1 //== llvm/Support/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 //===----------------------------------------------------------------------===//
10 // This file declares a class to represent arbitrary precision floating
11 // point values and provide a variety of arithmetic operations on them.
13 //===----------------------------------------------------------------------===//
15 /* A self-contained host- and target-independent arbitrary-precision
16 floating-point software implementation. It uses bignum integer
17 arithmetic as provided by static functions in the APInt class.
18 The library will work with bignum integers whose parts are any
19 unsigned type at least 16 bits wide, but 64 bits is recommended.
21 Written for clarity rather than speed, in particular with a view
22 to use in the front-end of a cross compiler so that target
23 arithmetic can be correctly performed on the host. Performance
24 should nonetheless be reasonable, particularly for its intended
25 use. It may be useful as a base implementation for a run-time
26 library during development of a faster target-specific one.
28 All 5 rounding modes in the IEEE-754R draft are handled correctly
29 for all implemented operations. Currently implemented operations
30 are add, subtract, multiply, divide, fused-multiply-add,
31 conversion-to-float, conversion-to-integer and
32 conversion-from-integer. New rounding modes (e.g. away from zero)
33 can be added with three or four lines of code.
35 Four formats are built-in: IEEE single precision, double
36 precision, quadruple precision, and x87 80-bit extended double
37 (when operating with full extended precision). Adding a new
38 format that obeys IEEE semantics only requires adding two lines of
39 code: a declaration and definition of the format.
41 All operations return the status of that operation as an exception
42 bit-mask, so multiple operations can be done consecutively with
43 their results or-ed together. The returned status can be useful
44 for compiler diagnostics; e.g., inexact, underflow and overflow
45 can be easily diagnosed on constant folding, and compiler
46 optimizers can determine what exceptions would be raised by
47 folding operations and optimize, or perhaps not optimize,
50 At present, underflow tininess is detected after rounding; it
51 should be straight forward to add support for the before-rounding
54 The library reads hexadecimal floating point numbers as per C99,
55 and correctly rounds if necessary according to the specified
56 rounding mode. Syntax is required to have been validated by the
57 caller. It also converts floating point numbers to hexadecimal
58 text as per the C99 %a and %A conversions. The output precision
59 (or alternatively the natural minimal precision) can be specified;
60 if the requested precision is less than the natural precision the
61 output is correctly rounded for the specified rounding mode.
63 It also reads decimal floating point numbers and correctly rounds
64 according to the specified rounding mode.
66 Conversion to decimal text is not currently implemented.
68 Non-zero finite numbers are represented internally as a sign bit,
69 a 16-bit signed exponent, and the significand as an array of
70 integer parts. After normalization of a number of precision P the
71 exponent is within the range of the format, and if the number is
72 not denormal the P-th bit of the significand is set as an explicit
73 integer bit. For denormals the most significant bit is shifted
74 right so that the exponent is maintained at the format's minimum,
75 so that the smallest denormal has just the least significant bit
76 of the significand set. The sign of zeroes and infinities is
77 significant; the exponent and significand of such numbers is not
78 stored, but has a known implicit (deterministic) value: 0 for the
79 significands, 0 for zero exponent, all 1 bits for infinity
80 exponent. For NaNs the sign and significand are deterministic,
81 although not really meaningful, and preserved in non-conversion
82 operations. The exponent is implicitly all 1 bits.
87 Some features that may or may not be worth adding:
89 Binary to decimal conversion (hard).
91 Optional ability to detect underflow tininess before rounding.
93 New formats: x87 in single and double precision mode (IEEE apart
94 from extended exponent range) (hard).
96 New operations: sqrt, IEEE remainder, C90 fmod, nextafter,
103 // APInt contains static functions implementing bignum arithmetic.
104 #include "llvm/ADT/APInt.h"
108 /* Exponents are stored as signed numbers. */
109 typedef signed short exponent_t;
114 /* When bits of a floating point number are truncated, this enum is
115 used to indicate what fraction of the LSB those bits represented.
116 It essentially combines the roles of guard and sticky bits. */
117 enum lostFraction { // Example of truncated bits:
118 lfExactlyZero, // 000000
119 lfLessThanHalf, // 0xxxxx x's not all zero
120 lfExactlyHalf, // 100000
121 lfMoreThanHalf // 1xxxxx x's not all zero
127 /* We support the following floating point semantics. */
128 static const fltSemantics IEEEhalf;
129 static const fltSemantics IEEEsingle;
130 static const fltSemantics IEEEdouble;
131 static const fltSemantics IEEEquad;
132 static const fltSemantics PPCDoubleDouble;
133 static const fltSemantics x87DoubleExtended;
134 /* And this pseudo, used to construct APFloats that cannot
135 conflict with anything real. */
136 static const fltSemantics Bogus;
138 static unsigned int semanticsPrecision(const fltSemantics &);
140 /* Floating point numbers have a four-state comparison relation. */
148 /* IEEE-754R gives five rounding modes. */
157 // Operation status. opUnderflow or opOverflow are always returned
158 // or-ed with opInexact.
