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 //===----------------------------------------------------------------------===//
17 #include "llvm/ADT/APFloat.h"
18 #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 COMPILE_TIME_ASSERT(integerPartWidth % 4 == 0);
30 /* Represents floating point arithmetic semantics. */
32 /* The largest E such that 2^E is representable; this matches the
33 definition of IEEE 754. */
34 exponent_t maxExponent;
36 /* The smallest E such that 2^E is a normalized number; this
37 matches the definition of IEEE 754. */
38 exponent_t minExponent;
40 /* Number of bits in the significand. This includes the integer
42 unsigned char precision;
44 /* If the target format has an implicit integer bit. */
45 bool implicitIntegerBit;
48 const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, true };
49 const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, true };
50 const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, true };
51 const fltSemantics APFloat::x87DoubleExtended = { 16383, -16382, 64, false };
52 const fltSemantics APFloat::Bogus = { 0, 0, 0, false };
55 /* Put a bunch of private, handy routines in an anonymous namespace. */
59 partCountForBits(unsigned int bits)
61 return ((bits) + integerPartWidth - 1) / integerPartWidth;
65 digitValue(unsigned int c)
77 hexDigitValue (unsigned int c)
96 /* This is ugly and needs cleaning up, but I don't immediately see
97 how whilst remaining safe. */
99 totalExponent(const char *p, int exponentAdjustment)
101 integerPart unsignedExponent;
102 bool negative, overflow;
105 /* Move past the exponent letter and sign to the digits. */
107 negative = *p == '-';
108 if(*p == '-' || *p == '+')
111 unsignedExponent = 0;
116 value = digitValue(*p);
121 unsignedExponent = unsignedExponent * 10 + value;
122 if(unsignedExponent > 65535)
126 if(exponentAdjustment > 65535 || exponentAdjustment < -65536)
130 exponent = unsignedExponent;
132 exponent = -exponent;
133 exponent += exponentAdjustment;
134 if(exponent > 65535 || exponent < -65536)
139 exponent = negative ? -65536: 65535;
145 skipLeadingZeroesAndAnyDot(const char *p, const char **dot)
160 /* Return the trailing fraction of a hexadecimal number.
161 DIGITVALUE is the first hex digit of the fraction, P points to
164 trailingHexadecimalFraction(const char *p, unsigned int digitValue)
166 unsigned int hexDigit;
168 /* If the first trailing digit isn't 0 or 8 we can work out the
169 fraction immediately. */
171 return lfMoreThanHalf;
172 else if(digitValue < 8 && digitValue > 0)
173 return lfLessThanHalf;
175 /* Otherwise we need to find the first non-zero digit. */
179 hexDigit = hexDigitValue(*p);
181 /* If we ran off the end it is exactly zero or one-half, otherwise
184 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf;
186 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf;
189 /* Return the fraction lost were a bignum truncated losing the least
190 significant BITS bits. */
192 lostFractionThroughTruncation(const integerPart *parts,
193 unsigned int partCount,
198 lsb = APInt::tcLSB(parts, partCount);
200 /* Note this is guaranteed true if bits == 0, or LSB == -1U. */
202 return lfExactlyZero;
204 return lfExactlyHalf;
205 if(bits <= partCount * integerPartWidth
206 && APInt::tcExtractBit(parts, bits - 1))
207 return lfMoreThanHalf;
209 return lfLessThanHalf;
212 /* Shift DST right BITS bits noting lost fraction. */
214 shiftRight(integerPart *dst, unsigned int parts, unsigned int bits)
216 lostFraction lost_fraction;
218 lost_fraction = lostFractionThroughTruncation(dst, parts, bits);
220 APInt::tcShiftRight(dst, parts, bits);
222 return lost_fraction;
225 /* Combine the effect of two lost fractions. */
227 combineLostFractions(lostFraction moreSignificant,
228 lostFraction lessSignificant)
230 if(lessSignificant != lfExactlyZero) {
231 if(moreSignificant == lfExactlyZero)
232 moreSignificant = lfLessThanHalf;
233 else if(moreSignificant == lfExactlyHalf)
234 moreSignificant = lfMoreThanHalf;
237 return moreSignificant;
240 /* Zero at the end to avoid modular arithmetic when adding one; used
241 when rounding up during hexadecimal output. */
242 static const char hexDigitsLower[] = "0123456789abcdef0";
243 static const char hexDigitsUpper[] = "0123456789ABCDEF0";
244 static const char infinityL[] = "infinity";
245 static const char infinityU[] = "INFINITY";
246 static const char NaNL[] = "nan";
247 static const char NaNU[] = "NAN";
249 /* Write out an integerPart in hexadecimal, starting with the most
250 significant nibble. Write out exactly COUNT hexdigits, return
253 partAsHex (char *dst, integerPart part, unsigned int count,
254 const char *hexDigitChars)
256 unsigned int result = count;
258 assert (count != 0 && count <= integerPartWidth / 4);
260 part >>= (integerPartWidth - 4 * count);
262 dst[count] = hexDigitChars[part & 0xf];
269 /* Write out an unsigned decimal integer. */
271 writeUnsignedDecimal (char *dst, unsigned int n)
287 /* Write out a signed decimal integer. */
289 writeSignedDecimal (char *dst, int value)
293 dst = writeUnsignedDecimal(dst, -(unsigned) value);
295 dst = writeUnsignedDecimal(dst, value);
303 APFloat::initialize(const fltSemantics *ourSemantics)
307 semantics = ourSemantics;
310 significand.parts = new integerPart[count];
314 APFloat::freeSignificand()
317 delete [] significand.parts;
321 APFloat::assign(const APFloat &rhs)
323 assert(semantics == rhs.semantics);
326 category = rhs.category;
327 exponent = rhs.exponent;
328 if(category == fcNormal || category == fcNaN)
329 copySignificand(rhs);
333 APFloat::copySignificand(const APFloat &rhs)
335 assert(category == fcNormal || category == fcNaN);
336 assert(rhs.partCount() >= partCount());
338 APInt::tcAssign(significandParts(), rhs.significandParts(),
343 APFloat::operator=(const APFloat &rhs)
346 if(semantics != rhs.semantics) {
348 initialize(rhs.semantics);
357 APFloat::bitwiseIsEqual(const APFloat &rhs) const {
360 if (semantics != rhs.semantics ||
361 category != rhs.category ||
364 if (category==fcZero || category==fcInfinity)
366 else if (category==fcNormal && exponent!=rhs.exponent)
370 const integerPart* p=significandParts();
371 const integerPart* q=rhs.significandParts();
372 for (; i>0; i--, p++, q++) {
380 APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value)
382 initialize(&ourSemantics);
385 exponent = ourSemantics.precision - 1;
386 significandParts()[0] = value;
387 normalize(rmNearestTiesToEven, lfExactlyZero);
390 APFloat::APFloat(const fltSemantics &ourSemantics,
391 fltCategory ourCategory, bool negative)
393 initialize(&ourSemantics);
394 category = ourCategory;
396 if(category == fcNormal)
400 APFloat::APFloat(const fltSemantics &ourSemantics, const char *text)
402 initialize(&ourSemantics);
403 convertFromString(text, rmNearestTiesToEven);
406 APFloat::APFloat(const APFloat &rhs)
408 initialize(rhs.semantics);
418 APFloat::partCount() const
420 return partCountForBits(semantics->precision + 1);
424 APFloat::semanticsPrecision(const fltSemantics &semantics)
426 return semantics.precision;
430 APFloat::significandParts() const
432 return const_cast<APFloat *>(this)->significandParts();
436 APFloat::significandParts()
438 assert(category == fcNormal || category == fcNaN);
441 return significand.parts;
443 return &significand.part;
447 APFloat::zeroSignificand()
450 APInt::tcSet(significandParts(), 0, partCount());
453 /* Increment an fcNormal floating point number's significand. */
455 APFloat::incrementSignificand()
459 carry = APInt::tcIncrement(significandParts(), partCount());
461 /* Our callers should never cause us to overflow. */
465 /* Add the significand of the RHS. Returns the carry flag. */
467 APFloat::addSignificand(const APFloat &rhs)
471 parts = significandParts();
473 assert(semantics == rhs.semantics);
474 assert(exponent == rhs.exponent);
476 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount());
479 /* Subtract the significand of the RHS with a borrow flag. Returns
482 APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow)
486 parts = significandParts();
488 assert(semantics == rhs.semantics);
489 assert(exponent == rhs.exponent);
491 return APInt::tcSubtract(parts, rhs.significandParts(), borrow,
495 /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it
496 on to the full-precision result of the multiplication. Returns the
499 APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend)
501 unsigned int omsb; // One, not zero, based MSB.
