1 //===- llvm/ADT/APFloat.h - Arbitrary Precision Floating Point ---*- C++ -*-==//
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 //===----------------------------------------------------------------------===//
12 /// This file declares a class to represent arbitrary precision floating point
13 /// values and provide a variety of arithmetic operations on them.
15 //===----------------------------------------------------------------------===//
17 #ifndef LLVM_ADT_APFLOAT_H
18 #define LLVM_ADT_APFLOAT_H
20 #include "llvm/ADT/APInt.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/Support/ErrorHandling.h"
25 #define APFLOAT_DISPATCH_ON_SEMANTICS(METHOD_CALL) \
27 if (usesLayout<IEEEFloat>(getSemantics())) \
28 return U.IEEE.METHOD_CALL; \
29 if (usesLayout<DoubleAPFloat>(getSemantics())) \
30 return U.Double.METHOD_CALL; \
31 llvm_unreachable("Unexpected semantics"); \
42 template <typename T> class SmallVectorImpl;
44 /// Enum that represents what fraction of the LSB truncated bits of an fp number
47 /// This essentially combines the roles of guard and sticky bits.
48 enum lostFraction { // Example of truncated bits:
49 lfExactlyZero, // 000000
50 lfLessThanHalf, // 0xxxxx x's not all zero
51 lfExactlyHalf, // 100000
52 lfMoreThanHalf // 1xxxxx x's not all zero
55 /// A self-contained host- and target-independent arbitrary-precision
56 /// floating-point software implementation.
58 /// APFloat uses bignum integer arithmetic as provided by static functions in
59 /// the APInt class. The library will work with bignum integers whose parts are
60 /// any unsigned type at least 16 bits wide, but 64 bits is recommended.
62 /// Written for clarity rather than speed, in particular with a view to use in
63 /// the front-end of a cross compiler so that target arithmetic can be correctly
64 /// performed on the host. Performance should nonetheless be reasonable,
65 /// particularly for its intended use. It may be useful as a base
66 /// implementation for a run-time library during development of a faster
67 /// target-specific one.
69 /// All 5 rounding modes in the IEEE-754R draft are handled correctly for all
70 /// implemented operations. Currently implemented operations are add, subtract,
71 /// multiply, divide, fused-multiply-add, conversion-to-float,
72 /// conversion-to-integer and conversion-from-integer. New rounding modes
73 /// (e.g. away from zero) can be added with three or four lines of code.
75 /// Four formats are built-in: IEEE single precision, double precision,
76 /// quadruple precision, and x87 80-bit extended double (when operating with
77 /// full extended precision). Adding a new format that obeys IEEE semantics
78 /// only requires adding two lines of code: a declaration and definition of the
81 /// All operations return the status of that operation as an exception bit-mask,
82 /// so multiple operations can be done consecutively with their results or-ed
83 /// together. The returned status can be useful for compiler diagnostics; e.g.,
84 /// inexact, underflow and overflow can be easily diagnosed on constant folding,
85 /// and compiler optimizers can determine what exceptions would be raised by
86 /// folding operations and optimize, or perhaps not optimize, accordingly.
88 /// At present, underflow tininess is detected after rounding; it should be
89 /// straight forward to add support for the before-rounding case too.
91 /// The library reads hexadecimal floating point numbers as per C99, and
92 /// correctly rounds if necessary according to the specified rounding mode.
93 /// Syntax is required to have been validated by the caller. It also converts
94 /// floating point numbers to hexadecimal text as per the C99 %a and %A
95 /// conversions. The output precision (or alternatively the natural minimal
96 /// precision) can be specified; if the requested precision is less than the
97 /// natural precision the output is correctly rounded for the specified rounding
100 /// It also reads decimal floating point numbers and correctly rounds according
101 /// to the specified rounding mode.
103 /// Conversion to decimal text is not currently implemented.
105 /// Non-zero finite numbers are represented internally as a sign bit, a 16-bit
106 /// signed exponent, and the significand as an array of integer parts. After
107 /// normalization of a number of precision P the exponent is within the range of
108 /// the format, and if the number is not denormal the P-th bit of the
109 /// significand is set as an explicit integer bit. For denormals the most
110 /// significant bit is shifted right so that the exponent is maintained at the
111 /// format's minimum, so that the smallest denormal has just the least
112 /// significant bit of the significand set. The sign of zeroes and infinities
113 /// is significant; the exponent and significand of such numbers is not stored,
114 /// but has a known implicit (deterministic) value: 0 for the significands, 0
115 /// for zero exponent, all 1 bits for infinity exponent. For NaNs the sign and
116 /// significand are deterministic, although not really meaningful, and preserved
117 /// in non-conversion operations. The exponent is implicitly all 1 bits.
119 /// APFloat does not provide any exception handling beyond default exception
120 /// handling. We represent Signaling NaNs via IEEE-754R 2008 6.2.1 should clause
121 /// by encoding Signaling NaNs with the first bit of its trailing significand as
127 /// Some features that may or may not be worth adding:
129 /// Binary to decimal conversion (hard).
131 /// Optional ability to detect underflow tininess before rounding.
133 /// New formats: x87 in single and double precision mode (IEEE apart from
134 /// extended exponent range) (hard).
136 /// New operations: sqrt, IEEE remainder, C90 fmod, nexttoward.
139 // This is the common type definitions shared by APFloat and its internal
140 // implementation classes. This struct should not define any non-static data
143 typedef APInt::WordType integerPart;
144 static const unsigned integerPartWidth = APInt::APINT_BITS_PER_WORD;
146 /// A signed type to represent a floating point numbers unbiased exponent.
147 typedef signed short ExponentType;
149 /// \name Floating Point Semantics.
