1 | //===- llvm/ADT/APFloat.h - Arbitrary Precision Floating Point ---*- C++ -*-==// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// |
9 | /// \file |
10 | /// This file declares a class to represent arbitrary precision floating point |
11 | /// values and provide a variety of arithmetic operations on them. |
12 | /// |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #ifndef LLVM_ADT_APFLOAT_H |
16 | #define LLVM_ADT_APFLOAT_H |
17 | |
18 | #include "llvm/ADT/APInt.h" |
19 | #include "llvm/ADT/ArrayRef.h" |
20 | #include "llvm/ADT/FloatingPointMode.h" |
21 | #include "llvm/Support/ErrorHandling.h" |
22 | #include <memory> |
23 | |
24 | #define APFLOAT_DISPATCH_ON_SEMANTICS(METHOD_CALL) \ |
25 | do { \ |
26 | if (usesLayout<IEEEFloat>(getSemantics())) \ |
27 | return U.IEEE.METHOD_CALL; \ |
28 | if (usesLayout<DoubleAPFloat>(getSemantics())) \ |
29 | return U.Double.METHOD_CALL; \ |
30 | llvm_unreachable("Unexpected semantics"); \ |
31 | } while (false) |
32 | |
33 | namespace llvm { |
34 | |
35 | struct fltSemantics; |
36 | class APSInt; |
37 | class StringRef; |
38 | class APFloat; |
39 | class raw_ostream; |
40 | |
41 | template <typename T> class Expected; |
42 | template <typename T> class SmallVectorImpl; |
43 | |
44 | /// Enum that represents what fraction of the LSB truncated bits of an fp number |
45 | /// represent. |
46 | /// |
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 |
53 | }; |
54 | |
55 | /// A self-contained host- and target-independent arbitrary-precision |
56 | /// floating-point software implementation. |
57 | /// |
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. |
61 | /// |
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. |
68 | /// |
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. |
74 | /// |
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 |
79 | /// format. |
80 | /// |
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. |
87 | /// |
88 | /// At present, underflow tininess is detected after rounding; it should be |
89 | /// straight forward to add support for the before-rounding case too. |
90 | /// |
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 |
98 | /// mode. |
99 | /// |
100 | /// It also reads decimal floating point numbers and correctly rounds according |
101 | /// to the specified rounding mode. |
102 | /// |
103 | /// Conversion to decimal text is not currently implemented. |
104 | /// |
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. |
118 | /// |
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 |
122 | /// 0. |
123 | /// |
124 | /// TODO |
125 | /// ==== |
126 | /// |
127 | /// Some features that may or may not be worth adding: |
128 | /// |
129 | /// Binary to decimal conversion (hard). |
130 | /// |
131 | /// Optional ability to detect underflow tininess before rounding. |
132 | /// |
133 | /// New formats: x87 in single and double precision mode (IEEE apart from |
134 | /// extended exponent range) (hard). |
135 | /// |
136 | /// New operations: sqrt, IEEE remainder, C90 fmod, nexttoward. |
137 | /// |
138 | |
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 |
141 | // members. |
142 | struct APFloatBase { |
143 | typedef APInt::WordType integerPart; |
144 | static constexpr unsigned integerPartWidth = APInt::APINT_BITS_PER_WORD; |
145 | |
146 | /// A signed type to represent a floating point numbers unbiased exponent. |
147 | typedef int32_t ExponentType; |
148 | |
149 | /// \name Floating Point Semantics. |
150 | /// @{ |
151 | enum Semantics { |
152 | S_IEEEhalf, |
153 | S_BFloat, |
154 | S_IEEEsingle, |
155 | S_IEEEdouble, |
156 | S_IEEEquad, |
157 | S_PPCDoubleDouble, |
158 | // 8-bit floating point number following IEEE-754 conventions with bit |
159 | // layout S1E5M2 as described in https://arxiv.org/abs/2209.05433. |
160 | S_Float8E5M2, |
161 | // 8-bit floating point number mostly following IEEE-754 conventions |
162 | // and bit layout S1E5M2 described in https://arxiv.org/abs/2206.02915, |
163 | // with expanded range and with no infinity or signed zero. |
164 | // NaN is represented as negative zero. (FN -> Finite, UZ -> unsigned zero). |
165 | // This format's exponent bias is 16, instead of the 15 (2 ** (5 - 1) - 1) |
166 | // that IEEE precedent would imply. |
167 | S_Float8E5M2FNUZ, |
168 | // 8-bit floating point number mostly following IEEE-754 conventions with |
169 | // bit layout S1E4M3 as described in https://arxiv.org/abs/2209.05433. |
170 | // Unlike IEEE-754 types, there are no infinity values, and NaN is |
171 | // represented with the exponent and mantissa bits set to all 1s. |
172 | S_Float8E4M3FN, |
173 | // 8-bit floating point number mostly following IEEE-754 conventions |
174 | // and bit layout S1E4M3 described in https://arxiv.org/abs/2206.02915, |
175 | // with expanded range and with no infinity or signed zero. |
176 | // NaN is represented as negative zero. (FN -> Finite, UZ -> unsigned zero). |
177 | // This format's exponent bias is 8, instead of the 7 (2 ** (4 - 1) - 1) |
178 | // that IEEE precedent would imply. |
179 | S_Float8E4M3FNUZ, |
180 | // 8-bit floating point number mostly following IEEE-754 conventions |
181 | // and bit layout S1E4M3 with expanded range and with no infinity or signed |
182 | // zero. |
183 | // NaN is represented as negative zero. (FN -> Finite, UZ -> unsigned zero). |
184 | // This format's exponent bias is 11, instead of the 7 (2 ** (4 - 1) - 1) |
185 | // that IEEE precedent would imply. |
186 | S_Float8E4M3B11FNUZ, |
187 | // Floating point number that occupies 32 bits or less of storage, providing |
188 | // improved range compared to half (16-bit) formats, at (potentially) |
189 | // greater throughput than single precision (32-bit) formats. |
190 | S_FloatTF32, |
191 | |
192 | S_x87DoubleExtended, |
193 | S_MaxSemantics = S_x87DoubleExtended, |
194 | }; |
195 | |
196 | static const llvm::fltSemantics &EnumToSemantics(Semantics S); |
197 | static Semantics SemanticsToEnum(const llvm::fltSemantics &Sem); |
198 | |
199 | static const fltSemantics &IEEEhalf() LLVM_READNONE; |
200 | static const fltSemantics &BFloat() LLVM_READNONE; |
201 | static const fltSemantics &IEEEsingle() LLVM_READNONE; |
202 | static const fltSemantics &IEEEdouble() LLVM_READNONE; |
203 | static const fltSemantics &IEEEquad() LLVM_READNONE; |
204 | static const fltSemantics &PPCDoubleDouble() LLVM_READNONE; |
205 | static const fltSemantics &Float8E5M2() LLVM_READNONE; |
206 | static const fltSemantics &Float8E5M2FNUZ() LLVM_READNONE; |
207 | static const fltSemantics &Float8E4M3FN() LLVM_READNONE; |
208 | static const fltSemantics &Float8E4M3FNUZ() LLVM_READNONE; |
209 | static const fltSemantics &Float8E4M3B11FNUZ() LLVM_READNONE; |
210 | static const fltSemantics &FloatTF32() LLVM_READNONE; |
211 | static const fltSemantics &x87DoubleExtended() LLVM_READNONE; |
212 | |
213 | /// A Pseudo fltsemantic used to construct APFloats that cannot conflict with |
214 | /// anything real. |
215 | static const fltSemantics &Bogus() LLVM_READNONE; |
216 | |
217 | /// @} |
218 | |
219 | /// IEEE-754R 5.11: Floating Point Comparison Relations. |
220 | enum cmpResult { |
221 | cmpLessThan, |
222 | cmpEqual, |
223 | cmpGreaterThan, |
224 | cmpUnordered |
225 | }; |
226 | |
227 | /// IEEE-754R 4.3: Rounding-direction attributes. |
