(multiplicand, &addend);
1170 fs = normalize(rounding_mode, lost_fraction);
1171 if(lost_fraction != lfExactlyZero)
1172 fs = (opStatus) (fs | opInexact);
1174 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1175 positive zero unless rounding to minus infinity, except that
1176 adding two like-signed zeroes gives that zero. */
1177 if(category == fcZero && sign != addend.sign)
1178 sign = (rounding_mode == rmTowardNegative);
1180 fs = multiplySpecials(multiplicand);
1182 /* FS can only be opOK or opInvalidOp. There is no more work
1183 to do in the latter case. The IEEE-754R standard says it is
1184 implementation-defined in this case whether, if ADDEND is a
1185 quiet NaN, we raise invalid op; this implementation does so.
1187 If we need to do the addition we can do so with normal
1190 fs = addOrSubtract(addend, rounding_mode, false);
1196 /* Comparison requires normalized numbers. */
1198 APFloat::compare(const APFloat &rhs) const
1202 assert(semantics == rhs.semantics);
1204 switch(convolve(category, rhs.category)) {
1208 case convolve(fcNaN, fcZero):
1209 case convolve(fcNaN, fcNormal):
1210 case convolve(fcNaN, fcInfinity):
1211 case convolve(fcNaN, fcNaN):
1212 case convolve(fcZero, fcNaN):
1213 case convolve(fcNormal, fcNaN):
1214 case convolve(fcInfinity, fcNaN):
1215 return cmpUnordered;
1217 case convolve(fcInfinity, fcNormal):
1218 case convolve(fcInfinity, fcZero):
1219 case convolve(fcNormal, fcZero):
1223 return cmpGreaterThan;
1225 case convolve(fcNormal, fcInfinity):
1226 case convolve(fcZero, fcInfinity):
1227 case convolve(fcZero, fcNormal):
1229 return cmpGreaterThan;
1233 case convolve(fcInfinity, fcInfinity):
1234 if(sign == rhs.sign)
1239 return cmpGreaterThan;
1241 case convolve(fcZero, fcZero):
1244 case convolve(fcNormal, fcNormal):
1248 /* Two normal numbers. Do they have the same sign? */
1249 if(sign != rhs.sign) {
1251 result = cmpLessThan;
1253 result = cmpGreaterThan;
1255 /* Compare absolute values; invert result if negative. */
1256 result = compareAbsoluteValue(rhs);
1259 if(result == cmpLessThan)
1260 result = cmpGreaterThan;
1261 else if(result == cmpGreaterThan)
1262 result = cmpLessThan;
1270 APFloat::convert(const fltSemantics &toSemantics,
1271 roundingMode rounding_mode)
1273 unsigned int newPartCount;
1276 newPartCount = partCountForBits(toSemantics.precision + 1);
1278 /* If our new form is wider, re-allocate our bit pattern into wider
1280 if(newPartCount > partCount()) {
1281 integerPart *newParts;
1283 newParts = new integerPart[newPartCount];
1284 APInt::tcSet(newParts, 0, newPartCount);
1285 APInt::tcAssign(newParts, significandParts(), partCount());
1287 significand.parts = newParts;
1290 if(category == fcNormal) {
1291 /* Re-interpret our bit-pattern. */
1292 exponent += toSemantics.precision - semantics->precision;
1293 semantics = &toSemantics;
1294 fs = normalize(rounding_mode, lfExactlyZero);
1296 semantics = &toSemantics;
1303 /* Convert a floating point number to an integer according to the
1304 rounding mode. If the rounded integer value is out of range this
1305 returns an invalid operation exception. If the rounded value is in
1306 range but the floating point number is not the exact integer, the C
1307 standard doesn't require an inexact exception to be raised. IEEE
1308 854 does require it so we do that.