502 unsigned int partsCount, newPartsCount, precision;
503 integerPart *lhsSignificand;
504 integerPart scratch[4];
505 integerPart *fullSignificand;
506 lostFraction lost_fraction;
508 assert(semantics == rhs.semantics);
510 precision = semantics->precision;
511 newPartsCount = partCountForBits(precision * 2);
513 if(newPartsCount > 4)
514 fullSignificand = new integerPart[newPartsCount];
516 fullSignificand = scratch;
518 lhsSignificand = significandParts();
519 partsCount = partCount();
521 APInt::tcFullMultiply(fullSignificand, lhsSignificand,
522 rhs.significandParts(), partsCount, partsCount);
524 lost_fraction = lfExactlyZero;
525 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
526 exponent += rhs.exponent;
529 Significand savedSignificand = significand;
530 const fltSemantics *savedSemantics = semantics;
531 fltSemantics extendedSemantics;
533 unsigned int extendedPrecision;
535 /* Normalize our MSB. */
536 extendedPrecision = precision + precision - 1;
537 if(omsb != extendedPrecision)
539 APInt::tcShiftLeft(fullSignificand, newPartsCount,
540 extendedPrecision - omsb);
541 exponent -= extendedPrecision - omsb;
544 /* Create new semantics. */
545 extendedSemantics = *semantics;
546 extendedSemantics.precision = extendedPrecision;
548 if(newPartsCount == 1)
549 significand.part = fullSignificand[0];
551 significand.parts = fullSignificand;
552 semantics = &extendedSemantics;
554 APFloat extendedAddend(*addend);
555 status = extendedAddend.convert(extendedSemantics, rmTowardZero);
556 assert(status == opOK);
557 lost_fraction = addOrSubtractSignificand(extendedAddend, false);
559 /* Restore our state. */
560 if(newPartsCount == 1)
561 fullSignificand[0] = significand.part;
562 significand = savedSignificand;
563 semantics = savedSemantics;
565 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
568 exponent -= (precision - 1);
570 if(omsb > precision) {
571 unsigned int bits, significantParts;
574 bits = omsb - precision;
575 significantParts = partCountForBits(omsb);
576 lf = shiftRight(fullSignificand, significantParts, bits);
577 lost_fraction = combineLostFractions(lf, lost_fraction);
581 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount);
583 if(newPartsCount > 4)
584 delete [] fullSignificand;
586 return lost_fraction;
589 /* Multiply the significands of LHS and RHS to DST. */
591 APFloat::divideSignificand(const APFloat &rhs)
593 unsigned int bit, i, partsCount;
594 const integerPart *rhsSignificand;
595 integerPart *lhsSignificand, *dividend, *divisor;
596 integerPart scratch[4];
597 lostFraction lost_fraction;
599 assert(semantics == rhs.semantics);
601 lhsSignificand = significandParts();
602 rhsSignificand = rhs.significandParts();
603 partsCount = partCount();
606 dividend = new integerPart[partsCount * 2];
610 divisor = dividend + partsCount;
612 /* Copy the dividend and divisor as they will be modified in-place. */
613 for(i = 0; i < partsCount; i++) {
614 dividend[i] = lhsSignificand[i];
615 divisor[i] = rhsSignificand[i];
616 lhsSignificand[i] = 0;
619 exponent -= rhs.exponent;
621 unsigned int precision = semantics->precision;
623 /* Normalize the divisor. */
624 bit = precision - APInt::tcMSB(divisor, partsCount) - 1;
627 APInt::tcShiftLeft(divisor, partsCount, bit);
630 /* Normalize the dividend. */
631 bit = precision - APInt::tcMSB(dividend, partsCount) - 1;
634 APInt::tcShiftLeft(dividend, partsCount, bit);
637 if(APInt::tcCompare(dividend, divisor, partsCount) < 0) {
639 APInt::tcShiftLeft(dividend, partsCount, 1);
640 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0);
644 for(bit = precision; bit; bit -= 1) {
645 if(APInt::tcCompare(dividend, divisor, partsCount) >= 0) {
646 APInt::tcSubtract(dividend, divisor, 0, partsCount);
647 APInt::tcSetBit(lhsSignificand, bit - 1);
650 APInt::tcShiftLeft(dividend, partsCount, 1);
653 /* Figure out the lost fraction. */
654 int cmp = APInt::tcCompare(dividend, divisor, partsCount);
657 lost_fraction = lfMoreThanHalf;
659 lost_fraction = lfExactlyHalf;
660 else if(APInt::tcIsZero(dividend, partsCount))
661 lost_fraction = lfExactlyZero;
663 lost_fraction = lfLessThanHalf;
668 return lost_fraction;
672 APFloat::significandMSB() const
674 return APInt::tcMSB(significandParts(), partCount());
678 APFloat::significandLSB() const
680 return APInt::tcLSB(significandParts(), partCount());
683 /* Note that a zero result is NOT normalized to fcZero. */
685 APFloat::shiftSignificandRight(unsigned int bits)
687 /* Our exponent should not overflow. */
688 assert((exponent_t) (exponent + bits) >= exponent);
692 return shiftRight(significandParts(), partCount(), bits);
695 /* Shift the significand left BITS bits, subtract BITS from its exponent. */
697 APFloat::shiftSignificandLeft(unsigned int bits)
699 assert(bits < semantics->precision);
702 unsigned int partsCount = partCount();
704 APInt::tcShiftLeft(significandParts(), partsCount, bits);
707 assert(!APInt::tcIsZero(significandParts(), partsCount));
712 APFloat::compareAbsoluteValue(const APFloat &rhs) const
716 assert(semantics == rhs.semantics);
717 assert(category == fcNormal);
718 assert(rhs.category == fcNormal);
720 compare = exponent - rhs.exponent;
722 /* If exponents are equal, do an unsigned bignum comparison of the
725 compare = APInt::tcCompare(significandParts(), rhs.significandParts(),
729 return cmpGreaterThan;
736 /* Handle overflow. Sign is preserved. We either become infinity or
737 the largest finite number. */
739 APFloat::handleOverflow(roundingMode rounding_mode)
742 if(rounding_mode == rmNearestTiesToEven
743 || rounding_mode == rmNearestTiesToAway
744 || (rounding_mode == rmTowardPositive && !sign)
745 || (rounding_mode == rmTowardNegative && sign))
747 category = fcInfinity;
748 return (opStatus) (opOverflow | opInexact);
751 /* Otherwise we become the largest finite number. */
753 exponent = semantics->maxExponent;
754 APInt::tcSetLeastSignificantBits(significandParts(), partCount(),
755 semantics->precision);
760 /* Returns TRUE if, when truncating the current number, with BIT the
761 new LSB, with the given lost fraction and rounding mode, the result