228 | using roundingMode = llvm::RoundingMode; |
229 | |
230 | static constexpr roundingMode rmNearestTiesToEven = |
231 | RoundingMode::NearestTiesToEven; |
232 | static constexpr roundingMode rmTowardPositive = RoundingMode::TowardPositive; |
233 | static constexpr roundingMode rmTowardNegative = RoundingMode::TowardNegative; |
234 | static constexpr roundingMode rmTowardZero = RoundingMode::TowardZero; |
235 | static constexpr roundingMode rmNearestTiesToAway = |
236 | RoundingMode::NearestTiesToAway; |
237 | |
238 | /// IEEE-754R 7: Default exception handling. |
239 | /// |
240 | /// opUnderflow or opOverflow are always returned or-ed with opInexact. |
241 | /// |
242 | /// APFloat models this behavior specified by IEEE-754: |
243 | /// "For operations producing results in floating-point format, the default |
244 | /// result of an operation that signals the invalid operation exception |
245 | /// shall be a quiet NaN." |
246 | enum opStatus { |
247 | opOK = 0x00, |
248 | opInvalidOp = 0x01, |
249 | opDivByZero = 0x02, |
250 | opOverflow = 0x04, |
251 | opUnderflow = 0x08, |
252 | opInexact = 0x10 |
253 | }; |
254 | |
255 | /// Category of internally-represented number. |
256 | enum fltCategory { |
257 | fcInfinity, |
258 | fcNaN, |
259 | fcNormal, |
260 | fcZero |
261 | }; |
262 | |
263 | /// Convenience enum used to construct an uninitialized APFloat. |
264 | enum uninitializedTag { |
265 | uninitialized |
266 | }; |
267 | |
268 | /// Enumeration of \c ilogb error results. |
269 | enum IlogbErrorKinds { |
270 | IEK_Zero = INT_MIN + 1, |
271 | IEK_NaN = INT_MIN, |
272 | IEK_Inf = INT_MAX |
273 | }; |
274 | |
275 | static unsigned int semanticsPrecision(const fltSemantics &); |
276 | static ExponentType semanticsMinExponent(const fltSemantics &); |
277 | static ExponentType semanticsMaxExponent(const fltSemantics &); |
278 | static unsigned int semanticsSizeInBits(const fltSemantics &); |
279 | static unsigned int semanticsIntSizeInBits(const fltSemantics&, bool); |
280 | |
281 | // Returns true if any number described by \p Src can be precisely represented |
282 | // by a normal (not subnormal) value in \p Dst. |
283 | static bool isRepresentableAsNormalIn(const fltSemantics &Src, |
284 | const fltSemantics &Dst); |
285 | |
286 | /// Returns the size of the floating point number (in bits) in the given |
287 | /// semantics. |
288 | static unsigned getSizeInBits(const fltSemantics &Sem); |
289 | }; |
290 | |
291 | namespace detail { |
292 | |
293 | class IEEEFloat final : public APFloatBase { |
294 | public: |
295 | /// \name Constructors |
296 | /// @{ |
297 | |
298 | IEEEFloat(const fltSemantics &); // Default construct to +0.0 |
299 | IEEEFloat(const fltSemantics &, integerPart); |
300 | IEEEFloat(const fltSemantics &, uninitializedTag); |
301 | IEEEFloat(const fltSemantics &, const APInt &); |
302 | explicit IEEEFloat(double d); |
303 | explicit IEEEFloat(float f); |
304 | IEEEFloat(const IEEEFloat &); |
305 | IEEEFloat(IEEEFloat &&); |
306 | ~IEEEFloat(); |
307 | |
308 | /// @} |
309 | |
310 | /// Returns whether this instance allocated memory. |
311 | bool needsCleanup() const { return partCount() > 1; } |
312 | |
313 | /// \name Convenience "constructors" |
314 | /// @{ |
315 | |
316 | /// @} |
317 | |
318 | /// \name Arithmetic |
319 | /// @{ |
320 | |
321 | opStatus add(const IEEEFloat &, roundingMode); |
322 | opStatus subtract(const IEEEFloat &, roundingMode); |
323 | opStatus multiply(const IEEEFloat &, roundingMode); |
324 | opStatus divide(const IEEEFloat &, roundingMode); |
325 | /// IEEE remainder. |
326 | opStatus remainder(const IEEEFloat &); |
327 | /// C fmod, or llvm frem. |
328 | opStatus mod(const IEEEFloat &); |
329 | opStatus fusedMultiplyAdd(const IEEEFloat &, const IEEEFloat &, roundingMode); |
330 | opStatus roundToIntegral(roundingMode); |
331 | /// IEEE-754R 5.3.1: nextUp/nextDown. |
332 | opStatus next(bool nextDown); |
333 | |
334 | /// @} |
335 | |
336 | /// \name Sign operations. |
337 | /// @{ |
338 | |
339 | void changeSign(); |
340 | |
341 | /// @} |
342 | |
343 | /// \name Conversions |
344 | /// @{ |
345 | |
346 | opStatus convert(const fltSemantics &, roundingMode, bool *); |
347 | opStatus convertToInteger(MutableArrayRef<integerPart>, unsigned int, bool, |
348 | roundingMode, bool *) const; |
349 | opStatus convertFromAPInt(const APInt &, bool, roundingMode); |
350 | opStatus convertFromSignExtendedInteger(const integerPart *, unsigned int, |
351 | bool, roundingMode); |
352 | opStatus convertFromZeroExtendedInteger(const integerPart *, unsigned int, |
353 | bool, roundingMode); |
354 | Expected<opStatus> convertFromString(StringRef, roundingMode); |
355 | APInt bitcastToAPInt() const; |
356 | double convertToDouble() const; |
357 | float convertToFloat() const; |
358 | |
359 | /// @} |
360 | |
361 | /// The definition of equality is not straightforward for floating point, so |
362 | /// we won't use operator==. Use one of the following, or write whatever it |
363 | /// is you really mean. |
364 | bool operator==(const IEEEFloat &) const = delete; |
365 | |
366 | /// IEEE comparison with another floating point number (NaNs compare |
367 | /// unordered, 0==-0). |
368 | cmpResult compare(const IEEEFloat &) const; |
369 | |
370 | /// Bitwise comparison for equality (QNaNs compare equal, 0!=-0). |
371 | bool bitwiseIsEqual(const IEEEFloat &) const; |
372 | |
373 | /// Write out a hexadecimal representation of the floating point value to DST, |
374 | /// which must be of sufficient size, in the C99 form [-]0xh.hhhhp[+-]d. |
375 | /// Return the number of characters written, excluding the terminating NUL. |
376 | unsigned int convertToHexString(char *dst, unsigned int hexDigits, |
377 | bool upperCase, roundingMode) const; |
378 | |
379 | /// \name IEEE-754R 5.7.2 General operations. |
380 | /// @{ |
381 | |
382 | /// IEEE-754R isSignMinus: Returns true if and only if the current value is |
383 | /// negative. |
384 | /// |
385 | /// This applies to zeros and NaNs as well. |
386 | bool isNegative() const { return sign; } |
387 | |
388 | /// IEEE-754R isNormal: Returns true if and only if the current value is normal. |
389 | /// |
390 | /// This implies that the current value of the float is not zero, subnormal, |
391 | /// infinite, or NaN following the definition of normality from IEEE-754R. |
392 | bool isNormal() const { return !isDenormal() && isFiniteNonZero(); } |
393 | |
394 | /// Returns true if and only if the current value is zero, subnormal, or |
395 | /// normal. |
396 | /// |
397 | /// This means that the value is not infinite or NaN. |
398 | bool isFinite() const { return !isNaN() && !isInfinity(); } |
399 | |
400 | /// Returns true if and only if the float is plus or minus zero. |
401 | bool isZero() const { return category == fcZero; } |
402 | |
403 | /// IEEE-754R isSubnormal(): Returns true if and only if the float is a |
404 | /// denormal. |
405 | bool isDenormal() const; |
406 | |
407 | /// IEEE-754R isInfinite(): Returns true if and only if the float is infinity. |
408 | bool isInfinity() const { return category == fcInfinity; } |
409 | |
410 | /// Returns true if and only if the float is a quiet or signaling NaN. |
411 | bool isNaN() const { return category == fcNaN; } |