1310 Note that for conversions to integer type the C standard requires
1311 round-to-zero to always be used. */
1313 APFloat::convertToInteger(integerPart *parts, unsigned int width,
1315 roundingMode rounding_mode) const
1317 lostFraction lost_fraction;
1318 unsigned int msb, partsCount;
1321 /* Handle the three special cases first. */
1322 if(category == fcInfinity || category == fcNaN)
1325 partsCount = partCountForBits(width);
1327 if(category == fcZero) {
1328 APInt::tcSet(parts, 0, partsCount);
1332 /* Shift the bit pattern so the fraction is lost. */
1335 bits = (int) semantics->precision - 1 - exponent;
1338 lost_fraction = tmp.shiftSignificandRight(bits);
1340 tmp.shiftSignificandLeft(-bits);
1341 lost_fraction = lfExactlyZero;
1344 if(lost_fraction != lfExactlyZero
1345 && tmp.roundAwayFromZero(rounding_mode, lost_fraction))
1346 tmp.incrementSignificand();
1348 msb = tmp.significandMSB();
1350 /* Negative numbers cannot be represented as unsigned. */
1351 if(!isSigned && tmp.sign && msb != -1U)
1354 /* It takes exponent + 1 bits to represent the truncated floating
1355 point number without its sign. We lose a bit for the sign, but
1356 the maximally negative integer is a special case. */
1357 if(msb + 1 > width) /* !! Not same as msb >= width !! */
1360 if(isSigned && msb + 1 == width
1361 && (!tmp.sign || tmp.significandLSB() != msb))
1364 APInt::tcAssign(parts, tmp.significandParts(), partsCount);
1367 APInt::tcNegate(parts, partsCount);
1369 if(lost_fraction == lfExactlyZero)
1376 APFloat::convertFromUnsignedInteger(integerPart *parts,
1377 unsigned int partCount,
1378 roundingMode rounding_mode)
1380 unsigned int msb, precision;
1381 lostFraction lost_fraction;
1383 msb = APInt::tcMSB(parts, partCount) + 1;
1384 precision = semantics->precision;
1386 category = fcNormal;
1387 exponent = precision - 1;
1389 if(msb > precision) {
1390 exponent += (msb - precision);
1391 lost_fraction = shiftRight(parts, partCount, msb - precision);
1394 lost_fraction = lfExactlyZero;
1396 /* Copy the bit image. */
1398 APInt::tcAssign(significandParts(), parts, partCountForBits(msb));
1400 return normalize(rounding_mode, lost_fraction);
1404 APFloat::convertFromInteger(const integerPart *parts,
1405 unsigned int partCount, bool isSigned,
1406 roundingMode rounding_mode)
1412 copy = new integerPart[partCount];
1413 APInt::tcAssign(copy, parts, partCount);
1415 width = partCount * integerPartWidth;
1418 if(isSigned && APInt::tcExtractBit(parts, width - 1)) {
1420 APInt::tcNegate(copy, partCount);
1423 status = convertFromUnsignedInteger(copy, partCount, rounding_mode);
1430 APFloat::convertFromHexadecimalString(const char *p,
1431 roundingMode rounding_mode)
1433 lostFraction lost_fraction;
1434 integerPart *significand;
1435 unsigned int bitPos, partsCount;
1436 const char *dot, *firstSignificantDigit;
1440 category = fcNormal;
1442 significand = significandParts();
1443 partsCount = partCount();
1444 bitPos = partsCount * integerPartWidth;
1446 /* Skip leading zeroes and any(hexa)decimal point. */
1447 p = skipLeadingZeroesAndAnyDot(p, &dot);
1448 firstSignificantDigit = p;
1451 integerPart hex_value;
1458 hex_value = hexDigitValue(*p);
1459 if(hex_value == -1U) {
1460 lost_fraction = lfExactlyZero;
1466 /* Store the number whilst 4-bit nibbles remain. */
1469 hex_value <<= bitPos % integerPartWidth;
1470 significand[bitPos / integerPartWidth] |= hex_value;
1472 lost_fraction = trailingHexadecimalFraction(p, hex_value);
1473 while(hexDigitValue(*p) != -1U)
1479 /* Hex floats require an exponent but not a hexadecimal point. */
1480 assert(*p == 'p' || *p == 'P');
1482 /* Ignore the exponent if we are zero. */
1483 if(p != firstSignificantDigit) {
1486 /* Implicit hexadecimal point? */
1490 /* Calculate the exponent adjustment implicit in the number of
1491 significant digits. */
1492 expAdjustment = dot - firstSignificantDigit;
1493 if(expAdjustment < 0)
1495 expAdjustment = expAdjustment * 4 - 1;
1497 /* Adjust for writing the significand starting at the most
1498 significant nibble. */
1499 expAdjustment += semantics->precision;
1500 expAdjustment -= partsCount * integerPartWidth;
1502 /* Adjust for the given exponent. */
1503 exponent = totalExponent(p, expAdjustment);
1506 return normalize(rounding_mode, lost_fraction);
1510 APFloat::convertFromString(const char *p, roundingMode rounding_mode) {
1511 /* Handle a leading minus sign. */
1517 if(p[0] == '0' && (p[1] == 'x' || p[1] == 'X'))
1518 return convertFromHexadecimalString(p + 2, rounding_mode);
1520 assert(0 && "Decimal to binary conversions not yet implemented");
1524 // For good performance it is desirable for different APFloats
1525 // to produce different integers.