168 // Category of internally-represented number.
177 APFloat(const fltSemantics &); // Default construct to 0.0
178 APFloat(const fltSemantics &, const StringRef &);
179 APFloat(const fltSemantics &, integerPart);
180 APFloat(const fltSemantics &, fltCategory, bool negative, unsigned type=0);
181 explicit APFloat(double d);
182 explicit APFloat(float f);
183 explicit APFloat(const APInt &, bool isIEEE = false);
184 APFloat(const APFloat &);
187 // Convenience "constructors"
188 static APFloat getZero(const fltSemantics &Sem, bool Negative = false) {
189 return APFloat(Sem, fcZero, Negative);
191 static APFloat getInf(const fltSemantics &Sem, bool Negative = false) {
192 return APFloat(Sem, fcInfinity, Negative);
195 /// getNaN - Factory for QNaN values.
197 /// \param Negative - True iff the NaN generated should be negative.
198 /// \param type - The unspecified fill bits for creating the NaN, 0 by
199 /// default. The value is truncated as necessary.
200 static APFloat getNaN(const fltSemantics &Sem, bool Negative = false,
202 return APFloat(Sem, fcNaN, Negative, type);
205 /// getLargest - Returns the largest finite number in the given
208 /// \param Negative - True iff the number should be negative
209 static APFloat getLargest(const fltSemantics &Sem, bool Negative = false);
211 /// getSmallest - Returns the smallest (by magnitude) finite number
212 /// in the given semantics. Might be denormalized, which implies a
213 /// relative loss of precision.
215 /// \param Negative - True iff the number should be negative
216 static APFloat getSmallest(const fltSemantics &Sem, bool Negative = false);
218 /// getSmallestNormalized - Returns the smallest (by magnitude)
219 /// normalized finite number in the given semantics.
221 /// \param Negative - True iff the number should be negative
222 static APFloat getSmallestNormalized(const fltSemantics &Sem,
223 bool Negative = false);
225 /// Profile - Used to insert APFloat objects, or objects that contain
226 /// APFloat objects, into FoldingSets.
227 void Profile(FoldingSetNodeID& NID) const;
229 /// @brief Used by the Bitcode serializer to emit APInts to Bitcode.
230 void Emit(Serializer& S) const;
232 /// @brief Used by the Bitcode deserializer to deserialize APInts.
233 static APFloat ReadVal(Deserializer& D);
236 opStatus add(const APFloat &, roundingMode);
237 opStatus subtract(const APFloat &, roundingMode);
238 opStatus multiply(const APFloat &, roundingMode);
239 opStatus divide(const APFloat &, roundingMode);
240 /* IEEE remainder. */
241 opStatus remainder(const APFloat &);
242 /* C fmod, or llvm frem. */
243 opStatus mod(const APFloat &, roundingMode);
244 opStatus fusedMultiplyAdd(const APFloat &, const APFloat &, roundingMode);
246 /* Sign operations. */
249 void copySign(const APFloat &);
252 opStatus convert(const fltSemantics &, roundingMode, bool *);
253 opStatus convertToInteger(integerPart *, unsigned int, bool,
254 roundingMode, bool *) const;
255 opStatus convertFromAPInt(const APInt &,
257 opStatus convertFromSignExtendedInteger(const integerPart *, unsigned int,
259 opStatus convertFromZeroExtendedInteger(const integerPart *, unsigned int,
261 opStatus convertFromString(const StringRef&, roundingMode);
262 APInt bitcastToAPInt() const;
263 double convertToDouble() const;
264 float convertToFloat() const;
266 /* The definition of equality is not straightforward for floating point,
267 so we won't use operator==. Use one of the following, or write
268 whatever it is you really mean. */
269 // bool operator==(const APFloat &) const; // DO NOT IMPLEMENT
271 /* IEEE comparison with another floating point number (NaNs
272 compare unordered, 0==-0). */
273 cmpResult compare(const APFloat &) const;
275 /* Bitwise comparison for equality (QNaNs compare equal, 0!=-0). */
276 bool bitwiseIsEqual(const APFloat &) const;
278 /* Write out a hexadecimal representation of the floating point
279 value to DST, which must be of sufficient size, in the C99 form
280 [-]0xh.hhhhp[+-]d. Return the number of characters written,
281 excluding the terminating NUL. */
282 unsigned int convertToHexString(char *dst, unsigned int hexDigits,
283 bool upperCase, roundingMode) const;
285 /* Simple queries. */
286 fltCategory getCategory() const { return category; }
287 const fltSemantics &getSemantics() const { return *semantics; }
288 bool isZero() const { return category == fcZero; }
289 bool isNonZero() const { return category != fcZero; }
290 bool isNaN() const { return category == fcNaN; }
291 bool isInfinity() const { return category == fcInfinity; }
292 bool isNegative() const { return sign; }
293 bool isPosZero() const { return isZero() && !isNegative(); }
294 bool isNegZero() const { return isZero() && isNegative(); }
296 APFloat& operator=(const APFloat &);
298 /* Return an arbitrary integer value usable for hashing. */
299 uint32_t getHashValue() const;
301 /// Converts this value into a decimal string.