152 static const fltSemantics &IEEEhalf() LLVM_READNONE;
153 static const fltSemantics &IEEEsingle() LLVM_READNONE;
154 static const fltSemantics &IEEEdouble() LLVM_READNONE;
155 static const fltSemantics &IEEEquad() LLVM_READNONE;
156 static const fltSemantics &PPCDoubleDouble() LLVM_READNONE;
157 static const fltSemantics &x87DoubleExtended() LLVM_READNONE;
159 /// A Pseudo fltsemantic used to construct APFloats that cannot conflict with
161 static const fltSemantics &Bogus() LLVM_READNONE;
165 /// IEEE-754R 5.11: Floating Point Comparison Relations.
173 /// IEEE-754R 4.3: Rounding-direction attributes.
182 /// IEEE-754R 7: Default exception handling.
184 /// opUnderflow or opOverflow are always returned or-ed with opInexact.
194 /// Category of internally-represented number.
202 /// Convenience enum used to construct an uninitialized APFloat.
203 enum uninitializedTag {
207 /// Enumeration of \c ilogb error results.
208 enum IlogbErrorKinds {
209 IEK_Zero = INT_MIN + 1,
214 static unsigned int semanticsPrecision(const fltSemantics &);
215 static ExponentType semanticsMinExponent(const fltSemantics &);
216 static ExponentType semanticsMaxExponent(const fltSemantics &);
217 static unsigned int semanticsSizeInBits(const fltSemantics &);
219 /// Returns the size of the floating point number (in bits) in the given
221 static unsigned getSizeInBits(const fltSemantics &Sem);
226 class IEEEFloat final : public APFloatBase {
228 /// \name Constructors
231 IEEEFloat(const fltSemantics &); // Default construct to 0.0
232 IEEEFloat(const fltSemantics &, integerPart);
233 IEEEFloat(const fltSemantics &, uninitializedTag);
234 IEEEFloat(const fltSemantics &, const APInt &);
235 explicit IEEEFloat(double d);
236 explicit IEEEFloat(float f);
237 IEEEFloat(const IEEEFloat &);
238 IEEEFloat(IEEEFloat &&);
243 /// Returns whether this instance allocated memory.
244 bool needsCleanup() const { return partCount() > 1; }
246 /// \name Convenience "constructors"
254 opStatus add(const IEEEFloat &, roundingMode);
255 opStatus subtract(const IEEEFloat &, roundingMode);
256 opStatus multiply(const IEEEFloat &, roundingMode);
257 opStatus divide(const IEEEFloat &, roundingMode);
259 opStatus remainder(const IEEEFloat &);
260 /// C fmod, or llvm frem.
261 opStatus mod(const IEEEFloat &);
262 opStatus fusedMultiplyAdd(const IEEEFloat &, const IEEEFloat &, roundingMode);
263 opStatus roundToIntegral(roundingMode);
264 /// IEEE-754R 5.3.1: nextUp/nextDown.
265 opStatus next(bool nextDown);
269 /// \name Sign operations.
276 /// \name Conversions
279 opStatus convert(const fltSemantics &, roundingMode, bool *);
280 opStatus convertToInteger(MutableArrayRef<integerPart>, unsigned int, bool,
281 roundingMode, bool *) const;
282 opStatus convertFromAPInt(const APInt &, bool, roundingMode);
283 opStatus convertFromSignExtendedInteger(const integerPart *, unsigned int,
285 opStatus convertFromZeroExtendedInteger(const integerPart *, unsigned int,
287 opStatus convertFromString(StringRef, roundingMode);
288 APInt bitcastToAPInt() const;
289 double convertToDouble() const;
290 float convertToFloat() const;
294 /// The definition of equality is not straightforward for floating point, so
295 /// we won't use operator==. Use one of the following, or write whatever it
296 /// is you really mean.
297 bool operator==(const IEEEFloat &) const = delete;
299 /// IEEE comparison with another floating point number (NaNs compare
300 /// unordered, 0==-0).
301 cmpResult compare(const IEEEFloat &) const;
303 /// Bitwise comparison for equality (QNaNs compare equal, 0!=-0).
304 bool bitwiseIsEqual(const IEEEFloat &) const;
306 /// Write out a hexadecimal representation of the floating point value to DST,
307 /// which must be of sufficient size, in the C99 form [-]0xh.hhhhp[+-]d.
308 /// Return the number of characters written, excluding the terminating NUL.
309 unsigned int convertToHexString(char *dst, unsigned int hexDigits,
310 bool upperCase, roundingMode) const;
312 /// \name IEEE-754R 5.7.2 General operations.
315 /// IEEE-754R isSignMinus: Returns true if and only if the current value is
318 /// This applies to zeros and NaNs as well.
319 bool isNegative() const { return sign; }
321 /// IEEE-754R isNormal: Returns true if and only if the current value is normal.
323 /// This implies that the current value of the float is not zero, subnormal,
324 /// infinite, or NaN following the definition of normality from IEEE-754R.
325 bool isNormal() const { return !isDenormal() && isFiniteNonZero(); }
327 /// Returns true if and only if the current value is zero, subnormal, or
330 /// This means that the value is not infinite or NaN.
331 bool isFinite() const { return !isNaN() && !isInfinity(); }
333 /// Returns true if and only if the float is plus or minus zero.
334 bool isZero() const { return category == fcZero; }
336 /// IEEE-754R isSubnormal(): Returns true if and only if the float is a
338 bool isDenormal() const;
340 /// IEEE-754R isInfinite(): Returns true if and only if the float is infinity.
341 bool isInfinity() const { return category == fcInfinity; }
343 /// Returns true if and only if the float is a quiet or signaling NaN.
344 bool isNaN() const { return category == fcNaN; }
346 /// Returns true if and only if the float is a signaling NaN.
347 bool isSignaling() const;
351 /// \name Simple Queries
354 fltCategory getCategory() const { return category; }
355 const fltSemantics &getSemantics() const { return *semantics; }
356 bool isNonZero() const { return category != fcZero; }
357 bool isFiniteNonZero() const { return isFinite() && !isZero(); }
358 bool isPosZero() const { return isZero() && !isNegative(); }
359 bool isNegZero() const { return isZero() && isNegative(); }
361 /// Returns true if and only if the number has the smallest possible non-zero
362 /// magnitude in the current semantics.