412 | |
413 | /// Returns true if and only if the float is a signaling NaN. |
414 | bool isSignaling() const; |
415 | |
416 | /// @} |
417 | |
418 | /// \name Simple Queries |
419 | /// @{ |
420 | |
421 | fltCategory getCategory() const { return category; } |
422 | const fltSemantics &getSemantics() const { return *semantics; } |
423 | bool isNonZero() const { return category != fcZero; } |
424 | bool isFiniteNonZero() const { return isFinite() && !isZero(); } |
425 | bool isPosZero() const { return isZero() && !isNegative(); } |
426 | bool isNegZero() const { return isZero() && isNegative(); } |
427 | |
428 | /// Returns true if and only if the number has the smallest possible non-zero |
429 | /// magnitude in the current semantics. |
430 | bool isSmallest() const; |
431 | |
432 | /// Returns true if this is the smallest (by magnitude) normalized finite |
433 | /// number in the given semantics. |
434 | bool isSmallestNormalized() const; |
435 | |
436 | /// Returns true if and only if the number has the largest possible finite |
437 | /// magnitude in the current semantics. |
438 | bool isLargest() const; |
439 | |
440 | /// Returns true if and only if the number is an exact integer. |
441 | bool isInteger() const; |
442 | |
443 | /// @} |
444 | |
445 | IEEEFloat &operator=(const IEEEFloat &); |
446 | IEEEFloat &operator=(IEEEFloat &&); |
447 | |
448 | /// Overload to compute a hash code for an APFloat value. |
449 | /// |
450 | /// Note that the use of hash codes for floating point values is in general |
451 | /// frought with peril. Equality is hard to define for these values. For |
452 | /// example, should negative and positive zero hash to different codes? Are |
453 | /// they equal or not? This hash value implementation specifically |
454 | /// emphasizes producing different codes for different inputs in order to |
455 | /// be used in canonicalization and memoization. As such, equality is |
456 | /// bitwiseIsEqual, and 0 != -0. |
457 | friend hash_code hash_value(const IEEEFloat &Arg); |
458 | |
459 | /// Converts this value into a decimal string. |
460 | /// |
461 | /// \param FormatPrecision The maximum number of digits of |
462 | /// precision to output. If there are fewer digits available, |
463 | /// zero padding will not be used unless the value is |
464 | /// integral and small enough to be expressed in |
465 | /// FormatPrecision digits. 0 means to use the natural |
466 | /// precision of the number. |
467 | /// \param FormatMaxPadding The maximum number of zeros to |
468 | /// consider inserting before falling back to scientific |
469 | /// notation. 0 means to always use scientific notation. |
470 | /// |
471 | /// \param TruncateZero Indicate whether to remove the trailing zero in |
472 | /// fraction part or not. Also setting this parameter to false forcing |
473 | /// producing of output more similar to default printf behavior. |
474 | /// Specifically the lower e is used as exponent delimiter and exponent |
475 | /// always contains no less than two digits. |
476 | /// |
477 | /// Number Precision MaxPadding Result |
478 | /// ------ --------- ---------- ------ |
479 | /// 1.01E+4 5 2 10100 |
480 | /// 1.01E+4 4 2 1.01E+4 |
481 | /// 1.01E+4 5 1 1.01E+4 |
482 | /// 1.01E-2 5 2 0.0101 |
483 | /// 1.01E-2 4 2 0.0101 |
484 | /// 1.01E-2 4 1 1.01E-2 |
485 | void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0, |
486 | unsigned FormatMaxPadding = 3, bool TruncateZero = true) const; |
487 | |
488 | /// If this value has an exact multiplicative inverse, store it in inv and |
489 | /// return true. |
490 | bool getExactInverse(APFloat *inv) const; |
491 | |
492 | /// Returns the exponent of the internal representation of the APFloat. |
493 | /// |
494 | /// Because the radix of APFloat is 2, this is equivalent to floor(log2(x)). |
495 | /// For special APFloat values, this returns special error codes: |
496 | /// |
497 | /// NaN -> \c IEK_NaN |
498 | /// 0 -> \c IEK_Zero |
499 | /// Inf -> \c IEK_Inf |
500 | /// |
501 | friend int ilogb(const IEEEFloat &Arg); |
502 | |
503 | /// Returns: X * 2^Exp for integral exponents. |
504 | friend IEEEFloat scalbn(IEEEFloat X, int Exp, roundingMode); |
505 | |
506 | friend IEEEFloat frexp(const IEEEFloat &X, int &Exp, roundingMode); |
507 | |
508 | /// \name Special value setters. |
509 | /// @{ |
510 | |
511 | void makeLargest(bool Neg = false); |
512 | void makeSmallest(bool Neg = false); |
513 | void makeNaN(bool SNaN = false, bool Neg = false, |
514 | const APInt *fill = nullptr); |
515 | void makeInf(bool Neg = false); |
516 | void makeZero(bool Neg = false); |
517 | void makeQuiet(); |
518 | |
519 | /// Returns the smallest (by magnitude) normalized finite number in the given |
520 | /// semantics. |
521 | /// |
522 | /// \param Negative - True iff the number should be negative |
523 | void makeSmallestNormalized(bool Negative = false); |
524 | |
525 | /// @} |
526 | |
527 | cmpResult compareAbsoluteValue(const IEEEFloat &) const; |
528 | |
529 | private: |
530 | /// \name Simple Queries |
531 | /// @{ |
532 | |
533 | integerPart *significandParts(); |
534 | const integerPart *significandParts() const; |
535 | unsigned int partCount() const; |
536 | |
537 | /// @} |
538 | |
539 | /// \name Significand operations. |
540 | /// @{ |
541 | |
542 | integerPart addSignificand(const IEEEFloat &); |
543 | integerPart subtractSignificand(const IEEEFloat &, integerPart); |
544 | lostFraction addOrSubtractSignificand(const IEEEFloat &, bool subtract); |
545 | lostFraction multiplySignificand(const IEEEFloat &, IEEEFloat); |
546 | lostFraction multiplySignificand(const IEEEFloat&); |
547 | lostFraction divideSignificand(const IEEEFloat &); |
548 | void incrementSignificand(); |
549 | void initialize(const fltSemantics *); |
550 | void shiftSignificandLeft(unsigned int); |
551 | lostFraction shiftSignificandRight(unsigned int); |
552 | unsigned int significandLSB() const; |
553 | unsigned int significandMSB() const; |
554 | void zeroSignificand(); |
555 | /// Return true if the significand excluding the integral bit is all ones. |
556 | bool isSignificandAllOnes() const; |
557 | bool isSignificandAllOnesExceptLSB() const; |
558 | /// Return true if the significand excluding the integral bit is all zeros. |
559 | bool isSignificandAllZeros() const; |
560 | bool isSignificandAllZerosExceptMSB() const; |
561 | |
562 | /// @} |
563 | |
564 | /// \name Arithmetic on special values. |
565 | /// @{ |
566 | |
567 | opStatus addOrSubtractSpecials(const IEEEFloat &, bool subtract); |
568 | opStatus divideSpecials(const IEEEFloat &); |
569 | opStatus multiplySpecials(const IEEEFloat &); |
570 | opStatus modSpecials(const IEEEFloat &); |
571 | opStatus remainderSpecials(const IEEEFloat&); |
572 | |
573 | /// @} |
574 | |
575 | /// \name Miscellany |
576 | /// @{ |
577 | |
578 | bool convertFromStringSpecials(StringRef str); |
579 | opStatus normalize(roundingMode, lostFraction); |
580 | opStatus addOrSubtract(const IEEEFloat &, roundingMode, bool subtract); |
581 | opStatus handleOverflow(roundingMode); |
582 | bool roundAwayFromZero(roundingMode, lostFraction, unsigned int) const; |
583 | opStatus convertToSignExtendedInteger(MutableArrayRef<integerPart>, |
584 | unsigned int, bool, roundingMode, |