762 would need to be rounded away from zero (i.e., by increasing the
763 signficand). This routine must work for fcZero of both signs, and
766 APFloat::roundAwayFromZero(roundingMode rounding_mode,
767 lostFraction lost_fraction,
768 unsigned int bit) const
770 /* NaNs and infinities should not have lost fractions. */
771 assert(category == fcNormal || category == fcZero);
773 /* Current callers never pass this so we don't handle it. */
774 assert(lost_fraction != lfExactlyZero);
776 switch(rounding_mode) {
780 case rmNearestTiesToAway:
781 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf;
783 case rmNearestTiesToEven:
784 if(lost_fraction == lfMoreThanHalf)
787 /* Our zeroes don't have a significand to test. */
788 if(lost_fraction == lfExactlyHalf && category != fcZero)
789 return APInt::tcExtractBit(significandParts(), bit);
796 case rmTowardPositive:
797 return sign == false;
799 case rmTowardNegative:
805 APFloat::normalize(roundingMode rounding_mode,
806 lostFraction lost_fraction)
808 unsigned int omsb; /* One, not zero, based MSB. */
811 if(category != fcNormal)
814 /* Before rounding normalize the exponent of fcNormal numbers. */
815 omsb = significandMSB() + 1;
818 /* OMSB is numbered from 1. We want to place it in the integer
819 bit numbered PRECISON if possible, with a compensating change in
821 exponentChange = omsb - semantics->precision;
823 /* If the resulting exponent is too high, overflow according to
824 the rounding mode. */
825 if(exponent + exponentChange > semantics->maxExponent)
826 return handleOverflow(rounding_mode);
828 /* Subnormal numbers have exponent minExponent, and their MSB
829 is forced based on that. */
830 if(exponent + exponentChange < semantics->minExponent)
831 exponentChange = semantics->minExponent - exponent;
833 /* Shifting left is easy as we don't lose precision. */
834 if(exponentChange < 0) {
835 assert(lost_fraction == lfExactlyZero);
837 shiftSignificandLeft(-exponentChange);
842 if(exponentChange > 0) {
845 /* Shift right and capture any new lost fraction. */
846 lf = shiftSignificandRight(exponentChange);
848 lost_fraction = combineLostFractions(lf, lost_fraction);
850 /* Keep OMSB up-to-date. */
851 if(omsb > (unsigned) exponentChange)
852 omsb -= (unsigned) exponentChange;
858 /* Now round the number according to rounding_mode given the lost
861 /* As specified in IEEE 754, since we do not trap we do not report
862 underflow for exact results. */
863 if(lost_fraction == lfExactlyZero) {
864 /* Canonicalize zeroes. */
871 /* Increment the significand if we're rounding away from zero. */
872 if(roundAwayFromZero(rounding_mode, lost_fraction, 0)) {
874 exponent = semantics->minExponent;
876 incrementSignificand();
877 omsb = significandMSB() + 1;
879 /* Did the significand increment overflow? */
880 if(omsb == (unsigned) semantics->precision + 1) {
881 /* Renormalize by incrementing the exponent and shifting our
882 significand right one. However if we already have the
883 maximum exponent we overflow to infinity. */
884 if(exponent == semantics->maxExponent) {
885 category = fcInfinity;
887 return (opStatus) (opOverflow | opInexact);
890 shiftSignificandRight(1);
896 /* The normal case - we were and are not denormal, and any
897 significand increment above didn't overflow. */
898 if(omsb == semantics->precision)
901 /* We have a non-zero denormal. */
902 assert(omsb < semantics->precision);
903 assert(exponent == semantics->minExponent);
905 /* Canonicalize zeroes. */
909 /* The fcZero case is a denormal that underflowed to zero. */
910 return (opStatus) (opUnderflow | opInexact);
914 APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract)
916 switch(convolve(category, rhs.category)) {
920 case convolve(fcNaN, fcZero):
921 case convolve(fcNaN, fcNormal):
922 case convolve(fcNaN, fcInfinity):
923 case convolve(fcNaN, fcNaN):
924 case convolve(fcNormal, fcZero):
925 case convolve(fcInfinity, fcNormal):
926 case convolve(fcInfinity, fcZero):
929 case convolve(fcZero, fcNaN):
930 case convolve(fcNormal, fcNaN):
931 case convolve(fcInfinity, fcNaN):
933 copySignificand(rhs);
936 case convolve(fcNormal, fcInfinity):
937 case convolve(fcZero, fcInfinity):
938 category = fcInfinity;
939 sign = rhs.sign ^ subtract;
942 case convolve(fcZero, fcNormal):
944 sign = rhs.sign ^ subtract;
947 case convolve(fcZero, fcZero):
948 /* Sign depends on rounding mode; handled by caller. */
951 case convolve(fcInfinity, fcInfinity):
952 /* Differently signed infinities can only be validly
954 if(sign ^ rhs.sign != subtract) {
956 // Arbitrary but deterministic value for significand
957 APInt::tcSet(significandParts(), ~0U, partCount());
963 case convolve(fcNormal, fcNormal):
968 /* Add or subtract two normal numbers. */
970 APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract)
973 lostFraction lost_fraction;
976 /* Determine if the operation on the absolute values is effectively
977 an addition or subtraction. */
978 subtract ^= (sign ^ rhs.sign);
980 /* Are we bigger exponent-wise than the RHS? */
981 bits = exponent - rhs.exponent;
983 /* Subtraction is more subtle than one might naively expect. */
985 APFloat temp_rhs(rhs);
989 reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan;
990 lost_fraction = lfExactlyZero;
991 } else if (bits > 0) {
992 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1);
993 shiftSignificandLeft(1);
996 lost_fraction = shiftSignificandRight(-bits - 1);
997 temp_rhs.shiftSignificandLeft(1);
1002 carry = temp_rhs.subtractSignificand
1003 (*this, lost_fraction != lfExactlyZero);
1004 copySignificand(temp_rhs);
1007 carry = subtractSignificand
1008 (temp_rhs, lost_fraction != lfExactlyZero);
1011 /* Invert the lost fraction - it was on the RHS and
1013 if(lost_fraction == lfLessThanHalf)
1014 lost_fraction = lfMoreThanHalf;
1015 else if(lost_fraction == lfMoreThanHalf)
1016 lost_fraction = lfLessThanHalf;
1018 /* The code above is intended to ensure that no borrow is
1023 APFloat temp_rhs(rhs);
1025 lost_fraction = temp_rhs.shiftSignificandRight(bits);