363 bool isSmallest() const;
365 /// Returns true if and only if the number has the largest possible finite
366 /// magnitude in the current semantics.
367 bool isLargest() const;
369 /// Returns true if and only if the number is an exact integer.
370 bool isInteger() const;
374 IEEEFloat &operator=(const IEEEFloat &);
375 IEEEFloat &operator=(IEEEFloat &&);
377 /// Overload to compute a hash code for an APFloat value.
379 /// Note that the use of hash codes for floating point values is in general
380 /// frought with peril. Equality is hard to define for these values. For
381 /// example, should negative and positive zero hash to different codes? Are
382 /// they equal or not? This hash value implementation specifically
383 /// emphasizes producing different codes for different inputs in order to
384 /// be used in canonicalization and memoization. As such, equality is
385 /// bitwiseIsEqual, and 0 != -0.
386 friend hash_code hash_value(const IEEEFloat &Arg);
388 /// Converts this value into a decimal string.
390 /// \param FormatPrecision The maximum number of digits of
391 /// precision to output. If there are fewer digits available,
392 /// zero padding will not be used unless the value is
393 /// integral and small enough to be expressed in
394 /// FormatPrecision digits. 0 means to use the natural
395 /// precision of the number.
396 /// \param FormatMaxPadding The maximum number of zeros to
397 /// consider inserting before falling back to scientific
398 /// notation. 0 means to always use scientific notation.
400 /// \param TruncateZero Indicate whether to remove the trailing zero in
401 /// fraction part or not. Also setting this parameter to false forcing
402 /// producing of output more similar to default printf behavior.
403 /// Specifically the lower e is used as exponent delimiter and exponent
404 /// always contains no less than two digits.
406 /// Number Precision MaxPadding Result
407 /// ------ --------- ---------- ------
408 /// 1.01E+4 5 2 10100
409 /// 1.01E+4 4 2 1.01E+4
410 /// 1.01E+4 5 1 1.01E+4
411 /// 1.01E-2 5 2 0.0101
412 /// 1.01E-2 4 2 0.0101
413 /// 1.01E-2 4 1 1.01E-2
414 void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0,
415 unsigned FormatMaxPadding = 3, bool TruncateZero = true) const;
417 /// If this value has an exact multiplicative inverse, store it in inv and
419 bool getExactInverse(APFloat *inv) const;
421 /// Returns the exponent of the internal representation of the APFloat.
423 /// Because the radix of APFloat is 2, this is equivalent to floor(log2(x)).
424 /// For special APFloat values, this returns special error codes:
426 /// NaN -> \c IEK_NaN
428 /// Inf -> \c IEK_Inf
430 friend int ilogb(const IEEEFloat &Arg);
432 /// Returns: X * 2^Exp for integral exponents.
433 friend IEEEFloat scalbn(IEEEFloat X, int Exp, roundingMode);
435 friend IEEEFloat frexp(const IEEEFloat &X, int &Exp, roundingMode);
437 /// \name Special value setters.
440 void makeLargest(bool Neg = false);
441 void makeSmallest(bool Neg = false);
442 void makeNaN(bool SNaN = false, bool Neg = false,
443 const APInt *fill = nullptr);
444 void makeInf(bool Neg = false);
445 void makeZero(bool Neg = false);
448 /// Returns the smallest (by magnitude) normalized finite number in the given
451 /// \param Negative - True iff the number should be negative
452 void makeSmallestNormalized(bool Negative = false);
456 cmpResult compareAbsoluteValue(const IEEEFloat &) const;
459 /// \name Simple Queries
462 integerPart *significandParts();
463 const integerPart *significandParts() const;
464 unsigned int partCount() const;
468 /// \name Significand operations.
471 integerPart addSignificand(const IEEEFloat &);
472 integerPart subtractSignificand(const IEEEFloat &, integerPart);
473 lostFraction addOrSubtractSignificand(const IEEEFloat &, bool subtract);
474 lostFraction multiplySignificand(const IEEEFloat &, const IEEEFloat *);
475 lostFraction divideSignificand(const IEEEFloat &);
476 void incrementSignificand();
477 void initialize(const fltSemantics *);
478 void shiftSignificandLeft(unsigned int);
479 lostFraction shiftSignificandRight(unsigned int);
480 unsigned int significandLSB() const;
481 unsigned int significandMSB() const;
482 void zeroSignificand();
483 /// Return true if the significand excluding the integral bit is all ones.
484 bool isSignificandAllOnes() const;
485 /// Return true if the significand excluding the integral bit is all zeros.
486 bool isSignificandAllZeros() const;
490 /// \name Arithmetic on special values.
493 opStatus addOrSubtractSpecials(const IEEEFloat &, bool subtract);
494 opStatus divideSpecials(const IEEEFloat &);
495 opStatus multiplySpecials(const IEEEFloat &);
496 opStatus modSpecials(const IEEEFloat &);
503 bool convertFromStringSpecials(StringRef str);
504 opStatus normalize(roundingMode, lostFraction);
505 opStatus addOrSubtract(const IEEEFloat &, roundingMode, bool subtract);
506 opStatus handleOverflow(roundingMode);
507 bool roundAwayFromZero(roundingMode, lostFraction, unsigned int) const;
508 opStatus convertToSignExtendedInteger(MutableArrayRef<integerPart>,
509 unsigned int, bool, roundingMode,
511 opStatus convertFromUnsignedParts(const integerPart *, unsigned int,
513 opStatus convertFromHexadecimalString(StringRef, roundingMode);
514 opStatus convertFromDecimalString(StringRef, roundingMode);
515 char *convertNormalToHexString(char *, unsigned int, bool,
517 opStatus roundSignificandWithExponent(const integerPart *, unsigned int, int,
522 APInt convertHalfAPFloatToAPInt() const;
523 APInt convertFloatAPFloatToAPInt() const;
524 APInt convertDoubleAPFloatToAPInt() const;
525 APInt convertQuadrupleAPFloatToAPInt() const;
526 APInt convertF80LongDoubleAPFloatToAPInt() const;
527 APInt convertPPCDoubleDoubleAPFloatToAPInt() const;
528 void initFromAPInt(const fltSemantics *Sem, const APInt &api);
529 void initFromHalfAPInt(const APInt &api);
530 void initFromFloatAPInt(const APInt &api);
531 void initFromDoubleAPInt(const APInt &api);
532 void initFromQuadrupleAPInt(const APInt &api);
533 void initFromF80LongDoubleAPInt(const APInt &api);
534 void initFromPPCDoubleDoubleAPInt(const APInt &api);
536 void assign(const IEEEFloat &);
537 void copySignificand(const IEEEFloat &);
538 void freeSignificand();
540 /// Note: this must be the first data member.