1527 APFloat::getHashValue() const {
1528 if (category==fcZero) return sign<<8 | semantics->precision ;
1529 else if (category==fcInfinity) return sign<<9 | semantics->precision;
1530 else if (category==fcNaN) return 1<<10 | semantics->precision;
1532 uint32_t hash = sign<<11 | semantics->precision | exponent<<12;
1533 const integerPart* p = significandParts();
1534 for (int i=partCount(); i>0; i--, p++)
1535 hash ^= ((uint32_t)*p) ^ (*p)>>32;
1540 // Conversion from APFloat to/from host float/double. It may eventually be
1541 // possible to eliminate these and have everybody deal with APFloats, but that
1542 // will take a while. This approach will not easily extend to long double.
1543 // Current implementation requires partCount()==1, which is correct at the
1544 // moment but could be made more general.
1547 APFloat::convertToDouble() const {
1548 assert(semantics == (const llvm::fltSemantics* const)&IEEEdouble);
1549 assert (partCount()==1);
1551 uint64_t myexponent, mysignificand;
1553 if (category==fcNormal) {
1554 mysignificand = *significandParts();
1555 myexponent = exponent+1023; //bias
1556 } else if (category==fcZero) {
1559 } else if (category==fcInfinity) {
1562 } else if (category==fcNaN) {
1564 mysignificand = *significandParts();
1568 return BitsToDouble((((uint64_t)sign & 1) << 63) |
1569 ((myexponent & 0x7ff) << 52) |
1570 (mysignificand & 0xfffffffffffffLL));
1574 APFloat::convertToFloat() const {
1575 assert(semantics == (const llvm::fltSemantics* const)&IEEEsingle);
1576 assert (partCount()==1);
1578 uint32_t myexponent, mysignificand;
1580 if (category==fcNormal) {
1581 myexponent = exponent+127; //bias
1582 mysignificand = *significandParts();
1583 } else if (category==fcZero) {
1586 } else if (category==fcInfinity) {
1589 } else if (category==fcNaN) {
1591 mysignificand = *significandParts();
1595 return BitsToFloat(((sign&1) << 31) | ((myexponent&0xff) << 23) |
1596 (mysignificand & 0x7fffff));
1599 APFloat::APFloat(double d) {
1600 uint64_t i = DoubleToBits(d);
1601 uint64_t myexponent = (i >> 52) & 0x7ff;
1602 uint64_t mysignificand = i & 0xfffffffffffffLL;
1604 initialize(&APFloat::IEEEdouble);
1605 assert(partCount()==1);
1608 if (myexponent==0 && mysignificand==0) {
1609 // exponent, significand meaningless
1611 } else if (myexponent==0x7ff && mysignificand==0) {
1612 // exponent, significand meaningless
1613 category = fcInfinity;
1614 } else if (myexponent==0x7ff && mysignificand!=0) {
1615 // exponent meaningless
1617 *significandParts() = mysignificand;
1619 category = fcNormal;
1620 exponent = myexponent - 1023;
1621 *significandParts() = mysignificand | 0x10000000000000LL;
1625 APFloat::APFloat(float f) {
1626 uint32_t i = FloatToBits(f);
1627 uint32_t myexponent = (i >> 23) & 0xff;
1628 uint32_t mysignificand = i & 0x7fffff;
1630 initialize(&APFloat::IEEEsingle);
1631 assert(partCount()==1);
1634 if (myexponent==0 && mysignificand==0) {
1635 // exponent, significand meaningless
1637 } else if (myexponent==0xff && mysignificand==0) {
1638 // exponent, significand meaningless
1639 category = fcInfinity;
1640 } else if (myexponent==0xff && (mysignificand & 0x400000)) {
1641 // sign, exponent, significand meaningless
1643 *significandParts() = mysignificand;
1645 category = fcNormal;
1646 exponent = myexponent - 127; //bias
1647 *significandParts() = mysignificand | 0x800000; // integer bit