1 //== llvm/Support/APFloat.h - Arbitrary Precision Floating Point -*- C++ -*-==//
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 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 Conversion to and from decimal text is not currently implemented.
65 Non-zero finite numbers are represented internally as a sign bit,
66 a 16-bit signed exponent, and the significand as an array of
67 integer parts. After normalization of a number of precision P the
68 exponent is within the range of the format, and if the number is
69 not denormal the P-th bit of the significand is set as an explicit
70 integer bit. For denormals the most significant bit is shifted
71 right so that the exponent is maintained at the format's minimum,
72 so that the smallest denormal has just the least significant bit
73 of the significand set. The sign of zeroes and infinities is
74 significant; the exponent and significand of such numbers is not
75 stored, but has a known implicit (deterministic) value: 0 for the
76 significands, 0 for zero exponent, all 1 bits for infinity
77 exponent. For NaNs the sign and significand are deterministic,
78 although not really meaningful, and preserved in non-conversion
79 operations. The exponent is implicitly all 1 bits.
84 Some features that may or may not be worth adding:
86 Conversions to and from decimal strings (hard).
88 Optional ability to detect underflow tininess before rounding.
90 New formats: x87 in single and double precision mode (IEEE apart
91 from extended exponent range) and IBM two-double extended
94 New operations: sqrt, IEEE remainder, C90 fmod, nextafter,
101 // APInt contains static functions implementing bignum arithmetic.
102 #include "llvm/ADT/APInt.h"
103 #include "llvm/CodeGen/ValueTypes.h"
107 /* Exponents are stored as signed numbers. */
108 typedef signed short exponent_t;
112 /* When bits of a floating point number are truncated, this enum is
113 used to indicate what fraction of the LSB those bits represented.
114 It essentially combines the roles of guard and sticky bits. */
115 enum lostFraction { // Example of truncated bits:
116 lfExactlyZero, // 000000
117 lfLessThanHalf, // 0xxxxx x's not all zero
118 lfExactlyHalf, // 100000
119 lfMoreThanHalf // 1xxxxx x's not all zero
125 /* We support the following floating point semantics. */
126 static const fltSemantics IEEEsingle;
127 static const fltSemantics IEEEdouble;
128 static const fltSemantics IEEEquad;
129 static const fltSemantics PPCDoubleDouble;
130 static const fltSemantics x87DoubleExtended;
131 /* And this psuedo, used to construct APFloats that cannot
132 conflict with anything real. */
133 static const fltSemantics Bogus;
135 static unsigned int semanticsPrecision(const fltSemantics &);
137 /* Floating point numbers have a four-state comparison relation. */
145 /* IEEE-754R gives five rounding modes. */
154 /* Operation status. opUnderflow or opOverflow are always returned
155 or-ed with opInexact. */
165 /* Category of internally-represented number. */
174 APFloat(const fltSemantics &, const char *);
175 APFloat(const fltSemantics &, integerPart);
176 APFloat(const fltSemantics &, fltCategory, bool negative);
177 explicit APFloat(double d);
178 explicit APFloat(float f);
179 explicit APFloat(const APInt &, bool isIEEE = false);
180 APFloat(const APFloat &);
184 opStatus add(const APFloat &, roundingMode);
185 opStatus subtract(const APFloat &, roundingMode);
186 opStatus multiply(const APFloat &, roundingMode);
187 opStatus divide(const APFloat &, roundingMode);
188 opStatus mod(const APFloat &, roundingMode);
189 void copySign(const APFloat &);
190 opStatus fusedMultiplyAdd(const APFloat &, const APFloat &, roundingMode);
191 void changeSign(); // neg
192 void clearSign(); // abs
195 opStatus convert(const fltSemantics &, roundingMode);
196 opStatus convertToInteger(integerPart *, unsigned int, bool,
198 opStatus convertFromSignExtendedInteger(const integerPart *, unsigned int,
200 opStatus convertFromZeroExtendedInteger(const integerPart *, unsigned int,
202 opStatus convertFromString(const char *, roundingMode);
203 APInt convertToAPInt() const;
204 double convertToDouble() const;
205 float convertToFloat() const;
207 /* The definition of equality is not straightforward for floating point,
208 so we won't use operator==. Use one of the following, or write