541 /// The semantics that this value obeys.
542 const fltSemantics *semantics;
544 /// A binary fraction with an explicit integer bit.
546 /// The significand must be at least one bit wider than the target precision.
552 /// The signed unbiased exponent of the value.
553 ExponentType exponent;
555 /// What kind of floating point number this is.
557 /// Only 2 bits are required, but VisualStudio incorrectly sign extends it.
558 /// Using the extra bit keeps it from failing under VisualStudio.
559 fltCategory category : 3;
561 /// Sign bit of the number.
562 unsigned int sign : 1;
565 hash_code hash_value(const IEEEFloat &Arg);
566 int ilogb(const IEEEFloat &Arg);
567 IEEEFloat scalbn(IEEEFloat X, int Exp, IEEEFloat::roundingMode);
568 IEEEFloat frexp(const IEEEFloat &Val, int &Exp, IEEEFloat::roundingMode RM);
570 // This mode implements more precise float in terms of two APFloats.
571 // The interface and layout is designed for arbitray underlying semantics,
572 // though currently only PPCDoubleDouble semantics are supported, whose
573 // corresponding underlying semantics are IEEEdouble.
574 class DoubleAPFloat final : public APFloatBase {
575 // Note: this must be the first data member.
576 const fltSemantics *Semantics;
577 std::unique_ptr<APFloat[]> Floats;
579 opStatus addImpl(const APFloat &a, const APFloat &aa, const APFloat &c,
580 const APFloat &cc, roundingMode RM);
582 opStatus addWithSpecial(const DoubleAPFloat &LHS, const DoubleAPFloat &RHS,
583 DoubleAPFloat &Out, roundingMode RM);
586 DoubleAPFloat(const fltSemantics &S);
587 DoubleAPFloat(const fltSemantics &S, uninitializedTag);
588 DoubleAPFloat(const fltSemantics &S, integerPart);
589 DoubleAPFloat(const fltSemantics &S, const APInt &I);
590 DoubleAPFloat(const fltSemantics &S, APFloat &&First, APFloat &&Second);
591 DoubleAPFloat(const DoubleAPFloat &RHS);
592 DoubleAPFloat(DoubleAPFloat &&RHS);
594 DoubleAPFloat &operator=(const DoubleAPFloat &RHS);
596 DoubleAPFloat &operator=(DoubleAPFloat &&RHS) {
598 this->~DoubleAPFloat();
599 new (this) DoubleAPFloat(std::move(RHS));
604 bool needsCleanup() const { return Floats != nullptr; }
606 APFloat &getFirst() { return Floats[0]; }
607 const APFloat &getFirst() const { return Floats[0]; }
608 APFloat &getSecond() { return Floats[1]; }
609 const APFloat &getSecond() const { return Floats[1]; }
611 opStatus add(const DoubleAPFloat &RHS, roundingMode RM);
612 opStatus subtract(const DoubleAPFloat &RHS, roundingMode RM);
613 opStatus multiply(const DoubleAPFloat &RHS, roundingMode RM);
614 opStatus divide(const DoubleAPFloat &RHS, roundingMode RM);
615 opStatus remainder(const DoubleAPFloat &RHS);
616 opStatus mod(const DoubleAPFloat &RHS);
617 opStatus fusedMultiplyAdd(const DoubleAPFloat &Multiplicand,
618 const DoubleAPFloat &Addend, roundingMode RM);
619 opStatus roundToIntegral(roundingMode RM);
621 cmpResult compareAbsoluteValue(const DoubleAPFloat &RHS) const;
623 fltCategory getCategory() const;
624 bool isNegative() const;
626 void makeInf(bool Neg);
627 void makeZero(bool Neg);
628 void makeLargest(bool Neg);
629 void makeSmallest(bool Neg);
630 void makeSmallestNormalized(bool Neg);
631 void makeNaN(bool SNaN, bool Neg, const APInt *fill);
633 cmpResult compare(const DoubleAPFloat &RHS) const;
634 bool bitwiseIsEqual(const DoubleAPFloat &RHS) const;
635 APInt bitcastToAPInt() const;
636 opStatus convertFromString(StringRef, roundingMode);
637 opStatus next(bool nextDown);
639 opStatus convertToInteger(MutableArrayRef<integerPart> Input,
640 unsigned int Width, bool IsSigned, roundingMode RM,
641 bool *IsExact) const;
642 opStatus convertFromAPInt(const APInt &Input, bool IsSigned, roundingMode RM);
643 opStatus convertFromSignExtendedInteger(const integerPart *Input,
644 unsigned int InputSize, bool IsSigned,
646 opStatus convertFromZeroExtendedInteger(const integerPart *Input,
647 unsigned int InputSize, bool IsSigned,
649 unsigned int convertToHexString(char *DST, unsigned int HexDigits,
650 bool UpperCase, roundingMode RM) const;
652 bool isDenormal() const;
653 bool isSmallest() const;
654 bool isLargest() const;
655 bool isInteger() const;
657 void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision,
658 unsigned FormatMaxPadding, bool TruncateZero = true) const;
660 bool getExactInverse(APFloat *inv) const;
662 friend int ilogb(const DoubleAPFloat &Arg);
663 friend DoubleAPFloat scalbn(DoubleAPFloat X, int Exp, roundingMode);
664 friend DoubleAPFloat frexp(const DoubleAPFloat &X, int &Exp, roundingMode);
665 friend hash_code hash_value(const DoubleAPFloat &Arg);
668 hash_code hash_value(const DoubleAPFloat &Arg);
670 } // End detail namespace
672 // This is a interface class that is currently forwarding functionalities from
673 // detail::IEEEFloat.
