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 | // If this is an exact power of two, return the exponent while ignoring the |
493 | // sign bit. If it's not an exact power of 2, return INT_MIN |
494 | LLVM_READONLY |
495 | int getExactLog2Abs() const; |
496 | |
497 | // If this is an exact power of two, return the exponent. If it's not an exact |
498 | // power of 2, return INT_MIN |
499 | LLVM_READONLY |
500 | int getExactLog2() const { |
501 | return isNegative() ? INT_MIN : getExactLog2Abs(); |
502 | } |
503 | |
504 | /// Returns the exponent of the internal representation of the APFloat. |
505 | /// |
506 | /// Because the radix of APFloat is 2, this is equivalent to floor(log2(x)). |
507 | /// For special APFloat values, this returns special error codes: |
508 | /// |
509 | /// NaN -> \c IEK_NaN |
510 | /// 0 -> \c IEK_Zero |
511 | /// Inf -> \c IEK_Inf |
512 | /// |
513 | friend int ilogb(const IEEEFloat &Arg); |
514 | |
515 | /// Returns: X * 2^Exp for integral exponents. |
516 | friend IEEEFloat scalbn(IEEEFloat X, int Exp, roundingMode); |
517 | |
518 | friend IEEEFloat frexp(const IEEEFloat &X, int &Exp, roundingMode); |
519 | |
520 | /// \name Special value setters. |
521 | /// @{ |
522 | |
523 | void makeLargest(bool Neg = false); |
524 | void makeSmallest(bool Neg = false); |
525 | void makeNaN(bool SNaN = false, bool Neg = false, |
526 | const APInt *fill = nullptr); |
527 | void makeInf(bool Neg = false); |
528 | void makeZero(bool Neg = false); |
529 | void makeQuiet(); |
530 | |
531 | /// Returns the smallest (by magnitude) normalized finite number in the given |
532 | /// semantics. |
533 | /// |
534 | /// \param Negative - True iff the number should be negative |
535 | void makeSmallestNormalized(bool Negative = false); |
536 | |
537 | /// @} |
538 | |
539 | cmpResult compareAbsoluteValue(const IEEEFloat &) const; |
540 | |
541 | private: |
542 | /// \name Simple Queries |
543 | /// @{ |
544 | |
545 | integerPart *significandParts(); |
546 | const integerPart *significandParts() const; |
547 | unsigned int partCount() const; |
548 | |
549 | /// @} |
550 | |
551 | /// \name Significand operations. |
552 | /// @{ |
553 | |
554 | integerPart addSignificand(const IEEEFloat &); |
555 | integerPart subtractSignificand(const IEEEFloat &, integerPart); |
556 | lostFraction addOrSubtractSignificand(const IEEEFloat &, bool subtract); |
557 | lostFraction multiplySignificand(const IEEEFloat &, IEEEFloat); |
558 | lostFraction multiplySignificand(const IEEEFloat&); |
559 | lostFraction divideSignificand(const IEEEFloat &); |
560 | void incrementSignificand(); |
561 | void initialize(const fltSemantics *); |
562 | void shiftSignificandLeft(unsigned int); |
563 | lostFraction shiftSignificandRight(unsigned int); |
564 | unsigned int significandLSB() const; |
565 | unsigned int significandMSB() const; |
566 | void zeroSignificand(); |
567 | /// Return true if the significand excluding the integral bit is all ones. |
568 | bool isSignificandAllOnes() const; |
569 | bool isSignificandAllOnesExceptLSB() const; |
570 | /// Return true if the significand excluding the integral bit is all zeros. |
571 | bool isSignificandAllZeros() const; |
572 | bool isSignificandAllZerosExceptMSB() const; |
573 | |
574 | /// @} |
575 | |
576 | /// \name Arithmetic on special values. |
577 | /// @{ |
578 | |
579 | opStatus addOrSubtractSpecials(const IEEEFloat &, bool subtract); |
580 | opStatus divideSpecials(const IEEEFloat &); |
581 | opStatus multiplySpecials(const IEEEFloat &); |
582 | opStatus modSpecials(const IEEEFloat &); |
583 | opStatus remainderSpecials(const IEEEFloat&); |
584 | |
585 | /// @} |
586 | |
587 | /// \name Miscellany |
588 | /// @{ |
589 | |
590 | bool convertFromStringSpecials(StringRef str); |
591 | opStatus normalize(roundingMode, lostFraction); |