1026 carry = addSignificand(temp_rhs);
1028 lost_fraction = shiftSignificandRight(-bits);
1029 carry = addSignificand(rhs);
1032 /* We have a guard bit; generating a carry cannot happen. */
1036 return lost_fraction;
1040 APFloat::multiplySpecials(const APFloat &rhs)
1042 switch(convolve(category, rhs.category)) {
1046 case convolve(fcNaN, fcZero):
1047 case convolve(fcNaN, fcNormal):
1048 case convolve(fcNaN, fcInfinity):
1049 case convolve(fcNaN, fcNaN):
1052 case convolve(fcZero, fcNaN):
1053 case convolve(fcNormal, fcNaN):
1054 case convolve(fcInfinity, fcNaN):
1056 copySignificand(rhs);
1059 case convolve(fcNormal, fcInfinity):
1060 case convolve(fcInfinity, fcNormal):
1061 case convolve(fcInfinity, fcInfinity):
1062 category = fcInfinity;
1065 case convolve(fcZero, fcNormal):
1066 case convolve(fcNormal, fcZero):
1067 case convolve(fcZero, fcZero):
1071 case convolve(fcZero, fcInfinity):
1072 case convolve(fcInfinity, fcZero):
1074 // Arbitrary but deterministic value for significand
1075 APInt::tcSet(significandParts(), ~0U, partCount());
1078 case convolve(fcNormal, fcNormal):
1084 APFloat::divideSpecials(const APFloat &rhs)
1086 switch(convolve(category, rhs.category)) {
1090 case convolve(fcNaN, fcZero):
1091 case convolve(fcNaN, fcNormal):
1092 case convolve(fcNaN, fcInfinity):
1093 case convolve(fcNaN, fcNaN):
1094 case convolve(fcInfinity, fcZero):
1095 case convolve(fcInfinity, fcNormal):
1096 case convolve(fcZero, fcInfinity):
1097 case convolve(fcZero, fcNormal):
1100 case convolve(fcZero, fcNaN):
1101 case convolve(fcNormal, fcNaN):
1102 case convolve(fcInfinity, fcNaN):
1104 copySignificand(rhs);
1107 case convolve(fcNormal, fcInfinity):
1111 case convolve(fcNormal, fcZero):
1112 category = fcInfinity;
1115 case convolve(fcInfinity, fcInfinity):
1116 case convolve(fcZero, fcZero):
1118 // Arbitrary but deterministic value for significand
1119 APInt::tcSet(significandParts(), ~0U, partCount());
1122 case convolve(fcNormal, fcNormal):
1129 APFloat::changeSign()
1131 /* Look mummy, this one's easy. */
1136 APFloat::clearSign()
1138 /* So is this one. */
1143 APFloat::copySign(const APFloat &rhs)
1149 /* Normalized addition or subtraction. */
1151 APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode,
1156 fs = addOrSubtractSpecials(rhs, subtract);
1158 /* This return code means it was not a simple case. */
1159 if(fs == opDivByZero) {
1160 lostFraction lost_fraction;
1162 lost_fraction = addOrSubtractSignificand(rhs, subtract);
1163 fs = normalize(rounding_mode, lost_fraction);
1165 /* Can only be zero if we lost no fraction. */
1166 assert(category != fcZero || lost_fraction == lfExactlyZero);
1169 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1170 positive zero unless rounding to minus infinity, except that
1171 adding two like-signed zeroes gives that zero. */
1172 if(category == fcZero) {
1173 if(rhs.category != fcZero || (sign == rhs.sign) == subtract)
1174 sign = (rounding_mode == rmTowardNegative);
1180 /* Normalized addition. */
1182 APFloat::add(const APFloat &rhs, roundingMode rounding_mode)
1184 return addOrSubtract(rhs, rounding_mode, false);
1187 /* Normalized subtraction. */
1189 APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode)
1191 return addOrSubtract(rhs, rounding_mode, true);
1194 /* Normalized multiply. */
1196 APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode)
1201 fs = multiplySpecials(rhs);
1203 if(category == fcNormal) {
1204 lostFraction lost_fraction = multiplySignificand(rhs, 0);
1205 fs = normalize(rounding_mode, lost_fraction);
1206 if(lost_fraction != lfExactlyZero)
1207 fs = (opStatus) (fs | opInexact);
1213 /* Normalized divide. */
1215 APFloat::divide(const APFloat &rhs, roundingMode rounding_mode)
1220 fs = divideSpecials(rhs);
1222 if(category == fcNormal) {
1223 lostFraction lost_fraction = divideSignificand(rhs);
1224 fs = normalize(rounding_mode, lost_fraction);
1225 if(lost_fraction != lfExactlyZero)
1226 fs = (opStatus) (fs | opInexact);
1232 /* Normalized remainder. This is not currently doing TRT. */
1234 APFloat::mod(const APFloat &rhs, roundingMode rounding_mode)
1238 unsigned int origSign = sign;
1239 fs = V.divide(rhs, rmNearestTiesToEven);
1240 if (fs == opDivByZero)
1243 int parts = partCount();
1244 integerPart *x = new integerPart[parts];
1245 fs = V.convertToInteger(x, parts * integerPartWidth, true,
1246 rmNearestTiesToEven);
1247 if (fs==opInvalidOp)
1250 fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true,
1251 rmNearestTiesToEven);
1252 assert(fs==opOK); // should always work
1254 fs = V.multiply(rhs, rounding_mode);
1255 assert(fs==opOK || fs==opInexact); // should not overflow or underflow
1257 fs = subtract(V, rounding_mode);
1258 assert(fs==opOK || fs==opInexact); // likewise
1261 sign = origSign; // IEEE754 requires this
1266 /* Normalized fused-multiply-add. */
1268 APFloat::fusedMultiplyAdd(const APFloat &multiplicand,
1269 const APFloat &addend,
1270 roundingMode rounding_mode)
1274 /* Post-multiplication sign, before addition. */
1275 sign ^= multiplicand.sign;
1277 /* If and only if all arguments are normal do we need to do an
1278 extended-precision calculation. */
1279 if(category == fcNormal
1280 && multiplicand.category == fcNormal
1281 && addend.category == fcNormal) {
1282 lostFraction lost_fraction;
1284 lost_fraction = multiplySignificand(multiplicand, &addend);
1285 fs = normalize(rounding_mode, lost_fraction);
1286 if(lost_fraction != lfExactlyZero)
1287 fs = (opStatus) (fs | opInexact);
1289 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
1290 positive zero unless rounding to minus infinity, except that
1291 adding two like-signed zeroes gives that zero. */
1292 if(category == fcZero && sign != addend.sign)
1293 sign = (rounding_mode == rmTowardNegative);
1295 fs = multiplySpecials(multiplicand);
1297 /* FS can only be opOK or opInvalidOp. There is no more work
1298 to do in the latter case. The IEEE-754R standard says it is
1299 implementation-defined in this case whether, if ADDEND is a
1300 quiet NaN, we raise invalid op; this implementation does so.