674 class APFloat : public APFloatBase {
675 typedef detail::IEEEFloat IEEEFloat;
676 typedef detail::DoubleAPFloat DoubleAPFloat;
678 static_assert(std::is_standard_layout<IEEEFloat>::value, "");
681 const fltSemantics *semantics;
683 DoubleAPFloat Double;
685 explicit Storage(IEEEFloat F, const fltSemantics &S);
686 explicit Storage(DoubleAPFloat F, const fltSemantics &S)
687 : Double(std::move(F)) {
688 assert(&S == &PPCDoubleDouble());
691 template <typename... ArgTypes>
692 Storage(const fltSemantics &Semantics, ArgTypes &&... Args) {
693 if (usesLayout<IEEEFloat>(Semantics)) {
694 new (&IEEE) IEEEFloat(Semantics, std::forward<ArgTypes>(Args)...);
697 if (usesLayout<DoubleAPFloat>(Semantics)) {
698 new (&Double) DoubleAPFloat(Semantics, std::forward<ArgTypes>(Args)...);
701 llvm_unreachable("Unexpected semantics");
705 if (usesLayout<IEEEFloat>(*semantics)) {
709 if (usesLayout<DoubleAPFloat>(*semantics)) {
710 Double.~DoubleAPFloat();
713 llvm_unreachable("Unexpected semantics");
716 Storage(const Storage &RHS) {
717 if (usesLayout<IEEEFloat>(*RHS.semantics)) {
718 new (this) IEEEFloat(RHS.IEEE);
721 if (usesLayout<DoubleAPFloat>(*RHS.semantics)) {
722 new (this) DoubleAPFloat(RHS.Double);
725 llvm_unreachable("Unexpected semantics");
728 Storage(Storage &&RHS) {
729 if (usesLayout<IEEEFloat>(*RHS.semantics)) {
730 new (this) IEEEFloat(std::move(RHS.IEEE));
733 if (usesLayout<DoubleAPFloat>(*RHS.semantics)) {
734 new (this) DoubleAPFloat(std::move(RHS.Double));
737 llvm_unreachable("Unexpected semantics");
740 Storage &operator=(const Storage &RHS) {
741 if (usesLayout<IEEEFloat>(*semantics) &&
742 usesLayout<IEEEFloat>(*RHS.semantics)) {
744 } else if (usesLayout<DoubleAPFloat>(*semantics) &&
745 usesLayout<DoubleAPFloat>(*RHS.semantics)) {
747 } else if (this != &RHS) {
749 new (this) Storage(RHS);
754 Storage &operator=(Storage &&RHS) {
755 if (usesLayout<IEEEFloat>(*semantics) &&
756 usesLayout<IEEEFloat>(*RHS.semantics)) {
757 IEEE = std::move(RHS.IEEE);
758 } else if (usesLayout<DoubleAPFloat>(*semantics) &&
759 usesLayout<DoubleAPFloat>(*RHS.semantics)) {
760 Double = std::move(RHS.Double);
761 } else if (this != &RHS) {
763 new (this) Storage(std::move(RHS));
769 template <typename T> static bool usesLayout(const fltSemantics &Semantics) {
770 static_assert(std::is_same<T, IEEEFloat>::value ||
771 std::is_same<T, DoubleAPFloat>::value, "");
772 if (std::is_same<T, DoubleAPFloat>::value) {
773 return &Semantics == &PPCDoubleDouble();
775 return &Semantics != &PPCDoubleDouble();
778 IEEEFloat &getIEEE() {
779 if (usesLayout<IEEEFloat>(*U.semantics))
781 if (usesLayout<DoubleAPFloat>(*U.semantics))
782 return U.Double.getFirst().U.IEEE;
783 llvm_unreachable("Unexpected semantics");
786 const IEEEFloat &getIEEE() const {
787 if (usesLayout<IEEEFloat>(*U.semantics))
789 if (usesLayout<DoubleAPFloat>(*U.semantics))
790 return U.Double.getFirst().U.IEEE;
791 llvm_unreachable("Unexpected semantics");
794 void makeZero(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeZero(Neg)); }
796 void makeInf(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeInf(Neg)); }
798 void makeNaN(bool SNaN, bool Neg, const APInt *fill) {
799 APFLOAT_DISPATCH_ON_SEMANTICS(makeNaN(SNaN, Neg, fill));
802 void makeLargest(bool Neg) {
803 APFLOAT_DISPATCH_ON_SEMANTICS(makeLargest(Neg));
806 void makeSmallest(bool Neg) {
807 APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallest(Neg));
810 void makeSmallestNormalized(bool Neg) {
811 APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallestNormalized(Neg));
814 // FIXME: This is due to clang 3.3 (or older version) always checks for the
815 // default constructor in an array aggregate initialization, even if no
816 // elements in the array is default initialized.