585 | bool *) const; |
586 | opStatus convertFromUnsignedParts(const integerPart *, unsigned int, |
587 | roundingMode); |
588 | Expected<opStatus> convertFromHexadecimalString(StringRef, roundingMode); |
589 | Expected<opStatus> convertFromDecimalString(StringRef, roundingMode); |
590 | char *convertNormalToHexString(char *, unsigned int, bool, |
591 | roundingMode) const; |
592 | opStatus roundSignificandWithExponent(const integerPart *, unsigned int, int, |
593 | roundingMode); |
594 | ExponentType exponentNaN() const; |
595 | ExponentType exponentInf() const; |
596 | ExponentType exponentZero() const; |
597 | |
598 | /// @} |
599 | |
600 | template <const fltSemantics &S> APInt convertIEEEFloatToAPInt() const; |
601 | APInt convertHalfAPFloatToAPInt() const; |
602 | APInt convertBFloatAPFloatToAPInt() const; |
603 | APInt convertFloatAPFloatToAPInt() const; |
604 | APInt convertDoubleAPFloatToAPInt() const; |
605 | APInt convertQuadrupleAPFloatToAPInt() const; |
606 | APInt convertF80LongDoubleAPFloatToAPInt() const; |
607 | APInt convertPPCDoubleDoubleAPFloatToAPInt() const; |
608 | APInt convertFloat8E5M2APFloatToAPInt() const; |
609 | APInt convertFloat8E5M2FNUZAPFloatToAPInt() const; |
610 | APInt convertFloat8E4M3FNAPFloatToAPInt() const; |
611 | APInt convertFloat8E4M3FNUZAPFloatToAPInt() const; |
612 | APInt convertFloat8E4M3B11FNUZAPFloatToAPInt() const; |
613 | APInt convertFloatTF32APFloatToAPInt() const; |
614 | void initFromAPInt(const fltSemantics *Sem, const APInt &api); |
615 | template <const fltSemantics &S> void initFromIEEEAPInt(const APInt &api); |
616 | void initFromHalfAPInt(const APInt &api); |
617 | void initFromBFloatAPInt(const APInt &api); |
618 | void initFromFloatAPInt(const APInt &api); |
619 | void initFromDoubleAPInt(const APInt &api); |
620 | void initFromQuadrupleAPInt(const APInt &api); |
621 | void initFromF80LongDoubleAPInt(const APInt &api); |
622 | void initFromPPCDoubleDoubleAPInt(const APInt &api); |
623 | void initFromFloat8E5M2APInt(const APInt &api); |
624 | void initFromFloat8E5M2FNUZAPInt(const APInt &api); |
625 | void initFromFloat8E4M3FNAPInt(const APInt &api); |
626 | void initFromFloat8E4M3FNUZAPInt(const APInt &api); |
627 | void initFromFloat8E4M3B11FNUZAPInt(const APInt &api); |
628 | void initFromFloatTF32APInt(const APInt &api); |
629 | |
630 | void assign(const IEEEFloat &); |
631 | void copySignificand(const IEEEFloat &); |
632 | void freeSignificand(); |
633 | |
634 | /// Note: this must be the first data member. |
635 | /// The semantics that this value obeys. |
636 | const fltSemantics *semantics; |
637 | |
638 | /// A binary fraction with an explicit integer bit. |
639 | /// |
640 | /// The significand must be at least one bit wider than the target precision. |
641 | union Significand { |
642 | integerPart part; |
643 | integerPart *parts; |
644 | } significand; |
645 | |
646 | /// The signed unbiased exponent of the value. |
647 | ExponentType exponent; |
648 | |
649 | /// What kind of floating point number this is. |
650 | /// |
651 | /// Only 2 bits are required, but VisualStudio incorrectly sign extends it. |
652 | /// Using the extra bit keeps it from failing under VisualStudio. |
653 | fltCategory category : 3; |
654 | |
655 | /// Sign bit of the number. |
656 | unsigned int sign : 1; |
657 | }; |
658 | |
659 | hash_code hash_value(const IEEEFloat &Arg); |
660 | int ilogb(const IEEEFloat &Arg); |
661 | IEEEFloat scalbn(IEEEFloat X, int Exp, IEEEFloat::roundingMode); |
662 | IEEEFloat frexp(const IEEEFloat &Val, int &Exp, IEEEFloat::roundingMode RM); |
663 | |
664 | // This mode implements more precise float in terms of two APFloats. |
665 | // The interface and layout is designed for arbitrary underlying semantics, |
666 | // though currently only PPCDoubleDouble semantics are supported, whose |
667 | // corresponding underlying semantics are IEEEdouble. |
668 | class DoubleAPFloat final : public APFloatBase { |
669 | // Note: this must be the first data member. |
670 | const fltSemantics *Semantics; |
671 | std::unique_ptr<APFloat[]> Floats; |
672 | |
673 | opStatus addImpl(const APFloat &a, const APFloat &aa, const APFloat &c, |
674 | const APFloat &cc, roundingMode RM); |
675 | |
676 | opStatus addWithSpecial(const DoubleAPFloat &LHS, const DoubleAPFloat &RHS, |
677 | DoubleAPFloat &Out, roundingMode RM); |
678 | |
679 | public: |
680 | DoubleAPFloat(const fltSemantics &S); |
681 | DoubleAPFloat(const fltSemantics &S, uninitializedTag); |
682 | DoubleAPFloat(const fltSemantics &S, integerPart); |
683 | DoubleAPFloat(const fltSemantics &S, const APInt &I); |
684 | DoubleAPFloat(const fltSemantics &S, APFloat &&First, APFloat &&Second); |
685 | DoubleAPFloat(const DoubleAPFloat &RHS); |
686 | DoubleAPFloat(DoubleAPFloat &&RHS); |
687 | |
688 | DoubleAPFloat &operator=(const DoubleAPFloat &RHS); |
689 | inline DoubleAPFloat &operator=(DoubleAPFloat &&RHS); |
690 | |
691 | bool needsCleanup() const { return Floats != nullptr; } |
692 | |
693 | inline APFloat &getFirst(); |
694 | inline const APFloat &getFirst() const; |
695 | inline APFloat &getSecond(); |
696 | inline const APFloat &getSecond() const; |
697 | |
698 | opStatus add(const DoubleAPFloat &RHS, roundingMode RM); |
699 | opStatus subtract(const DoubleAPFloat &RHS, roundingMode RM); |
700 | opStatus multiply(const DoubleAPFloat &RHS, roundingMode RM); |
701 | opStatus divide(const DoubleAPFloat &RHS, roundingMode RM); |
702 | opStatus remainder(const DoubleAPFloat &RHS); |
703 | opStatus mod(const DoubleAPFloat &RHS); |
704 | opStatus fusedMultiplyAdd(const DoubleAPFloat &Multiplicand, |
705 | const DoubleAPFloat &Addend, roundingMode RM); |
706 | opStatus roundToIntegral(roundingMode RM); |
707 | void changeSign(); |
708 | cmpResult compareAbsoluteValue(const DoubleAPFloat &RHS) const; |
709 | |
710 | fltCategory getCategory() const; |
711 | bool isNegative() const; |
712 | |
713 | void makeInf(bool Neg); |
714 | void makeZero(bool Neg); |
715 | void makeLargest(bool Neg); |
716 | void makeSmallest(bool Neg); |
717 | void makeSmallestNormalized(bool Neg); |
718 | void makeNaN(bool SNaN, bool Neg, const APInt *fill); |
719 | |
720 | cmpResult compare(const DoubleAPFloat &RHS) const; |
721 | bool bitwiseIsEqual(const DoubleAPFloat &RHS) const; |
722 | APInt bitcastToAPInt() const; |
723 | Expected<opStatus> convertFromString(StringRef, roundingMode); |
724 | opStatus next(bool nextDown); |
725 | |
726 | opStatus convertToInteger(MutableArrayRef<integerPart> Input, |
727 | unsigned int Width, bool IsSigned, roundingMode RM, |
728 | bool *IsExact) const; |
729 | opStatus convertFromAPInt(const APInt &Input, bool IsSigned, roundingMode RM); |
730 | opStatus convertFromSignExtendedInteger(const integerPart *Input, |
731 | unsigned int InputSize, bool IsSigned, |
732 | roundingMode RM); |
733 | opStatus convertFromZeroExtendedInteger(const integerPart *Input, |
734 | unsigned int InputSize, bool IsSigned, |
735 | roundingMode RM); |
736 | unsigned int convertToHexString(char *DST, unsigned int HexDigits, |
737 | bool UpperCase, roundingMode RM) const; |
738 | |
739 | bool isDenormal() const; |
740 | bool isSmallest() const; |
741 | bool isSmallestNormalized() const; |
742 | bool isLargest() const; |
743 | bool isInteger() const; |