592 | opStatus addOrSubtract(const IEEEFloat &, roundingMode, bool subtract); |
593 | opStatus handleOverflow(roundingMode); |
594 | bool roundAwayFromZero(roundingMode, lostFraction, unsigned int) const; |
595 | opStatus convertToSignExtendedInteger(MutableArrayRef<integerPart>, |
596 | unsigned int, bool, roundingMode, |
597 | bool *) const; |
598 | opStatus convertFromUnsignedParts(const integerPart *, unsigned int, |
599 | roundingMode); |
600 | Expected<opStatus> convertFromHexadecimalString(StringRef, roundingMode); |
601 | Expected<opStatus> convertFromDecimalString(StringRef, roundingMode); |
602 | char *convertNormalToHexString(char *, unsigned int, bool, |
603 | roundingMode) const; |
604 | opStatus roundSignificandWithExponent(const integerPart *, unsigned int, int, |
605 | roundingMode); |
606 | ExponentType exponentNaN() const; |
607 | ExponentType exponentInf() const; |
608 | ExponentType exponentZero() const; |
609 | |
610 | /// @} |
611 | |
612 | template <const fltSemantics &S> APInt convertIEEEFloatToAPInt() const; |
613 | APInt convertHalfAPFloatToAPInt() const; |
614 | APInt convertBFloatAPFloatToAPInt() const; |
615 | APInt convertFloatAPFloatToAPInt() const; |
616 | APInt convertDoubleAPFloatToAPInt() const; |
617 | APInt convertQuadrupleAPFloatToAPInt() const; |
618 | APInt convertF80LongDoubleAPFloatToAPInt() const; |
619 | APInt convertPPCDoubleDoubleAPFloatToAPInt() const; |
620 | APInt convertFloat8E5M2APFloatToAPInt() const; |
621 | APInt convertFloat8E5M2FNUZAPFloatToAPInt() const; |
622 | APInt convertFloat8E4M3FNAPFloatToAPInt() const; |
623 | APInt convertFloat8E4M3FNUZAPFloatToAPInt() const; |
624 | APInt convertFloat8E4M3B11FNUZAPFloatToAPInt() const; |
625 | APInt convertFloatTF32APFloatToAPInt() const; |
626 | void initFromAPInt(const fltSemantics *Sem, const APInt &api); |
627 | template <const fltSemantics &S> void initFromIEEEAPInt(const APInt &api); |
628 | void initFromHalfAPInt(const APInt &api); |
629 | void initFromBFloatAPInt(const APInt &api); |
630 | void initFromFloatAPInt(const APInt &api); |
631 | void initFromDoubleAPInt(const APInt &api); |
632 | void initFromQuadrupleAPInt(const APInt &api); |
633 | void initFromF80LongDoubleAPInt(const APInt &api); |
634 | void initFromPPCDoubleDoubleAPInt(const APInt &api); |
635 | void initFromFloat8E5M2APInt(const APInt &api); |
636 | void initFromFloat8E5M2FNUZAPInt(const APInt &api); |
637 | void initFromFloat8E4M3FNAPInt(const APInt &api); |
638 | void initFromFloat8E4M3FNUZAPInt(const APInt &api); |
639 | void initFromFloat8E4M3B11FNUZAPInt(const APInt &api); |
640 | void initFromFloatTF32APInt(const APInt &api); |
641 | |
642 | void assign(const IEEEFloat &); |
643 | void copySignificand(const IEEEFloat &); |
644 | void freeSignificand(); |
645 | |
646 | /// Note: this must be the first data member. |
647 | /// The semantics that this value obeys. |
648 | const fltSemantics *semantics; |
649 | |
650 | /// A binary fraction with an explicit integer bit. |
651 | /// |
652 | /// The significand must be at least one bit wider than the target precision. |
653 | union Significand { |
654 | integerPart part; |
655 | integerPart *parts; |
656 | } significand; |
657 | |
658 | /// The signed unbiased exponent of the value. |
659 | ExponentType exponent; |
660 | |
661 | /// What kind of floating point number this is. |
662 | /// |
663 | /// Only 2 bits are required, but VisualStudio incorrectly sign extends it. |
664 | /// Using the extra bit keeps it from failing under VisualStudio. |
665 | fltCategory category : 3; |
666 | |
667 | /// Sign bit of the number. |
668 | unsigned int sign : 1; |
669 | }; |
670 | |
671 | hash_code hash_value(const IEEEFloat &Arg); |
672 | int ilogb(const IEEEFloat &Arg); |
673 | IEEEFloat scalbn(IEEEFloat X, int Exp, IEEEFloat::roundingMode); |
674 | IEEEFloat frexp(const IEEEFloat &Val, int &Exp, IEEEFloat::roundingMode RM); |