817 APFloat() : U(IEEEdouble()) {
818 llvm_unreachable("This is a workaround for old clang.");
821 explicit APFloat(IEEEFloat F, const fltSemantics &S) : U(std::move(F), S) {}
822 explicit APFloat(DoubleAPFloat F, const fltSemantics &S)
823 : U(std::move(F), S) {}
825 cmpResult compareAbsoluteValue(const APFloat &RHS) const {
826 assert(&getSemantics() == &RHS.getSemantics() &&
827 "Should only compare APFloats with the same semantics");
828 if (usesLayout<IEEEFloat>(getSemantics()))
829 return U.IEEE.compareAbsoluteValue(RHS.U.IEEE);
830 if (usesLayout<DoubleAPFloat>(getSemantics()))
831 return U.Double.compareAbsoluteValue(RHS.U.Double);
832 llvm_unreachable("Unexpected semantics");
836 APFloat(const fltSemantics &Semantics) : U(Semantics) {}
837 APFloat(const fltSemantics &Semantics, StringRef S);
838 APFloat(const fltSemantics &Semantics, integerPart I) : U(Semantics, I) {}
839 // TODO: Remove this constructor. This isn't faster than the first one.
840 APFloat(const fltSemantics &Semantics, uninitializedTag)
841 : U(Semantics, uninitialized) {}
842 APFloat(const fltSemantics &Semantics, const APInt &I) : U(Semantics, I) {}
843 explicit APFloat(double d) : U(IEEEFloat(d), IEEEdouble()) {}
844 explicit APFloat(float f) : U(IEEEFloat(f), IEEEsingle()) {}
845 APFloat(const APFloat &RHS) = default;
846 APFloat(APFloat &&RHS) = default;
848 ~APFloat() = default;
850 bool needsCleanup() const { APFLOAT_DISPATCH_ON_SEMANTICS(needsCleanup()); }
852 /// Factory for Positive and Negative Zero.
854 /// \param Negative True iff the number should be negative.
855 static APFloat getZero(const fltSemantics &Sem, bool Negative = false) {
856 APFloat Val(Sem, uninitialized);
857 Val.makeZero(Negative);
861 /// Factory for Positive and Negative Infinity.
863 /// \param Negative True iff the number should be negative.
864 static APFloat getInf(const fltSemantics &Sem, bool Negative = false) {
865 APFloat Val(Sem, uninitialized);
866 Val.makeInf(Negative);
870 /// Factory for NaN values.
872 /// \param Negative - True iff the NaN generated should be negative.
873 /// \param type - The unspecified fill bits for creating the NaN, 0 by
874 /// default. The value is truncated as necessary.
875 static APFloat getNaN(const fltSemantics &Sem, bool Negative = false,
878 APInt fill(64, type);
879 return getQNaN(Sem, Negative, &fill);
881 return getQNaN(Sem, Negative, nullptr);
885 /// Factory for QNaN values.
886 static APFloat getQNaN(const fltSemantics &Sem, bool Negative = false,
887 const APInt *payload = nullptr) {
888 APFloat Val(Sem, uninitialized);
889 Val.makeNaN(false, Negative, payload);
893 /// Factory for SNaN values.
894 static APFloat getSNaN(const fltSemantics &Sem, bool Negative = false,
895 const APInt *payload = nullptr) {
896 APFloat Val(Sem, uninitialized);
897 Val.makeNaN(true, Negative, payload);
901 /// Returns the largest finite number in the given semantics.
903 /// \param Negative - True iff the number should be negative
904 static APFloat getLargest(const fltSemantics &Sem, bool Negative = false) {
905 APFloat Val(Sem, uninitialized);
906 Val.makeLargest(Negative);
910 /// Returns the smallest (by magnitude) finite number in the given semantics.
911 /// Might be denormalized, which implies a relative loss of precision.
913 /// \param Negative - True iff the number should be negative
914 static APFloat getSmallest(const fltSemantics &Sem, bool Negative = false) {
915 APFloat Val(Sem, uninitialized);
916 Val.makeSmallest(Negative);
920 /// Returns the smallest (by magnitude) normalized finite number in the given
923 /// \param Negative - True iff the number should be negative
924 static APFloat getSmallestNormalized(const fltSemantics &Sem,
925 bool Negative = false) {
926 APFloat Val(Sem, uninitialized);
927 Val.makeSmallestNormalized(Negative);
931 /// Returns a float which is bitcasted from an all one value int.
933 /// \param BitWidth - Select float type
934 /// \param isIEEE - If 128 bit number, select between PPC and IEEE
935 static APFloat getAllOnesValue(unsigned BitWidth, bool isIEEE = false);
937 /// Used to insert APFloat objects, or objects that contain APFloat objects,
938 /// into FoldingSets.