744 | |
745 | void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision, |
746 | unsigned FormatMaxPadding, bool TruncateZero = true) const; |
747 | |
748 | bool getExactInverse(APFloat *inv) const; |
749 | |
750 | friend DoubleAPFloat scalbn(const DoubleAPFloat &X, int Exp, roundingMode); |
751 | friend DoubleAPFloat frexp(const DoubleAPFloat &X, int &Exp, roundingMode); |
752 | friend hash_code hash_value(const DoubleAPFloat &Arg); |
753 | }; |
754 | |
755 | hash_code hash_value(const DoubleAPFloat &Arg); |
756 | |
757 | } // End detail namespace |
758 | |
759 | // This is a interface class that is currently forwarding functionalities from |
760 | // detail::IEEEFloat. |
761 | class APFloat : public APFloatBase { |
762 | typedef detail::IEEEFloat IEEEFloat; |
763 | typedef detail::DoubleAPFloat DoubleAPFloat; |
764 | |
765 | static_assert(std::is_standard_layout<IEEEFloat>::value); |
766 | |
767 | union Storage { |
768 | const fltSemantics *semantics; |
769 | IEEEFloat IEEE; |
770 | DoubleAPFloat Double; |
771 | |
772 | explicit Storage(IEEEFloat F, const fltSemantics &S); |
773 | explicit Storage(DoubleAPFloat F, const fltSemantics &S) |
774 | : Double(std::move(F)) { |
775 | assert(&S == &PPCDoubleDouble()); |
776 | } |
777 | |
778 | template <typename... ArgTypes> |
779 | Storage(const fltSemantics &Semantics, ArgTypes &&... Args) { |
780 | if (usesLayout<IEEEFloat>(Semantics)) { |
781 | new (&IEEE) IEEEFloat(Semantics, std::forward<ArgTypes>(Args)...); |
782 | return; |
783 | } |
784 | if (usesLayout<DoubleAPFloat>(Semantics)) { |
785 | new (&Double) DoubleAPFloat(Semantics, std::forward<ArgTypes>(Args)...); |
786 | return; |
787 | } |
788 | llvm_unreachable("Unexpected semantics" ); |
789 | } |
790 | |
791 | ~Storage() { |
792 | if (usesLayout<IEEEFloat>(Semantics: *semantics)) { |
793 | IEEE.~IEEEFloat(); |
794 | return; |
795 | } |
796 | if (usesLayout<DoubleAPFloat>(Semantics: *semantics)) { |
797 | Double.~DoubleAPFloat(); |
798 | return; |
799 | } |
800 | llvm_unreachable("Unexpected semantics" ); |
801 | } |
802 | |
803 | Storage(const Storage &RHS) { |
804 | if (usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
805 | new (this) IEEEFloat(RHS.IEEE); |
806 | return; |
807 | } |
808 | if (usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
809 | new (this) DoubleAPFloat(RHS.Double); |
810 | return; |
811 | } |
812 | llvm_unreachable("Unexpected semantics" ); |
813 | } |
814 | |
815 | Storage(Storage &&RHS) { |
816 | if (usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
817 | new (this) IEEEFloat(std::move(RHS.IEEE)); |
818 | return; |
819 | } |
820 | if (usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
821 | new (this) DoubleAPFloat(std::move(RHS.Double)); |
822 | return; |
823 | } |
824 | llvm_unreachable("Unexpected semantics" ); |
825 | } |
826 | |
827 | Storage &operator=(const Storage &RHS) { |
828 | if (usesLayout<IEEEFloat>(Semantics: *semantics) && |
829 | usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
830 | IEEE = RHS.IEEE; |
831 | } else if (usesLayout<DoubleAPFloat>(Semantics: *semantics) && |
832 | usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
833 | Double = RHS.Double; |
834 | } else if (this != &RHS) { |
835 | this->~Storage(); |
836 | new (this) Storage(RHS); |
837 | } |
838 | return *this; |
839 | } |
840 | |
841 | Storage &operator=(Storage &&RHS) { |
842 | if (usesLayout<IEEEFloat>(Semantics: *semantics) && |
843 | usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
844 | IEEE = std::move(RHS.IEEE); |
845 | } else if (usesLayout<DoubleAPFloat>(Semantics: *semantics) && |
846 | usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
847 | Double = std::move(RHS.Double); |
848 | } else if (this != &RHS) { |
849 | this->~Storage(); |
850 | new (this) Storage(std::move(RHS)); |
851 | } |
852 | return *this; |
853 | } |
854 | } U; |
855 | |
856 | template <typename T> static bool usesLayout(const fltSemantics &Semantics) { |
857 | static_assert(std::is_same<T, IEEEFloat>::value || |
858 | std::is_same<T, DoubleAPFloat>::value); |
859 | if (std::is_same<T, DoubleAPFloat>::value) { |
860 | return &Semantics == &PPCDoubleDouble(); |
861 | } |
862 | return &Semantics != &PPCDoubleDouble(); |
863 | } |
864 | |
865 | IEEEFloat &getIEEE() { |
866 | if (usesLayout<IEEEFloat>(Semantics: *U.semantics)) |
867 | return U.IEEE; |
868 | if (usesLayout<DoubleAPFloat>(Semantics: *U.semantics)) |
869 | return U.Double.getFirst().U.IEEE; |
870 | llvm_unreachable("Unexpected semantics" ); |
871 | } |
872 | |
873 | const IEEEFloat &getIEEE() const { |
874 | if (usesLayout<IEEEFloat>(Semantics: *U.semantics)) |
875 | return U.IEEE; |
876 | if (usesLayout<DoubleAPFloat>(Semantics: *U.semantics)) |
877 | return U.Double.getFirst().U.IEEE; |
878 | llvm_unreachable("Unexpected semantics" ); |
879 | } |
880 | |
881 | void makeZero(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeZero(Neg)); } |
882 | |
883 | void makeInf(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeInf(Neg)); } |
884 | |
885 | void makeNaN(bool SNaN, bool Neg, const APInt *fill) { |
886 | APFLOAT_DISPATCH_ON_SEMANTICS(makeNaN(SNaN, Neg, fill)); |
887 | } |
888 | |
889 | void makeLargest(bool Neg) { |
890 | APFLOAT_DISPATCH_ON_SEMANTICS(makeLargest(Neg)); |
891 | } |
892 | |
893 | void makeSmallest(bool Neg) { |
894 | APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallest(Neg)); |
895 | } |
896 | |
897 | void makeSmallestNormalized(bool Neg) { |
898 | APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallestNormalized(Neg)); |
899 | } |
900 | |
901 | explicit APFloat(IEEEFloat F, const fltSemantics &S) : U(std::move(F), S) {} |
902 | explicit APFloat(DoubleAPFloat F, const fltSemantics &S) |
903 | : U(std::move(F), S) {} |
904 | |
905 | cmpResult compareAbsoluteValue(const APFloat &RHS) const { |
906 | assert(&getSemantics() == &RHS.getSemantics() && |
907 | "Should only compare APFloats with the same semantics" ); |
908 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
909 | return U.IEEE.compareAbsoluteValue(RHS.U.IEEE); |
910 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
911 | return U.Double.compareAbsoluteValue(RHS: RHS.U.Double); |
912 | llvm_unreachable("Unexpected semantics" ); |
913 | } |
914 | |
915 | public: |
916 | APFloat(const fltSemantics &Semantics) : U(Semantics) {} |
917 | APFloat(const fltSemantics &Semantics, StringRef S); |
918 | APFloat(const fltSemantics &Semantics, integerPart I) : U(Semantics, I) {} |
919 | template <typename T, |
920 | typename = std::enable_if_t<std::is_floating_point<T>::value>> |
921 | APFloat(const fltSemantics &Semantics, T V) = delete; |
922 | // TODO: Remove this constructor. This isn't faster than the first one. |
923 | APFloat(const fltSemantics &Semantics, uninitializedTag) |
924 | : U(Semantics, uninitialized) {} |
925 | APFloat(const fltSemantics &Semantics, const APInt &I) : U(Semantics, I) {} |
926 | explicit APFloat(double d) : U(IEEEFloat(d), IEEEdouble()) {} |
927 | explicit APFloat(float f) : U(IEEEFloat(f), IEEEsingle()) {} |
928 | APFloat(const APFloat &RHS) = default; |
929 | APFloat(APFloat &&RHS) = default; |
930 | |
931 | ~APFloat() = default; |
932 | |
933 | bool needsCleanup() const { APFLOAT_DISPATCH_ON_SEMANTICS(needsCleanup()); } |