1302 If we need to do the addition we can do so with normal
1305 fs = addOrSubtract(addend, rounding_mode, false);
1311 /* Comparison requires normalized numbers. */
1313 APFloat::compare(const APFloat &rhs) const
1317 assert(semantics == rhs.semantics);
1319 switch(convolve(category, rhs.category)) {
1323 case convolve(fcNaN, fcZero):
1324 case convolve(fcNaN, fcNormal):
1325 case convolve(fcNaN, fcInfinity):
1326 case convolve(fcNaN, fcNaN):
1327 case convolve(fcZero, fcNaN):
1328 case convolve(fcNormal, fcNaN):
1329 case convolve(fcInfinity, fcNaN):
1330 return cmpUnordered;
1332 case convolve(fcInfinity, fcNormal):
1333 case convolve(fcInfinity, fcZero):
1334 case convolve(fcNormal, fcZero):
1338 return cmpGreaterThan;
1340 case convolve(fcNormal, fcInfinity):
1341 case convolve(fcZero, fcInfinity):
1342 case convolve(fcZero, fcNormal):
1344 return cmpGreaterThan;
1348 case convolve(fcInfinity, fcInfinity):
1349 if(sign == rhs.sign)
1354 return cmpGreaterThan;
1356 case convolve(fcZero, fcZero):
1359 case convolve(fcNormal, fcNormal):
1363 /* Two normal numbers. Do they have the same sign? */
1364 if(sign != rhs.sign) {
1366 result = cmpLessThan;
1368 result = cmpGreaterThan;
1370 /* Compare absolute values; invert result if negative. */
1371 result = compareAbsoluteValue(rhs);
1374 if(result == cmpLessThan)
1375 result = cmpGreaterThan;
1376 else if(result == cmpGreaterThan)
1377 result = cmpLessThan;
1385 APFloat::convert(const fltSemantics &toSemantics,
1386 roundingMode rounding_mode)
1388 lostFraction lostFraction;
1389 unsigned int newPartCount, oldPartCount;
1392 lostFraction = lfExactlyZero;
1393 newPartCount = partCountForBits(toSemantics.precision + 1);
1394 oldPartCount = partCount();
1396 /* Handle storage complications. If our new form is wider,
1397 re-allocate our bit pattern into wider storage. If it is
1398 narrower, we ignore the excess parts, but if narrowing to a
1399 single part we need to free the old storage.
1400 Be careful not to reference significandParts for zeroes
1401 and infinities, since it aborts. */
1402 if (newPartCount > oldPartCount) {
1403 integerPart *newParts;
1404 newParts = new integerPart[newPartCount];
1405 APInt::tcSet(newParts, 0, newPartCount);
1406 if (category==fcNormal || category==fcNaN)
1407 APInt::tcAssign(newParts, significandParts(), oldPartCount);
1409 significand.parts = newParts;
1410 } else if (newPartCount < oldPartCount) {
1411 /* Capture any lost fraction through truncation of parts so we get
1412 correct rounding whilst normalizing. */
1413 if (category==fcNormal)
1414 lostFraction = lostFractionThroughTruncation
1415 (significandParts(), oldPartCount, toSemantics.precision);
1416 if (newPartCount == 1) {
1417 integerPart newPart = 0;
1418 if (category==fcNormal || category==fcNaN)
1419 newPart = significandParts()[0];
1421 significand.part = newPart;
1425 if(category == fcNormal) {
1426 /* Re-interpret our bit-pattern. */
1427 exponent += toSemantics.precision - semantics->precision;
1428 semantics = &toSemantics;
1429 fs = normalize(rounding_mode, lostFraction);
1430 } else if (category == fcNaN) {
1431 int shift = toSemantics.precision - semantics->precision;
1432 // No normalization here, just truncate
1434 APInt::tcShiftLeft(significandParts(), newPartCount, shift);
1436 APInt::tcShiftRight(significandParts(), newPartCount, -shift);
1437 // gcc forces the Quiet bit on, which means (float)(double)(float_sNan)
1438 // does not give you back the same bits. This is dubious, and we
1439 // don't currently do it. You're really supposed to get
1440 // an invalid operation signal at runtime, but nobody does that.
1441 semantics = &toSemantics;
1444 semantics = &toSemantics;
1451 /* Convert a floating point number to an integer according to the
1452 rounding mode. If the rounded integer value is out of range this
1453 returns an invalid operation exception. If the rounded value is in
1454 range but the floating point number is not the exact integer, the C
1455 standard doesn't require an inexact exception to be raised. IEEE
1456 854 does require it so we do that.
1458 Note that for conversions to integer type the C standard requires
1459 round-to-zero to always be used. */
1461 APFloat::convertToInteger(integerPart *parts, unsigned int width,
1463 roundingMode rounding_mode) const
1465 lostFraction lost_fraction;
1466 unsigned int msb, partsCount;
1469 partsCount = partCountForBits(width);
1471 /* Handle the three special cases first. We produce
1472 a deterministic result even for the Invalid cases. */
1473 if (category == fcNaN) {
1474 // Neither sign nor isSigned affects this.