209 whatever it is you really mean. */
210 // bool operator==(const APFloat &) const; // DO NOT IMPLEMENT
212 /* IEEE comparison with another floating point number (NaNs
213 compare unordered, 0==-0). */
214 cmpResult compare(const APFloat &) const;
216 /* Write out a hexadecimal representation of the floating point
217 value to DST, which must be of sufficient size, in the C99 form
218 [-]0xh.hhhhp[+-]d. Return the number of characters written,
219 excluding the terminating NUL. */
220 unsigned int convertToHexString(char *dst, unsigned int hexDigits,
221 bool upperCase, roundingMode) const;
223 /* Bitwise comparison for equality (QNaNs compare equal, 0!=-0). */
224 bool bitwiseIsEqual(const APFloat &) const;
226 /* Simple queries. */
227 fltCategory getCategory() const { return category; }
228 const fltSemantics &getSemantics() const { return *semantics; }
229 bool isZero() const { return category == fcZero; }
230 bool isNonZero() const { return category != fcZero; }
231 bool isNegative() const { return sign; }
232 bool isPosZero() const { return isZero() && !isNegative(); }
233 bool isNegZero() const { return isZero() && isNegative(); }
235 APFloat& operator=(const APFloat &);
237 /* Return an arbitrary integer value usable for hashing. */
238 uint32_t getHashValue() const;
242 /* Trivial queries. */
243 integerPart *significandParts();
244 const integerPart *significandParts() const;
245 unsigned int partCount() const;
247 /* Significand operations. */
248 integerPart addSignificand(const APFloat &);
249 integerPart subtractSignificand(const APFloat &, integerPart);
250 lostFraction addOrSubtractSignificand(const APFloat &, bool subtract);
251 lostFraction multiplySignificand(const APFloat &, const APFloat *);
252 lostFraction divideSignificand(const APFloat &);
253 void incrementSignificand();
254 void initialize(const fltSemantics *);
255 void shiftSignificandLeft(unsigned int);
256 lostFraction shiftSignificandRight(unsigned int);
257 unsigned int significandLSB() const;
258 unsigned int significandMSB() const;
259 void zeroSignificand();
261 /* Arithmetic on special values. */
262 opStatus addOrSubtractSpecials(const APFloat &, bool subtract);
263 opStatus divideSpecials(const APFloat &);
264 opStatus multiplySpecials(const APFloat &);
267 opStatus normalize(roundingMode, lostFraction);
268 opStatus addOrSubtract(const APFloat &, roundingMode, bool subtract);
269 cmpResult compareAbsoluteValue(const APFloat &) const;
270 opStatus handleOverflow(roundingMode);
271 bool roundAwayFromZero(roundingMode, lostFraction, unsigned int) const;
272 opStatus convertFromUnsignedParts(const integerPart *, unsigned int,
274 opStatus convertFromHexadecimalString(const char *, roundingMode);
275 char *convertNormalToHexString(char *, unsigned int, bool,
277 APInt convertFloatAPFloatToAPInt() const;
278 APInt convertDoubleAPFloatToAPInt() const;
279 APInt convertF80LongDoubleAPFloatToAPInt() const;
280 APInt convertPPCDoubleDoubleAPFloatToAPInt() const;
281 void initFromAPInt(const APInt& api, bool isIEEE = false);
282 void initFromFloatAPInt(const APInt& api);
283 void initFromDoubleAPInt(const APInt& api);
284 void initFromF80LongDoubleAPInt(const APInt& api);
285 void initFromPPCDoubleDoubleAPInt(const APInt& api);
287 void assign(const APFloat &);
288 void copySignificand(const APFloat &);
289 void freeSignificand();
291 /* What kind of semantics does this value obey? */
292 const fltSemantics *semantics;
294 /* Significand - the fraction with an explicit integer bit. Must be
295 at least one bit wider than the target precision. */
302 /* The exponent - a signed number. */
305 /* What kind of floating point number this is. */
306 /* Only 2 bits are required, but VisualStudio incorrectly sign extends
307 it. Using the extra bit keeps it from failing under VisualStudio */
308 fltCategory category: 3;
310 /* The sign bit of this number. */
311 unsigned int sign: 1;
313 /* For PPCDoubleDouble, we have a second exponent and sign (the second
314 significand is appended to the first one, although it would be wrong to
315 regard these as a single number for arithmetic purposes). These fields
316 are not meaningful for any other type. */
317 exponent_t exponent2 : 11;
318 unsigned int sign2: 1;
320 } /* namespace llvm */
322 #endif /* LLVM_FLOAT_H */