303 /// \param FormatPrecision The maximum number of digits of
304 /// precision to output. If there are fewer digits available,
305 /// zero padding will not be used unless the value is
306 /// integral and small enough to be expressed in
307 /// FormatPrecision digits. 0 means to use the natural
308 /// precision of the number.
309 /// \param FormatMaxPadding The maximum number of zeros to
310 /// consider inserting before falling back to scientific
311 /// notation. 0 means to always use scientific notation.
313 /// Number Precision MaxPadding Result
314 /// ------ --------- ---------- ------
315 /// 1.01E+4 5 2 10100
316 /// 1.01E+4 4 2 1.01E+4
317 /// 1.01E+4 5 1 1.01E+4
318 /// 1.01E-2 5 2 0.0101
319 /// 1.01E-2 4 2 0.0101
320 /// 1.01E-2 4 1 1.01E-2
321 void toString(SmallVectorImpl<char> &Str,
322 unsigned FormatPrecision = 0,
323 unsigned FormatMaxPadding = 3);
327 /* Trivial queries. */
328 integerPart *significandParts();
329 const integerPart *significandParts() const;
330 unsigned int partCount() const;
332 /* Significand operations. */
333 integerPart addSignificand(const APFloat &);
334 integerPart subtractSignificand(const APFloat &, integerPart);
335 lostFraction addOrSubtractSignificand(const APFloat &, bool subtract);
336 lostFraction multiplySignificand(const APFloat &, const APFloat *);
337 lostFraction divideSignificand(const APFloat &);
338 void incrementSignificand();
339 void initialize(const fltSemantics *);
340 void shiftSignificandLeft(unsigned int);
341 lostFraction shiftSignificandRight(unsigned int);
342 unsigned int significandLSB() const;
343 unsigned int significandMSB() const;
344 void zeroSignificand();
346 /* Arithmetic on special values. */
347 opStatus addOrSubtractSpecials(const APFloat &, bool subtract);
348 opStatus divideSpecials(const APFloat &);
349 opStatus multiplySpecials(const APFloat &);
350 opStatus modSpecials(const APFloat &);
353 void makeNaN(unsigned = 0);
354 opStatus normalize(roundingMode, lostFraction);
355 opStatus addOrSubtract(const APFloat &, roundingMode, bool subtract);
356 cmpResult compareAbsoluteValue(const APFloat &) const;
357 opStatus handleOverflow(roundingMode);
358 bool roundAwayFromZero(roundingMode, lostFraction, unsigned int) const;
359 opStatus convertToSignExtendedInteger(integerPart *, unsigned int, bool,
360 roundingMode, bool *) const;
361 opStatus convertFromUnsignedParts(const integerPart *, unsigned int,
363 opStatus convertFromHexadecimalString(const StringRef&, roundingMode);
364 opStatus convertFromDecimalString (const StringRef&, roundingMode);
365 char *convertNormalToHexString(char *, unsigned int, bool,
367 opStatus roundSignificandWithExponent(const integerPart *, unsigned int,
370 APInt convertHalfAPFloatToAPInt() const;
371 APInt convertFloatAPFloatToAPInt() const;
372 APInt convertDoubleAPFloatToAPInt() const;
373 APInt convertQuadrupleAPFloatToAPInt() const;
374 APInt convertF80LongDoubleAPFloatToAPInt() const;
375 APInt convertPPCDoubleDoubleAPFloatToAPInt() const;
376 void initFromAPInt(const APInt& api, bool isIEEE = false);
377 void initFromHalfAPInt(const APInt& api);
378 void initFromFloatAPInt(const APInt& api);
379 void initFromDoubleAPInt(const APInt& api);
380 void initFromQuadrupleAPInt(const APInt &api);
381 void initFromF80LongDoubleAPInt(const APInt& api);
382 void initFromPPCDoubleDoubleAPInt(const APInt& api);
384 void assign(const APFloat &);
385 void copySignificand(const APFloat &);
386 void freeSignificand();
388 /* What kind of semantics does this value obey? */
389 const fltSemantics *semantics;
391 /* Significand - the fraction with an explicit integer bit. Must be
392 at least one bit wider than the target precision. */
399 /* The exponent - a signed number. */
402 /* What kind of floating point number this is. */
403 /* Only 2 bits are required, but VisualStudio incorrectly sign extends
404 it. Using the extra bit keeps it from failing under VisualStudio */
405 fltCategory category: 3;
407 /* The sign bit of this number. */
408 unsigned int sign: 1;
410 /* For PPCDoubleDouble, we have a second exponent and sign (the second
411 significand is appended to the first one, although it would be wrong to
412 regard these as a single number for arithmetic purposes). These fields
413 are not meaningful for any other type. */
414 exponent_t exponent2 : 11;
415 unsigned int sign2: 1;
417 } /* namespace llvm */
419 #endif /* LLVM_FLOAT_H */