1475 APInt::tcSet(parts, 0, partsCount);
1478 if (category == fcInfinity) {
1479 if (!sign && isSigned)
1480 APInt::tcSetLeastSignificantBits(parts, partsCount, width-1);
1481 else if (!sign && !isSigned)
1482 APInt::tcSetLeastSignificantBits(parts, partsCount, width);
1483 else if (sign && isSigned) {
1484 APInt::tcSetLeastSignificantBits(parts, partsCount, 1);
1485 APInt::tcShiftLeft(parts, partsCount, width-1);
1486 } else // sign && !isSigned
1487 APInt::tcSet(parts, 0, partsCount);
1490 if (category == fcZero) {
1491 APInt::tcSet(parts, 0, partsCount);
1495 /* Shift the bit pattern so the fraction is lost. */
1498 bits = (int) semantics->precision - 1 - exponent;
1501 lost_fraction = tmp.shiftSignificandRight(bits);
1503 if (-bits >= semantics->precision) {
1504 // Unrepresentably large.
1505 if (!sign && isSigned)
1506 APInt::tcSetLeastSignificantBits(parts, partsCount, width-1);
1507 else if (!sign && !isSigned)
1508 APInt::tcSetLeastSignificantBits(parts, partsCount, width);
1509 else if (sign && isSigned) {
1510 APInt::tcSetLeastSignificantBits(parts, partsCount, 1);
1511 APInt::tcShiftLeft(parts, partsCount, width-1);
1512 } else // sign && !isSigned
1513 APInt::tcSet(parts, 0, partsCount);
1514 return (opStatus)(opOverflow | opInexact);
1516 tmp.shiftSignificandLeft(-bits);
1517 lost_fraction = lfExactlyZero;
1520 if(lost_fraction != lfExactlyZero
1521 && tmp.roundAwayFromZero(rounding_mode, lost_fraction, 0))
1522 tmp.incrementSignificand();
1524 msb = tmp.significandMSB();
1526 /* Negative numbers cannot be represented as unsigned. */
1527 if(!isSigned && tmp.sign && msb != -1U)
1530 /* It takes exponent + 1 bits to represent the truncated floating
1531 point number without its sign. We lose a bit for the sign, but
1532 the maximally negative integer is a special case. */
1533 if(msb + 1 > width) /* !! Not same as msb >= width !! */
1536 if(isSigned && msb + 1 == width
1537 && (!tmp.sign || tmp.significandLSB() != msb))
1540 APInt::tcAssign(parts, tmp.significandParts(), partsCount);
1543 APInt::tcNegate(parts, partsCount);
1545 if(lost_fraction == lfExactlyZero)
1551 /* Convert an unsigned integer SRC to a floating point number,
1552 rounding according to ROUNDING_MODE. The sign of the floating
1553 point number is not modified. */
1555 APFloat::convertFromUnsignedParts(const integerPart *src,
1556 unsigned int srcCount,
1557 roundingMode rounding_mode)
1559 unsigned int dstCount;
1560 lostFraction lost_fraction;
1563 category = fcNormal;
1564 exponent = semantics->precision - 1;
1566 dst = significandParts();
1567 dstCount = partCount();
1569 /* We need to capture the non-zero most significant parts. */
1570 while (srcCount > dstCount && src[srcCount - 1] == 0)
1573 /* Copy the bit image of as many parts as we can. If we are wider,
1574 zero-out remaining parts. */
1575 if (dstCount >= srcCount) {
1576 APInt::tcAssign(dst, src, srcCount);
1577 while (srcCount < dstCount)
1578 dst[srcCount++] = 0;
1579 lost_fraction = lfExactlyZero;
1581 exponent += (srcCount - dstCount) * integerPartWidth;
1582 APInt::tcAssign(dst, src + (srcCount - dstCount), dstCount);
1583 lost_fraction = lostFractionThroughTruncation(src, srcCount,
1584 dstCount * integerPartWidth);
1587 return normalize(rounding_mode, lost_fraction);
1590 /* FIXME: should this just take a const APInt reference? */
1592 APFloat::convertFromZeroExtendedInteger(const integerPart *parts,
1593 unsigned int width, bool isSigned,
1594 roundingMode rounding_mode)
1596 unsigned int partCount = partCountForBits(width);
1598 APInt api = APInt(width, partCount, parts);
1599 integerPart *copy = new integerPart[partCount];
1602 if(isSigned && APInt::tcExtractBit(parts, width - 1)) {
1607 APInt::tcAssign(copy, api.getRawData(), partCount);
1608 status = convertFromUnsignedParts(copy, partCount, rounding_mode);
1613 APFloat::convertFromHexadecimalString(const char *p,
1614 roundingMode rounding_mode)
1616 lostFraction lost_fraction;
1617 integerPart *significand;
1618 unsigned int bitPos, partsCount;
1619 const char *dot, *firstSignificantDigit;
1623 category = fcNormal;
1625 significand = significandParts();
1626 partsCount = partCount();
1627 bitPos = partsCount * integerPartWidth;
1629 /* Skip leading zeroes and any (hexa)decimal point. */
1630 p = skipLeadingZeroesAndAnyDot(p, &dot);
1631 firstSignificantDigit = p;
1634 integerPart hex_value;
1641 hex_value = hexDigitValue(*p);
1642 if(hex_value == -1U) {
1643 lost_fraction = lfExactlyZero;
1649 /* Store the number whilst 4-bit nibbles remain. */
1652 hex_value <<= bitPos % integerPartWidth;
1653 significand[bitPos / integerPartWidth] |= hex_value;
1655 lost_fraction = trailingHexadecimalFraction(p, hex_value);
1656 while(hexDigitValue(*p) != -1U)
1662 /* Hex floats require an exponent but not a hexadecimal point. */
1663 assert(*p == 'p' || *p == 'P');
1665 /* Ignore the exponent if we are zero. */
1666 if(p != firstSignificantDigit) {
1669 /* Implicit hexadecimal point? */
1673 /* Calculate the exponent adjustment implicit in the number of
1674 significant digits. */
1675 expAdjustment = dot - firstSignificantDigit;
1676 if(expAdjustment < 0)
1678 expAdjustment = expAdjustment * 4 - 1;
1680 /* Adjust for writing the significand starting at the most
1681 significant nibble. */
1682 expAdjustment += semantics->precision;
1683 expAdjustment -= partsCount * integerPartWidth;
1685 /* Adjust for the given exponent. */
1686 exponent = totalExponent(p, expAdjustment);
1689 return normalize(rounding_mode, lost_fraction);
1693 APFloat::convertFromString(const char *p, roundingMode rounding_mode)
1695 /* Handle a leading minus sign. */
1701 if(p[0] == '0' && (p[1] == 'x' || p[1] == 'X'))
1702 return convertFromHexadecimalString(p + 2, rounding_mode);
1704 assert(0 && "Decimal to binary conversions not yet implemented");
1708 /* Write out a hexadecimal representation of the floating point value
1709 to DST, which must be of sufficient size, in the C99 form
1710 [-]0xh.hhhhp[+-]d. Return the number of characters written,
1711 excluding the terminating NUL.