934 | |
935 | /// Factory for Positive and Negative Zero. |
936 | /// |
937 | /// \param Negative True iff the number should be negative. |
938 | static APFloat getZero(const fltSemantics &Sem, bool Negative = false) { |
939 | APFloat Val(Sem, uninitialized); |
940 | Val.makeZero(Neg: Negative); |
941 | return Val; |
942 | } |
943 | |
944 | /// Factory for Positive and Negative Infinity. |
945 | /// |
946 | /// \param Negative True iff the number should be negative. |
947 | static APFloat getInf(const fltSemantics &Sem, bool Negative = false) { |
948 | APFloat Val(Sem, uninitialized); |
949 | Val.makeInf(Neg: Negative); |
950 | return Val; |
951 | } |
952 | |
953 | /// Factory for NaN values. |
954 | /// |
955 | /// \param Negative - True iff the NaN generated should be negative. |
956 | /// \param payload - The unspecified fill bits for creating the NaN, 0 by |
957 | /// default. The value is truncated as necessary. |
958 | static APFloat getNaN(const fltSemantics &Sem, bool Negative = false, |
959 | uint64_t payload = 0) { |
960 | if (payload) { |
961 | APInt intPayload(64, payload); |
962 | return getQNaN(Sem, Negative, payload: &intPayload); |
963 | } else { |
964 | return getQNaN(Sem, Negative, payload: nullptr); |
965 | } |
966 | } |
967 | |
968 | /// Factory for QNaN values. |
969 | static APFloat getQNaN(const fltSemantics &Sem, bool Negative = false, |
970 | const APInt *payload = nullptr) { |
971 | APFloat Val(Sem, uninitialized); |
972 | Val.makeNaN(SNaN: false, Neg: Negative, fill: payload); |
973 | return Val; |
974 | } |
975 | |
976 | /// Factory for SNaN values. |
977 | static APFloat getSNaN(const fltSemantics &Sem, bool Negative = false, |
978 | const APInt *payload = nullptr) { |
979 | APFloat Val(Sem, uninitialized); |
980 | Val.makeNaN(SNaN: true, Neg: Negative, fill: payload); |
981 | return Val; |
982 | } |
983 | |
984 | /// Returns the largest finite number in the given semantics. |
985 | /// |
986 | /// \param Negative - True iff the number should be negative |
987 | static APFloat getLargest(const fltSemantics &Sem, bool Negative = false) { |
988 | APFloat Val(Sem, uninitialized); |
989 | Val.makeLargest(Neg: Negative); |
990 | return Val; |
991 | } |
992 | |
993 | /// Returns the smallest (by magnitude) finite number in the given semantics. |
994 | /// Might be denormalized, which implies a relative loss of precision. |
995 | /// |
996 | /// \param Negative - True iff the number should be negative |
997 | static APFloat getSmallest(const fltSemantics &Sem, bool Negative = false) { |
998 | APFloat Val(Sem, uninitialized); |
999 | Val.makeSmallest(Neg: Negative); |
1000 | return Val; |
1001 | } |
1002 | |
1003 | /// Returns the smallest (by magnitude) normalized finite number in the given |
1004 | /// semantics. |
1005 | /// |
1006 | /// \param Negative - True iff the number should be negative |
1007 | static APFloat getSmallestNormalized(const fltSemantics &Sem, |
1008 | bool Negative = false) { |
1009 | APFloat Val(Sem, uninitialized); |
1010 | Val.makeSmallestNormalized(Neg: Negative); |
1011 | return Val; |
1012 | } |
1013 | |
1014 | /// Returns a float which is bitcasted from an all one value int. |
1015 | /// |
1016 | /// \param Semantics - type float semantics |
1017 | static APFloat getAllOnesValue(const fltSemantics &Semantics); |
1018 | |
1019 | /// Used to insert APFloat objects, or objects that contain APFloat objects, |
1020 | /// into FoldingSets. |
1021 | void Profile(FoldingSetNodeID &NID) const; |
1022 | |
1023 | opStatus add(const APFloat &RHS, roundingMode RM) { |
1024 | assert(&getSemantics() == &RHS.getSemantics() && |
1025 | "Should only call on two APFloats with the same semantics" ); |
1026 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1027 | return U.IEEE.add(RHS.U.IEEE, RM); |
1028 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1029 | return U.Double.add(RHS: RHS.U.Double, RM); |
1030 | llvm_unreachable("Unexpected semantics" ); |
1031 | } |
1032 | opStatus subtract(const APFloat &RHS, roundingMode RM) { |
1033 | assert(&getSemantics() == &RHS.getSemantics() && |
1034 | "Should only call on two APFloats with the same semantics" ); |
1035 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1036 | return U.IEEE.subtract(RHS.U.IEEE, RM); |
1037 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1038 | return U.Double.subtract(RHS: RHS.U.Double, RM); |
1039 | llvm_unreachable("Unexpected semantics" ); |
1040 | } |
1041 | opStatus multiply(const APFloat &RHS, roundingMode RM) { |
1042 | assert(&getSemantics() == &RHS.getSemantics() && |
1043 | "Should only call on two APFloats with the same semantics" ); |
1044 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1045 | return U.IEEE.multiply(RHS.U.IEEE, RM); |
1046 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1047 | return U.Double.multiply(RHS: RHS.U.Double, RM); |
1048 | llvm_unreachable("Unexpected semantics" ); |
1049 | } |
1050 | opStatus divide(const APFloat &RHS, roundingMode RM) { |
1051 | assert(&getSemantics() == &RHS.getSemantics() && |
1052 | "Should only call on two APFloats with the same semantics" ); |
1053 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1054 | return U.IEEE.divide(RHS.U.IEEE, RM); |
1055 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1056 | return U.Double.divide(RHS: RHS.U.Double, RM); |
1057 | llvm_unreachable("Unexpected semantics" ); |
1058 | } |
1059 | opStatus remainder(const APFloat &RHS) { |
1060 | assert(&getSemantics() == &RHS.getSemantics() && |
1061 | "Should only call on two APFloats with the same semantics" ); |
1062 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1063 | return U.IEEE.remainder(RHS.U.IEEE); |
1064 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1065 | return U.Double.remainder(RHS: RHS.U.Double); |
1066 | llvm_unreachable("Unexpected semantics" ); |
1067 | } |
1068 | opStatus mod(const APFloat &RHS) { |
1069 | assert(&getSemantics() == &RHS.getSemantics() && |
1070 | "Should only call on two APFloats with the same semantics" ); |
1071 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1072 | return U.IEEE.mod(RHS.U.IEEE); |
1073 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1074 | return U.Double.mod(RHS: RHS.U.Double); |
1075 | llvm_unreachable("Unexpected semantics" ); |
1076 | } |
1077 | opStatus fusedMultiplyAdd(const APFloat &Multiplicand, const APFloat &Addend, |
1078 | roundingMode RM) { |
1079 | assert(&getSemantics() == &Multiplicand.getSemantics() && |
1080 | "Should only call on APFloats with the same semantics" ); |
1081 | assert(&getSemantics() == &Addend.getSemantics() && |
1082 | "Should only call on APFloats with the same semantics" ); |
1083 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1084 | return U.IEEE.fusedMultiplyAdd(Multiplicand.U.IEEE, Addend.U.IEEE, RM); |
1085 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1086 | return U.Double.fusedMultiplyAdd(Multiplicand: Multiplicand.U.Double, Addend: Addend.U.Double, |
1087 | RM); |
1088 | llvm_unreachable("Unexpected semantics" ); |
1089 | } |
1090 | opStatus roundToIntegral(roundingMode RM) { |
1091 | APFLOAT_DISPATCH_ON_SEMANTICS(roundToIntegral(RM)); |
1092 | } |
1093 | |
1094 | // TODO: bool parameters are not readable and a source of bugs. |