675 | |
676 | // This mode implements more precise float in terms of two APFloats. |
677 | // The interface and layout is designed for arbitrary underlying semantics, |
678 | // though currently only PPCDoubleDouble semantics are supported, whose |
679 | // corresponding underlying semantics are IEEEdouble. |
680 | class DoubleAPFloat final : public APFloatBase { |
681 | // Note: this must be the first data member. |
682 | const fltSemantics *Semantics; |
683 | std::unique_ptr<APFloat[]> Floats; |
684 | |
685 | opStatus addImpl(const APFloat &a, const APFloat &aa, const APFloat &c, |
686 | const APFloat &cc, roundingMode RM); |
687 | |
688 | opStatus addWithSpecial(const DoubleAPFloat &LHS, const DoubleAPFloat &RHS, |
689 | DoubleAPFloat &Out, roundingMode RM); |
690 | |
691 | public: |
692 | DoubleAPFloat(const fltSemantics &S); |
693 | DoubleAPFloat(const fltSemantics &S, uninitializedTag); |
694 | DoubleAPFloat(const fltSemantics &S, integerPart); |
695 | DoubleAPFloat(const fltSemantics &S, const APInt &I); |
696 | DoubleAPFloat(const fltSemantics &S, APFloat &&First, APFloat &&Second); |
697 | DoubleAPFloat(const DoubleAPFloat &RHS); |
698 | DoubleAPFloat(DoubleAPFloat &&RHS); |
699 | |
700 | DoubleAPFloat &operator=(const DoubleAPFloat &RHS); |
701 | inline DoubleAPFloat &operator=(DoubleAPFloat &&RHS); |
702 | |
703 | bool needsCleanup() const { return Floats != nullptr; } |
704 | |
705 | inline APFloat &getFirst(); |
706 | inline const APFloat &getFirst() const; |
707 | inline APFloat &getSecond(); |
708 | inline const APFloat &getSecond() const; |
709 | |
710 | opStatus add(const DoubleAPFloat &RHS, roundingMode RM); |
711 | opStatus subtract(const DoubleAPFloat &RHS, roundingMode RM); |
712 | opStatus multiply(const DoubleAPFloat &RHS, roundingMode RM); |
713 | opStatus divide(const DoubleAPFloat &RHS, roundingMode RM); |
714 | opStatus remainder(const DoubleAPFloat &RHS); |
715 | opStatus mod(const DoubleAPFloat &RHS); |
716 | opStatus fusedMultiplyAdd(const DoubleAPFloat &Multiplicand, |
717 | const DoubleAPFloat &Addend, roundingMode RM); |
718 | opStatus roundToIntegral(roundingMode RM); |
719 | void changeSign(); |
720 | cmpResult compareAbsoluteValue(const DoubleAPFloat &RHS) const; |
721 | |
722 | fltCategory getCategory() const; |
723 | bool isNegative() const; |
724 | |
725 | void makeInf(bool Neg); |
726 | void makeZero(bool Neg); |
727 | void makeLargest(bool Neg); |
728 | void makeSmallest(bool Neg); |
729 | void makeSmallestNormalized(bool Neg); |
730 | void makeNaN(bool SNaN, bool Neg, const APInt *fill); |
731 | |
732 | cmpResult compare(const DoubleAPFloat &RHS) const; |
733 | bool bitwiseIsEqual(const DoubleAPFloat &RHS) const; |
734 | APInt bitcastToAPInt() const; |
735 | Expected<opStatus> convertFromString(StringRef, roundingMode); |
736 | opStatus next(bool nextDown); |
737 | |
738 | opStatus convertToInteger(MutableArrayRef<integerPart> Input, |
739 | unsigned int Width, bool IsSigned, roundingMode RM, |
740 | bool *IsExact) const; |
741 | opStatus convertFromAPInt(const APInt &Input, bool IsSigned, roundingMode RM); |
742 | opStatus convertFromSignExtendedInteger(const integerPart *Input, |
743 | unsigned int InputSize, bool IsSigned, |
744 | roundingMode RM); |
745 | opStatus convertFromZeroExtendedInteger(const integerPart *Input, |
746 | unsigned int InputSize, bool IsSigned, |
747 | roundingMode RM); |
748 | unsigned int convertToHexString(char *DST, unsigned int HexDigits, |
749 | bool UpperCase, roundingMode RM) const; |
750 | |
751 | bool isDenormal() const; |
752 | bool isSmallest() const; |
753 | bool isSmallestNormalized() const; |
754 | bool isLargest() const; |
755 | bool isInteger() const; |
756 | |
757 | void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision, |
758 | unsigned FormatMaxPadding, bool TruncateZero = true) const; |