1713 If UPPERCASE, the output is in upper case, otherwise in lower case.
1715 HEXDIGITS digits appear altogether, rounding the value if
1716 necessary. If HEXDIGITS is 0, the minimal precision to display the
1717 number precisely is used instead. If nothing would appear after
1718 the decimal point it is suppressed.
1720 The decimal exponent is always printed and has at least one digit.
1721 Zero values display an exponent of zero. Infinities and NaNs
1722 appear as "infinity" or "nan" respectively.
1724 The above rules are as specified by C99. There is ambiguity about
1725 what the leading hexadecimal digit should be. This implementation
1726 uses whatever is necessary so that the exponent is displayed as
1727 stored. This implies the exponent will fall within the IEEE format
1728 range, and the leading hexadecimal digit will be 0 (for denormals),
1729 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with
1730 any other digits zero).
1733 APFloat::convertToHexString(char *dst, unsigned int hexDigits,
1734 bool upperCase, roundingMode rounding_mode) const
1744 memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1);
1745 dst += sizeof infinityL - 1;
1749 memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1);
1750 dst += sizeof NaNU - 1;
1755 *dst++ = upperCase ? 'X': 'x';
1757 if (hexDigits > 1) {
1759 memset (dst, '0', hexDigits - 1);
1760 dst += hexDigits - 1;
1762 *dst++ = upperCase ? 'P': 'p';
1767 dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode);
1776 /* Does the hard work of outputting the correctly rounded hexadecimal
1777 form of a normal floating point number with the specified number of
1778 hexadecimal digits. If HEXDIGITS is zero the minimum number of
1779 digits necessary to print the value precisely is output. */
1781 APFloat::convertNormalToHexString(char *dst, unsigned int hexDigits,
1783 roundingMode rounding_mode) const
1785 unsigned int count, valueBits, shift, partsCount, outputDigits;
1786 const char *hexDigitChars;
1787 const integerPart *significand;
1792 *dst++ = upperCase ? 'X': 'x';
1795 hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower;
1797 significand = significandParts();
1798 partsCount = partCount();
1800 /* +3 because the first digit only uses the single integer bit, so
1801 we have 3 virtual zero most-significant-bits. */
1802 valueBits = semantics->precision + 3;
1803 shift = integerPartWidth - valueBits % integerPartWidth;
1805 /* The natural number of digits required ignoring trailing
1806 insignificant zeroes. */
1807 outputDigits = (valueBits - significandLSB () + 3) / 4;
1809 /* hexDigits of zero means use the required number for the
1810 precision. Otherwise, see if we are truncating. If we are,
1811 find out if we need to round away from zero. */
1813 if (hexDigits < outputDigits) {
1814 /* We are dropping non-zero bits, so need to check how to round.
1815 "bits" is the number of dropped bits. */
1817 lostFraction fraction;
1819 bits = valueBits - hexDigits * 4;
1820 fraction = lostFractionThroughTruncation (significand, partsCount, bits);
1821 roundUp = roundAwayFromZero(rounding_mode, fraction, bits);
1823 outputDigits = hexDigits;
1826 /* Write the digits consecutively, and start writing in the location
1827 of the hexadecimal point. We move the most significant digit
1828 left and add the hexadecimal point later. */
1831 count = (valueBits + integerPartWidth - 1) / integerPartWidth;
1833 while (outputDigits && count) {
1836 /* Put the most significant integerPartWidth bits in "part". */
1837 if (--count == partsCount)
1838 part = 0; /* An imaginary higher zero part. */
1840 part = significand[count] << shift;
1843 part |= significand[count - 1] >> (integerPartWidth - shift);
1845 /* Convert as much of "part" to hexdigits as we can. */
1846 unsigned int curDigits = integerPartWidth / 4;
1848 if (curDigits > outputDigits)
1849 curDigits = outputDigits;
1850 dst += partAsHex (dst, part, curDigits, hexDigitChars);
1851 outputDigits -= curDigits;
1857 /* Note that hexDigitChars has a trailing '0'. */
1860 *q = hexDigitChars[hexDigitValue (*q) + 1];
1861 } while (*q == '0');
1864 /* Add trailing zeroes. */
1865 memset (dst, '0', outputDigits);
1866 dst += outputDigits;
1869 /* Move the most significant digit to before the point, and if there
1870 is something after the decimal point add it. This must come
1871 after rounding above. */
1878 /* Finally output the exponent. */
1879 *dst++ = upperCase ? 'P': 'p';
1881 return writeSignedDecimal (dst, exponent);
1884 // For good performance it is desirable for different APFloats
1885 // to produce different integers.
1887 APFloat::getHashValue() const
1889 if (category==fcZero) return sign<<8 | semantics->precision ;
1890 else if (category==fcInfinity) return sign<<9 | semantics->precision;
1891 else if (category==fcNaN) return 1<<10 | semantics->precision;
1893 uint32_t hash = sign<<11 | semantics->precision | exponent<<12;
1894 const integerPart* p = significandParts();
1895 for (int i=partCount(); i>0; i--, p++)
1896 hash ^= ((uint32_t)*p) ^ (*p)>>32;
1901 // Conversion from APFloat to/from host float/double. It may eventually be
1902 // possible to eliminate these and have everybody deal with APFloats, but that
1903 // will take a while. This approach will not easily extend to long double.
1904 // Current implementation requires integerPartWidth==64, which is correct at
1905 // the moment but could be made more general.
1907 // Denormals have exponent minExponent in APFloat, but minExponent-1 in
1908 // the actual IEEE respresentations. We compensate for that here.