939 void Profile(FoldingSetNodeID &NID) const;
941 opStatus add(const APFloat &RHS, roundingMode RM) {
942 assert(&getSemantics() == &RHS.getSemantics() &&
943 "Should only call on two APFloats with the same semantics");
944 if (usesLayout<IEEEFloat>(getSemantics()))
945 return U.IEEE.add(RHS.U.IEEE, RM);
946 if (usesLayout<DoubleAPFloat>(getSemantics()))
947 return U.Double.add(RHS.U.Double, RM);
948 llvm_unreachable("Unexpected semantics");
950 opStatus subtract(const APFloat &RHS, roundingMode RM) {
951 assert(&getSemantics() == &RHS.getSemantics() &&
952 "Should only call on two APFloats with the same semantics");
953 if (usesLayout<IEEEFloat>(getSemantics()))
954 return U.IEEE.subtract(RHS.U.IEEE, RM);
955 if (usesLayout<DoubleAPFloat>(getSemantics()))
956 return U.Double.subtract(RHS.U.Double, RM);
957 llvm_unreachable("Unexpected semantics");
959 opStatus multiply(const APFloat &RHS, roundingMode RM) {
960 assert(&getSemantics() == &RHS.getSemantics() &&
961 "Should only call on two APFloats with the same semantics");
962 if (usesLayout<IEEEFloat>(getSemantics()))
963 return U.IEEE.multiply(RHS.U.IEEE, RM);
964 if (usesLayout<DoubleAPFloat>(getSemantics()))
965 return U.Double.multiply(RHS.U.Double, RM);
966 llvm_unreachable("Unexpected semantics");
968 opStatus divide(const APFloat &RHS, roundingMode RM) {
969 assert(&getSemantics() == &RHS.getSemantics() &&
970 "Should only call on two APFloats with the same semantics");
971 if (usesLayout<IEEEFloat>(getSemantics()))
972 return U.IEEE.divide(RHS.U.IEEE, RM);
973 if (usesLayout<DoubleAPFloat>(getSemantics()))
974 return U.Double.divide(RHS.U.Double, RM);
975 llvm_unreachable("Unexpected semantics");
977 opStatus remainder(const APFloat &RHS) {
978 assert(&getSemantics() == &RHS.getSemantics() &&
979 "Should only call on two APFloats with the same semantics");
980 if (usesLayout<IEEEFloat>(getSemantics()))
981 return U.IEEE.remainder(RHS.U.IEEE);
982 if (usesLayout<DoubleAPFloat>(getSemantics()))
983 return U.Double.remainder(RHS.U.Double);
984 llvm_unreachable("Unexpected semantics");
986 opStatus mod(const APFloat &RHS) {
987 assert(&getSemantics() == &RHS.getSemantics() &&
988 "Should only call on two APFloats with the same semantics");
989 if (usesLayout<IEEEFloat>(getSemantics()))
990 return U.IEEE.mod(RHS.U.IEEE);
991 if (usesLayout<DoubleAPFloat>(getSemantics()))
992 return U.Double.mod(RHS.U.Double);
993 llvm_unreachable("Unexpected semantics");
995 opStatus fusedMultiplyAdd(const APFloat &Multiplicand, const APFloat &Addend,
997 assert(&getSemantics() == &Multiplicand.getSemantics() &&
998 "Should only call on APFloats with the same semantics");
999 assert(&getSemantics() == &Addend.getSemantics() &&
1000 "Should only call on APFloats with the same semantics");
1001 if (usesLayout<IEEEFloat>(getSemantics()))
1002 return U.IEEE.fusedMultiplyAdd(Multiplicand.U.IEEE, Addend.U.IEEE, RM);
1003 if (usesLayout<DoubleAPFloat>(getSemantics()))
1004 return U.Double.fusedMultiplyAdd(Multiplicand.U.Double, Addend.U.Double,
1006 llvm_unreachable("Unexpected semantics");
1008 opStatus roundToIntegral(roundingMode RM) {
1009 APFLOAT_DISPATCH_ON_SEMANTICS(roundToIntegral(RM));
1012 // TODO: bool parameters are not readable and a source of bugs.
1014 opStatus next(bool nextDown) {
1015 APFLOAT_DISPATCH_ON_SEMANTICS(next(nextDown));
1018 /// Add two APFloats, rounding ties to the nearest even.
1019 /// No error checking.
1020 APFloat operator+(const APFloat &RHS) const {
1021 APFloat Result(*this);
1022 (void)Result.add(RHS, rmNearestTiesToEven);
1026 /// Subtract two APFloats, rounding ties to the nearest even.
1027 /// No error checking.
1028 APFloat operator-(const APFloat &RHS) const {
1029 APFloat Result(*this);
1030 (void)Result.subtract(RHS, rmNearestTiesToEven);
1034 /// Multiply two APFloats, rounding ties to the nearest even.
1035 /// No error checking.
1036 APFloat operator*(const APFloat &RHS) const {
1037 APFloat Result(*this);
1038 (void)Result.multiply(RHS, rmNearestTiesToEven);
1042 /// Divide the first APFloat by the second, rounding ties to the nearest even.
1043 /// No error checking.
1044 APFloat operator/(const APFloat &RHS) const {
1045 APFloat Result(*this);
1046 (void)Result.divide(RHS, rmNearestTiesToEven);
1050 void changeSign() { APFLOAT_DISPATCH_ON_SEMANTICS(changeSign()); }
1055 void copySign(const APFloat &RHS) {
1056 if (isNegative() != RHS.isNegative())
1060 /// A static helper to produce a copy of an APFloat value with its sign
1061 /// copied from some other APFloat.
1062 static APFloat copySign(APFloat Value, const APFloat &Sign) {
1063 Value.copySign(Sign);
1067 opStatus convert(const fltSemantics &ToSemantics, roundingMode RM,
1069 opStatus convertToInteger(MutableArrayRef<integerPart> Input,
1070 unsigned int Width, bool IsSigned, roundingMode RM,
1071 bool *IsExact) const {
1072 APFLOAT_DISPATCH_ON_SEMANTICS(
1073 convertToInteger(Input, Width, IsSigned, RM, IsExact));
1075 opStatus convertToInteger(APSInt &Result, roundingMode RM,
1076 bool *IsExact) const;
1077 opStatus convertFromAPInt(const APInt &Input, bool IsSigned,
1079 APFLOAT_DISPATCH_ON_SEMANTICS(convertFromAPInt(Input, IsSigned, RM));
1081 opStatus convertFromSignExtendedInteger(const integerPart *Input,
1082 unsigned int InputSize, bool IsSigned,
1084 APFLOAT_DISPATCH_ON_SEMANTICS(
1085 convertFromSignExtendedInteger(Input, InputSize, IsSigned, RM));
1087 opStatus convertFromZeroExtendedInteger(const integerPart *Input,