1095 | // Do something. |
1096 | opStatus next(bool nextDown) { |
1097 | APFLOAT_DISPATCH_ON_SEMANTICS(next(nextDown)); |
1098 | } |
1099 | |
1100 | /// Negate an APFloat. |
1101 | APFloat operator-() const { |
1102 | APFloat Result(*this); |
1103 | Result.changeSign(); |
1104 | return Result; |
1105 | } |
1106 | |
1107 | /// Add two APFloats, rounding ties to the nearest even. |
1108 | /// No error checking. |
1109 | APFloat operator+(const APFloat &RHS) const { |
1110 | APFloat Result(*this); |
1111 | (void)Result.add(RHS, RM: rmNearestTiesToEven); |
1112 | return Result; |
1113 | } |
1114 | |
1115 | /// Subtract two APFloats, rounding ties to the nearest even. |
1116 | /// No error checking. |
1117 | APFloat operator-(const APFloat &RHS) const { |
1118 | APFloat Result(*this); |
1119 | (void)Result.subtract(RHS, RM: rmNearestTiesToEven); |
1120 | return Result; |
1121 | } |
1122 | |
1123 | /// Multiply two APFloats, rounding ties to the nearest even. |
1124 | /// No error checking. |
1125 | APFloat operator*(const APFloat &RHS) const { |
1126 | APFloat Result(*this); |
1127 | (void)Result.multiply(RHS, RM: rmNearestTiesToEven); |
1128 | return Result; |
1129 | } |
1130 | |
1131 | /// Divide the first APFloat by the second, rounding ties to the nearest even. |
1132 | /// No error checking. |
1133 | APFloat operator/(const APFloat &RHS) const { |
1134 | APFloat Result(*this); |
1135 | (void)Result.divide(RHS, RM: rmNearestTiesToEven); |
1136 | return Result; |
1137 | } |
1138 | |
1139 | void changeSign() { APFLOAT_DISPATCH_ON_SEMANTICS(changeSign()); } |
1140 | void clearSign() { |
1141 | if (isNegative()) |
1142 | changeSign(); |
1143 | } |
1144 | void copySign(const APFloat &RHS) { |
1145 | if (isNegative() != RHS.isNegative()) |
1146 | changeSign(); |
1147 | } |
1148 | |
1149 | /// A static helper to produce a copy of an APFloat value with its sign |
1150 | /// copied from some other APFloat. |
1151 | static APFloat copySign(APFloat Value, const APFloat &Sign) { |
1152 | Value.copySign(RHS: Sign); |
1153 | return Value; |
1154 | } |
1155 | |
1156 | /// Assuming this is an IEEE-754 NaN value, quiet its signaling bit. |
1157 | /// This preserves the sign and payload bits. |
1158 | APFloat makeQuiet() const { |
1159 | APFloat Result(*this); |
1160 | Result.getIEEE().makeQuiet(); |
1161 | return Result; |
1162 | } |
1163 | |
1164 | opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, |
1165 | bool *losesInfo); |
1166 | opStatus convertToInteger(MutableArrayRef<integerPart> Input, |
1167 | unsigned int Width, bool IsSigned, roundingMode RM, |
1168 | bool *IsExact) const { |
1169 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1170 | convertToInteger(Input, Width, IsSigned, RM, IsExact)); |
1171 | } |
1172 | opStatus convertToInteger(APSInt &Result, roundingMode RM, |
1173 | bool *IsExact) const; |
1174 | opStatus convertFromAPInt(const APInt &Input, bool IsSigned, |
1175 | roundingMode RM) { |
1176 | APFLOAT_DISPATCH_ON_SEMANTICS(convertFromAPInt(Input, IsSigned, RM)); |
1177 | } |
1178 | opStatus convertFromSignExtendedInteger(const integerPart *Input, |
1179 | unsigned int InputSize, bool IsSigned, |
1180 | roundingMode RM) { |
1181 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1182 | convertFromSignExtendedInteger(Input, InputSize, IsSigned, RM)); |
1183 | } |
1184 | opStatus convertFromZeroExtendedInteger(const integerPart *Input, |
1185 | unsigned int InputSize, bool IsSigned, |
1186 | roundingMode RM) { |
1187 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1188 | convertFromZeroExtendedInteger(Input, InputSize, IsSigned, RM)); |
1189 | } |
1190 | Expected<opStatus> convertFromString(StringRef, roundingMode); |
1191 | APInt bitcastToAPInt() const { |
1192 | APFLOAT_DISPATCH_ON_SEMANTICS(bitcastToAPInt()); |
1193 | } |
1194 | |
1195 | /// Converts this APFloat to host double value. |
1196 | /// |
1197 | /// \pre The APFloat must be built using semantics, that can be represented by |
1198 | /// the host double type without loss of precision. It can be IEEEdouble and |
1199 | /// shorter semantics, like IEEEsingle and others. |
1200 | double convertToDouble() const; |
1201 | |
1202 | /// Converts this APFloat to host float value. |
1203 | /// |
1204 | /// \pre The APFloat must be built using semantics, that can be represented by |
1205 | /// the host float type without loss of precision. It can be IEEEsingle and |
1206 | /// shorter semantics, like IEEEhalf. |
1207 | float convertToFloat() const; |
1208 | |
1209 | bool operator==(const APFloat &RHS) const { return compare(RHS) == cmpEqual; } |
1210 | |
1211 | bool operator!=(const APFloat &RHS) const { return compare(RHS) != cmpEqual; } |
1212 | |
1213 | bool operator<(const APFloat &RHS) const { |
1214 | return compare(RHS) == cmpLessThan; |
1215 | } |
1216 | |
1217 | bool operator>(const APFloat &RHS) const { |
1218 | return compare(RHS) == cmpGreaterThan; |
1219 | } |
1220 | |
1221 | bool operator<=(const APFloat &RHS) const { |
1222 | cmpResult Res = compare(RHS); |
1223 | return Res == cmpLessThan || Res == cmpEqual; |
1224 | } |
1225 | |
1226 | bool operator>=(const APFloat &RHS) const { |
1227 | cmpResult Res = compare(RHS); |
1228 | return Res == cmpGreaterThan || Res == cmpEqual; |
1229 | } |
1230 | |
1231 | cmpResult compare(const APFloat &RHS) const { |
1232 | assert(&getSemantics() == &RHS.getSemantics() && |
1233 | "Should only compare APFloats with the same semantics" ); |
1234 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1235 | return U.IEEE.compare(RHS.U.IEEE); |
1236 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1237 | return U.Double.compare(RHS: RHS.U.Double); |
1238 | llvm_unreachable("Unexpected semantics" ); |
1239 | } |
1240 | |
1241 | bool bitwiseIsEqual(const APFloat &RHS) const { |
1242 | if (&getSemantics() != &RHS.getSemantics()) |
1243 | return false; |
1244 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1245 | return U.IEEE.bitwiseIsEqual(RHS.U.IEEE); |
1246 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1247 | return U.Double.bitwiseIsEqual(RHS: RHS.U.Double); |
1248 | llvm_unreachable("Unexpected semantics" ); |
1249 | } |
1250 | |
1251 | /// We don't rely on operator== working on double values, as |
1252 | /// it returns true for things that are clearly not equal, like -0.0 and 0.0. |
1253 | /// As such, this method can be used to do an exact bit-for-bit comparison of |
1254 | /// two floating point values. |
1255 | /// |
1256 | /// We leave the version with the double argument here because it's just so |
1257 | /// convenient to write "2.0" and the like. Without this function we'd |
1258 | /// have to duplicate its logic everywhere it's called. |
1259 | bool isExactlyValue(double V) const { |
1260 | bool ignored; |
1261 | APFloat Tmp(V); |
1262 | Tmp.convert(ToSemantics: getSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &ignored); |
1263 | return bitwiseIsEqual(RHS: Tmp); |
1264 | } |
1265 | |
1266 | unsigned int convertToHexString(char *DST, unsigned int HexDigits, |
1267 | bool UpperCase, roundingMode RM) const { |
1268 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1269 | convertToHexString(DST, HexDigits, UpperCase, RM)); |
1270 | } |
1271 | |
1272 | bool isZero() const { return getCategory() == fcZero; } |