1911 APFloat::convertF80LongDoubleAPFloatToAPInt() const
1913 assert(semantics == (const llvm::fltSemantics* const)&x87DoubleExtended);
1914 assert (partCount()==2);
1916 uint64_t myexponent, mysignificand;
1918 if (category==fcNormal) {
1919 myexponent = exponent+16383; //bias
1920 mysignificand = significandParts()[0];
1921 if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL))
1922 myexponent = 0; // denormal
1923 } else if (category==fcZero) {
1926 } else if (category==fcInfinity) {
1927 myexponent = 0x7fff;
1928 mysignificand = 0x8000000000000000ULL;
1930 assert(category == fcNaN && "Unknown category");
1931 myexponent = 0x7fff;
1932 mysignificand = significandParts()[0];
1936 words[0] = (((uint64_t)sign & 1) << 63) |
1937 ((myexponent & 0x7fff) << 48) |
1938 ((mysignificand >>16) & 0xffffffffffffLL);
1939 words[1] = mysignificand & 0xffff;
1940 return APInt(80, 2, words);
1944 APFloat::convertDoubleAPFloatToAPInt() const
1946 assert(semantics == (const llvm::fltSemantics*)&IEEEdouble);
1947 assert (partCount()==1);
1949 uint64_t myexponent, mysignificand;
1951 if (category==fcNormal) {
1952 myexponent = exponent+1023; //bias
1953 mysignificand = *significandParts();
1954 if (myexponent==1 && !(mysignificand & 0x10000000000000LL))
1955 myexponent = 0; // denormal
1956 } else if (category==fcZero) {
1959 } else if (category==fcInfinity) {
1963 assert(category == fcNaN && "Unknown category!");
1965 mysignificand = *significandParts();
1968 return APInt(64, (((((uint64_t)sign & 1) << 63) |
1969 ((myexponent & 0x7ff) << 52) |
1970 (mysignificand & 0xfffffffffffffLL))));
1974 APFloat::convertFloatAPFloatToAPInt() const
1976 assert(semantics == (const llvm::fltSemantics*)&IEEEsingle);
1977 assert (partCount()==1);
1979 uint32_t myexponent, mysignificand;
1981 if (category==fcNormal) {
1982 myexponent = exponent+127; //bias
1983 mysignificand = *significandParts();
1984 if (myexponent == 1 && !(mysignificand & 0x400000))
1985 myexponent = 0; // denormal
1986 } else if (category==fcZero) {
1989 } else if (category==fcInfinity) {
1993 assert(category == fcNaN && "Unknown category!");
1995 mysignificand = *significandParts();
1998 return APInt(32, (((sign&1) << 31) | ((myexponent&0xff) << 23) |
1999 (mysignificand & 0x7fffff)));
2003 APFloat::convertToAPInt() const
2005 if (semantics == (const llvm::fltSemantics* const)&IEEEsingle)
2006 return convertFloatAPFloatToAPInt();
2008 if (semantics == (const llvm::fltSemantics* const)&IEEEdouble)
2009 return convertDoubleAPFloatToAPInt();
2011 assert(semantics == (const llvm::fltSemantics* const)&x87DoubleExtended &&
2013 return convertF80LongDoubleAPFloatToAPInt();
2017 APFloat::convertToFloat() const
2019 assert(semantics == (const llvm::fltSemantics* const)&IEEEsingle);
2020 APInt api = convertToAPInt();
2021 return api.bitsToFloat();
2025 APFloat::convertToDouble() const
2027 assert(semantics == (const llvm::fltSemantics* const)&IEEEdouble);
2028 APInt api = convertToAPInt();
2029 return api.bitsToDouble();
2032 /// Integer bit is explicit in this format. Current Intel book does not
2033 /// define meaning of:
2034 /// exponent = all 1's, integer bit not set.
2035 /// exponent = 0, integer bit set. (formerly "psuedodenormals")
2036 /// exponent!=0 nor all 1's, integer bit not set. (formerly "unnormals")
2038 APFloat::initFromF80LongDoubleAPInt(const APInt &api)
2040 assert(api.getBitWidth()==80);
2041 uint64_t i1 = api.getRawData()[0];
2042 uint64_t i2 = api.getRawData()[1];
2043 uint64_t myexponent = (i1 >> 48) & 0x7fff;
2044 uint64_t mysignificand = ((i1 << 16) & 0xffffffffffff0000ULL) |
2047 initialize(&APFloat::x87DoubleExtended);
2048 assert(partCount()==2);
2051 if (myexponent==0 && mysignificand==0) {
2052 // exponent, significand meaningless
2054 } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) {
2055 // exponent, significand meaningless
2056 category = fcInfinity;
2057 } else if (myexponent==0x7fff && mysignificand!=0x8000000000000000ULL) {
2058 // exponent meaningless
2060 significandParts()[0] = mysignificand;
2061 significandParts()[1] = 0;
2063 category = fcNormal;
2064 exponent = myexponent - 16383;
2065 significandParts()[0] = mysignificand;
2066 significandParts()[1] = 0;
2067 if (myexponent==0) // denormal
2073 APFloat::initFromDoubleAPInt(const APInt &api)
2075 assert(api.getBitWidth()==64);
2076 uint64_t i = *api.getRawData();
2077 uint64_t myexponent = (i >> 52) & 0x7ff;
2078 uint64_t mysignificand = i & 0xfffffffffffffLL;
2080 initialize(&APFloat::IEEEdouble);
2081 assert(partCount()==1);
2084 if (myexponent==0 && mysignificand==0) {
2085 // exponent, significand meaningless
2087 } else if (myexponent==0x7ff && mysignificand==0) {
2088 // exponent, significand meaningless
2089 category = fcInfinity;
2090 } else if (myexponent==0x7ff && mysignificand!=0) {
2091 // exponent meaningless
2093 *significandParts() = mysignificand;
2095 category = fcNormal;
2096 exponent = myexponent - 1023;
2097 *significandParts() = mysignificand;
2098 if (myexponent==0) // denormal
2101 *significandParts() |= 0x10000000000000LL; // integer bit
2106 APFloat::initFromFloatAPInt(const APInt & api)
2108 assert(api.getBitWidth()==32);
2109 uint32_t i = (uint32_t)*api.getRawData();
2110 uint32_t myexponent = (i >> 23) & 0xff;
2111 uint32_t mysignificand = i & 0x7fffff;
2113 initialize(&APFloat::IEEEsingle);
2114 assert(partCount()==1);
2117 if (myexponent==0 && mysignificand==0) {
2118 // exponent, significand meaningless
2120 } else if (myexponent==0xff && mysignificand==0) {
2121 // exponent, significand meaningless
2122 category = fcInfinity;
2123 } else if (myexponent==0xff && mysignificand!=0) {
2124 // sign, exponent, significand meaningless
2126 *significandParts() = mysignificand;
2128 category = fcNormal;
2129 exponent = myexponent - 127; //bias
2130 *significandParts() = mysignificand;
2131 if (myexponent==0) // denormal
2134 *significandParts() |= 0x800000; // integer bit
2138 /// Treat api as containing the bits of a floating point number. Currently
2139 /// we infer the floating point type from the size of the APInt. FIXME: This
2140 /// breaks when we get to PPC128 and IEEE128 (but both cannot exist in the
2141 /// same compile...)
2143 APFloat::initFromAPInt(const APInt& api)
2145 if (api.getBitWidth() == 32)
2146 return initFromFloatAPInt(api);
2147 else if (api.getBitWidth()==64)
2148 return initFromDoubleAPInt(api);
2149 else if (api.getBitWidth()==80)
2150 return initFromF80LongDoubleAPInt(api);
2155 APFloat::APFloat(const APInt& api)
2160 APFloat::APFloat(float f)
2162 APInt api = APInt(32, 0);
2163 initFromAPInt(api.floatToBits(f));
2166 APFloat::APFloat(double d)
2168 APInt api = APInt(64, 0);
2169 initFromAPInt(api.doubleToBits(d));