1088 unsigned int InputSize, bool IsSigned,
1090 APFLOAT_DISPATCH_ON_SEMANTICS(
1091 convertFromZeroExtendedInteger(Input, InputSize, IsSigned, RM));
1093 opStatus convertFromString(StringRef, roundingMode);
1094 APInt bitcastToAPInt() const {
1095 APFLOAT_DISPATCH_ON_SEMANTICS(bitcastToAPInt());
1097 double convertToDouble() const { return getIEEE().convertToDouble(); }
1098 float convertToFloat() const { return getIEEE().convertToFloat(); }
1100 bool operator==(const APFloat &) const = delete;
1102 cmpResult compare(const APFloat &RHS) const {
1103 assert(&getSemantics() == &RHS.getSemantics() &&
1104 "Should only compare APFloats with the same semantics");
1105 if (usesLayout<IEEEFloat>(getSemantics()))
1106 return U.IEEE.compare(RHS.U.IEEE);
1107 if (usesLayout<DoubleAPFloat>(getSemantics()))
1108 return U.Double.compare(RHS.U.Double);
1109 llvm_unreachable("Unexpected semantics");
1112 bool bitwiseIsEqual(const APFloat &RHS) const {
1113 if (&getSemantics() != &RHS.getSemantics())
1115 if (usesLayout<IEEEFloat>(getSemantics()))
1116 return U.IEEE.bitwiseIsEqual(RHS.U.IEEE);
1117 if (usesLayout<DoubleAPFloat>(getSemantics()))
1118 return U.Double.bitwiseIsEqual(RHS.U.Double);
1119 llvm_unreachable("Unexpected semantics");
1122 unsigned int convertToHexString(char *DST, unsigned int HexDigits,
1123 bool UpperCase, roundingMode RM) const {
1124 APFLOAT_DISPATCH_ON_SEMANTICS(
1125 convertToHexString(DST, HexDigits, UpperCase, RM));
1128 bool isZero() const { return getCategory() == fcZero; }
1129 bool isInfinity() const { return getCategory() == fcInfinity; }
1130 bool isNaN() const { return getCategory() == fcNaN; }
1132 bool isNegative() const { return getIEEE().isNegative(); }
1133 bool isDenormal() const { APFLOAT_DISPATCH_ON_SEMANTICS(isDenormal()); }
1134 bool isSignaling() const { return getIEEE().isSignaling(); }
1136 bool isNormal() const { return !isDenormal() && isFiniteNonZero(); }
1137 bool isFinite() const { return !isNaN() && !isInfinity(); }
1139 fltCategory getCategory() const { return getIEEE().getCategory(); }
1140 const fltSemantics &getSemantics() const { return *U.semantics; }
1141 bool isNonZero() const { return !isZero(); }
1142 bool isFiniteNonZero() const { return isFinite() && !isZero(); }
1143 bool isPosZero() const { return isZero() && !isNegative(); }
1144 bool isNegZero() const { return isZero() && isNegative(); }
1145 bool isSmallest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isSmallest()); }
1146 bool isLargest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isLargest()); }
1147 bool isInteger() const { APFLOAT_DISPATCH_ON_SEMANTICS(isInteger()); }
1149 APFloat &operator=(const APFloat &RHS) = default;
1150 APFloat &operator=(APFloat &&RHS) = default;
1152 void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0,
1153 unsigned FormatMaxPadding = 3, bool TruncateZero = true) const {
1154 APFLOAT_DISPATCH_ON_SEMANTICS(
1155 toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero));
1158 void print(raw_ostream &) const;
1161 bool getExactInverse(APFloat *inv) const {
1162 APFLOAT_DISPATCH_ON_SEMANTICS(getExactInverse(inv));
1165 friend hash_code hash_value(const APFloat &Arg);
1166 friend int ilogb(const APFloat &Arg) { return ilogb(Arg.getIEEE()); }
1167 friend APFloat scalbn(APFloat X, int Exp, roundingMode RM);
1168 friend APFloat frexp(const APFloat &X, int &Exp, roundingMode RM);
1170 friend DoubleAPFloat;
1173 /// See friend declarations above.
1175 /// These additional declarations are required in order to compile LLVM with IBM
1177 hash_code hash_value(const APFloat &Arg);
1178 inline APFloat scalbn(APFloat X, int Exp, APFloat::roundingMode RM) {
1179 if (APFloat::usesLayout<detail::IEEEFloat>(X.getSemantics()))
1180 return APFloat(scalbn(X.U.IEEE, Exp, RM), X.getSemantics());
1181 if (APFloat::usesLayout<detail::DoubleAPFloat>(X.getSemantics()))
1182 return APFloat(scalbn(X.U.Double, Exp, RM), X.getSemantics());
1183 llvm_unreachable("Unexpected semantics");
1186 /// Equivalent of C standard library function.
1188 /// While the C standard says Exp is an unspecified value for infinity and nan,
1189 /// this returns INT_MAX for infinities, and INT_MIN for NaNs.
1190 inline APFloat frexp(const APFloat &X, int &Exp, APFloat::roundingMode RM) {
1191 if (APFloat::usesLayout<detail::IEEEFloat>(X.getSemantics()))
1192 return APFloat(frexp(X.U.IEEE, Exp, RM), X.getSemantics());
1193 if (APFloat::usesLayout<detail::DoubleAPFloat>(X.getSemantics()))
1194 return APFloat(frexp(X.U.Double, Exp, RM), X.getSemantics());
1195 llvm_unreachable("Unexpected semantics");
1197 /// Returns the absolute value of the argument.
1198 inline APFloat abs(APFloat X) {
1203 /// \brief Returns the negated value of the argument.
1204 inline APFloat neg(APFloat X) {
1209 /// Implements IEEE minNum semantics. Returns the smaller of the 2 arguments if
1210 /// both are not NaN. If either argument is a NaN, returns the other argument.
1212 inline APFloat minnum(const APFloat &A, const APFloat &B) {
1217 return (B.compare(A) == APFloat::cmpLessThan) ? B : A;
1220 /// Implements IEEE maxNum semantics. Returns the larger of the 2 arguments if
1221 /// both are not NaN. If either argument is a NaN, returns the other argument.
1223 inline APFloat maxnum(const APFloat &A, const APFloat &B) {
1228 return (A.compare(B) == APFloat::cmpLessThan) ? B : A;
1233 #undef APFLOAT_DISPATCH_ON_SEMANTICS
1234 #endif // LLVM_ADT_APFLOAT_H