1273 | bool isInfinity() const { return getCategory() == fcInfinity; } |
1274 | bool isNaN() const { return getCategory() == fcNaN; } |
1275 | |
1276 | bool isNegative() const { return getIEEE().isNegative(); } |
1277 | bool isDenormal() const { APFLOAT_DISPATCH_ON_SEMANTICS(isDenormal()); } |
1278 | bool isSignaling() const { return getIEEE().isSignaling(); } |
1279 | |
1280 | bool isNormal() const { return !isDenormal() && isFiniteNonZero(); } |
1281 | bool isFinite() const { return !isNaN() && !isInfinity(); } |
1282 | |
1283 | fltCategory getCategory() const { return getIEEE().getCategory(); } |
1284 | const fltSemantics &getSemantics() const { return *U.semantics; } |
1285 | bool isNonZero() const { return !isZero(); } |
1286 | bool isFiniteNonZero() const { return isFinite() && !isZero(); } |
1287 | bool isPosZero() const { return isZero() && !isNegative(); } |
1288 | bool isNegZero() const { return isZero() && isNegative(); } |
1289 | bool isPosInfinity() const { return isInfinity() && !isNegative(); } |
1290 | bool isNegInfinity() const { return isInfinity() && isNegative(); } |
1291 | bool isSmallest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isSmallest()); } |
1292 | bool isLargest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isLargest()); } |
1293 | bool isInteger() const { APFLOAT_DISPATCH_ON_SEMANTICS(isInteger()); } |
1294 | bool isIEEE() const { return usesLayout<IEEEFloat>(Semantics: getSemantics()); } |
1295 | |
1296 | bool isSmallestNormalized() const { |
1297 | APFLOAT_DISPATCH_ON_SEMANTICS(isSmallestNormalized()); |
1298 | } |
1299 | |
1300 | /// Return the FPClassTest which will return true for the value. |
1301 | FPClassTest classify() const; |
1302 | |
1303 | APFloat &operator=(const APFloat &RHS) = default; |
1304 | APFloat &operator=(APFloat &&RHS) = default; |
1305 | |
1306 | void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0, |
1307 | unsigned FormatMaxPadding = 3, bool TruncateZero = true) const { |
1308 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1309 | toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero)); |
1310 | } |
1311 | |
1312 | void print(raw_ostream &) const; |
1313 | void dump() const; |
1314 | |
1315 | bool getExactInverse(APFloat *inv) const { |
1316 | APFLOAT_DISPATCH_ON_SEMANTICS(getExactInverse(inv)); |
1317 | } |
1318 | |
1319 | friend hash_code hash_value(const APFloat &Arg); |
1320 | friend int ilogb(const APFloat &Arg) { return ilogb(Arg: Arg.getIEEE()); } |
1321 | friend APFloat scalbn(APFloat X, int Exp, roundingMode RM); |
1322 | friend APFloat frexp(const APFloat &X, int &Exp, roundingMode RM); |
1323 | friend IEEEFloat; |
1324 | friend DoubleAPFloat; |
1325 | }; |
1326 | |
1327 | /// See friend declarations above. |
1328 | /// |
1329 | /// These additional declarations are required in order to compile LLVM with IBM |
1330 | /// xlC compiler. |
1331 | hash_code hash_value(const APFloat &Arg); |
1332 | inline APFloat scalbn(APFloat X, int Exp, APFloat::roundingMode RM) { |
1333 | if (APFloat::usesLayout<detail::IEEEFloat>(Semantics: X.getSemantics())) |
1334 | return APFloat(scalbn(X: X.U.IEEE, Exp, RM), X.getSemantics()); |
1335 | if (APFloat::usesLayout<detail::DoubleAPFloat>(Semantics: X.getSemantics())) |
1336 | return APFloat(scalbn(X: X.U.Double, Exp, RM), X.getSemantics()); |
1337 | llvm_unreachable("Unexpected semantics" ); |
1338 | } |
1339 | |
1340 | /// Equivalent of C standard library function. |
1341 | /// |
1342 | /// While the C standard says Exp is an unspecified value for infinity and nan, |
1343 | /// this returns INT_MAX for infinities, and INT_MIN for NaNs. |
1344 | inline APFloat frexp(const APFloat &X, int &Exp, APFloat::roundingMode RM) { |
1345 | if (APFloat::usesLayout<detail::IEEEFloat>(Semantics: X.getSemantics())) |
1346 | return APFloat(frexp(Val: X.U.IEEE, Exp, RM), X.getSemantics()); |
1347 | if (APFloat::usesLayout<detail::DoubleAPFloat>(Semantics: X.getSemantics())) |
1348 | return APFloat(frexp(X: X.U.Double, Exp, RM), X.getSemantics()); |
1349 | llvm_unreachable("Unexpected semantics" ); |
1350 | } |
1351 | /// Returns the absolute value of the argument. |
1352 | inline APFloat abs(APFloat X) { |
1353 | X.clearSign(); |
1354 | return X; |
1355 | } |
1356 | |
1357 | /// Returns the negated value of the argument. |
1358 | inline APFloat neg(APFloat X) { |
1359 | X.changeSign(); |
1360 | return X; |
1361 | } |
1362 | |
1363 | /// Implements IEEE minNum semantics. Returns the smaller of the 2 arguments if |
1364 | /// both are not NaN. If either argument is a NaN, returns the other argument. |
1365 | LLVM_READONLY |
1366 | inline APFloat minnum(const APFloat &A, const APFloat &B) { |
1367 | if (A.isNaN()) |
1368 | return B; |
1369 | if (B.isNaN()) |
1370 | return A; |
1371 | return B < A ? B : A; |
1372 | } |
1373 | |
1374 | /// Implements IEEE maxNum semantics. Returns the larger of the 2 arguments if |
1375 | /// both are not NaN. If either argument is a NaN, returns the other argument. |
1376 | LLVM_READONLY |
1377 | inline APFloat maxnum(const APFloat &A, const APFloat &B) { |
1378 | if (A.isNaN()) |
1379 | return B; |
1380 | if (B.isNaN()) |
1381 | return A; |
1382 | return A < B ? B : A; |
1383 | } |
1384 | |
1385 | /// Implements IEEE 754-2018 minimum semantics. Returns the smaller of 2 |
1386 | /// arguments, propagating NaNs and treating -0 as less than +0. |
1387 | LLVM_READONLY |
1388 | inline APFloat minimum(const APFloat &A, const APFloat &B) { |
1389 | if (A.isNaN()) |
1390 | return A; |
1391 | if (B.isNaN()) |
1392 | return B; |
1393 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1394 | return A.isNegative() ? A : B; |
1395 | return B < A ? B : A; |
1396 | } |
1397 | |
1398 | /// Implements IEEE 754-2018 maximum semantics. Returns the larger of 2 |
1399 | /// arguments, propagating NaNs and treating -0 as less than +0. |
1400 | LLVM_READONLY |
1401 | inline APFloat maximum(const APFloat &A, const APFloat &B) { |
1402 | if (A.isNaN()) |
1403 | return A; |
1404 | if (B.isNaN()) |
1405 | return B; |
1406 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1407 | return A.isNegative() ? B : A; |
1408 | return A < B ? B : A; |
1409 | } |
1410 | |
1411 | // We want the following functions to be available in the header for inlining. |
1412 | // We cannot define them inline in the class definition of `DoubleAPFloat` |
1413 | // because doing so would instantiate `std::unique_ptr<APFloat[]>` before |
1414 | // `APFloat` is defined, and that would be undefined behavior. |
1415 | namespace detail { |
1416 | |
1417 | DoubleAPFloat &DoubleAPFloat::operator=(DoubleAPFloat &&RHS) { |
1418 | if (this != &RHS) { |
1419 | this->~DoubleAPFloat(); |
1420 | new (this) DoubleAPFloat(std::move(RHS)); |
1421 | } |
1422 | return *this; |
1423 | } |
1424 | |
1425 | APFloat &DoubleAPFloat::getFirst() { return Floats[0]; } |
1426 | const APFloat &DoubleAPFloat::getFirst() const { return Floats[0]; } |
1427 | APFloat &DoubleAPFloat::getSecond() { return Floats[1]; } |
1428 | const APFloat &DoubleAPFloat::getSecond() const { return Floats[1]; } |
1429 | |
1430 | } // namespace detail |
1431 | |
1432 | } // namespace llvm |
1433 | |
1434 | #undef APFLOAT_DISPATCH_ON_SEMANTICS |
1435 | #endif // LLVM_ADT_APFLOAT_H |
1436 | |