759 | |
760 | bool getExactInverse(APFloat *inv) const; |
761 | |
762 | LLVM_READONLY |
763 | int getExactLog2() const; |
764 | LLVM_READONLY |
765 | int getExactLog2Abs() const; |
766 | |
767 | friend DoubleAPFloat scalbn(const DoubleAPFloat &X, int Exp, roundingMode); |
768 | friend DoubleAPFloat frexp(const DoubleAPFloat &X, int &Exp, roundingMode); |
769 | friend hash_code hash_value(const DoubleAPFloat &Arg); |
770 | }; |
771 | |
772 | hash_code hash_value(const DoubleAPFloat &Arg); |
773 | DoubleAPFloat scalbn(const DoubleAPFloat &Arg, int Exp, IEEEFloat::roundingMode RM); |
774 | DoubleAPFloat frexp(const DoubleAPFloat &X, int &Exp, IEEEFloat::roundingMode); |
775 | |
776 | } // End detail namespace |
777 | |
778 | // This is a interface class that is currently forwarding functionalities from |
779 | // detail::IEEEFloat. |
780 | class APFloat : public APFloatBase { |
781 | typedef detail::IEEEFloat IEEEFloat; |
782 | typedef detail::DoubleAPFloat DoubleAPFloat; |
783 | |
784 | static_assert(std::is_standard_layout<IEEEFloat>::value); |
785 | |
786 | union Storage { |
787 | const fltSemantics *semantics; |
788 | IEEEFloat IEEE; |
789 | DoubleAPFloat Double; |
790 | |
791 | explicit Storage(IEEEFloat F, const fltSemantics &S); |
792 | explicit Storage(DoubleAPFloat F, const fltSemantics &S) |
793 | : Double(std::move(F)) { |
794 | assert(&S == &PPCDoubleDouble()); |
795 | } |
796 | |
797 | template <typename... ArgTypes> |
798 | Storage(const fltSemantics &Semantics, ArgTypes &&... Args) { |
799 | if (usesLayout<IEEEFloat>(Semantics)) { |
800 | new (&IEEE) IEEEFloat(Semantics, std::forward<ArgTypes>(Args)...); |
801 | return; |
802 | } |
803 | if (usesLayout<DoubleAPFloat>(Semantics)) { |
804 | new (&Double) DoubleAPFloat(Semantics, std::forward<ArgTypes>(Args)...); |
805 | return; |
806 | } |
807 | llvm_unreachable("Unexpected semantics" ); |
808 | } |
809 | |
810 | ~Storage() { |
811 | if (usesLayout<IEEEFloat>(Semantics: *semantics)) { |
812 | IEEE.~IEEEFloat(); |
813 | return; |
814 | } |
815 | if (usesLayout<DoubleAPFloat>(Semantics: *semantics)) { |
816 | Double.~DoubleAPFloat(); |
817 | return; |
818 | } |
819 | llvm_unreachable("Unexpected semantics" ); |
820 | } |
821 | |
822 | Storage(const Storage &RHS) { |
823 | if (usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
824 | new (this) IEEEFloat(RHS.IEEE); |
825 | return; |
826 | } |
827 | if (usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
828 | new (this) DoubleAPFloat(RHS.Double); |
829 | return; |
830 | } |
831 | llvm_unreachable("Unexpected semantics" ); |
832 | } |
833 | |
834 | Storage(Storage &&RHS) { |
835 | if (usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
836 | new (this) IEEEFloat(std::move(RHS.IEEE)); |
837 | return; |
838 | } |
839 | if (usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
840 | new (this) DoubleAPFloat(std::move(RHS.Double)); |
841 | return; |
842 | } |
843 | llvm_unreachable("Unexpected semantics" ); |
844 | } |
845 | |
846 | Storage &operator=(const Storage &RHS) { |
847 | if (usesLayout<IEEEFloat>(Semantics: *semantics) && |
848 | usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
849 | IEEE = RHS.IEEE; |
850 | } else if (usesLayout<DoubleAPFloat>(Semantics: *semantics) && |
851 | usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
852 | Double = RHS.Double; |
853 | } else if (this != &RHS) { |
854 | this->~Storage(); |
855 | new (this) Storage(RHS); |
856 | } |
857 | return *this; |
858 | } |
859 | |
860 | Storage &operator=(Storage &&RHS) { |
861 | if (usesLayout<IEEEFloat>(Semantics: *semantics) && |
862 | usesLayout<IEEEFloat>(Semantics: *RHS.semantics)) { |
863 | IEEE = std::move(RHS.IEEE); |
864 | } else if (usesLayout<DoubleAPFloat>(Semantics: *semantics) && |
865 | usesLayout<DoubleAPFloat>(Semantics: *RHS.semantics)) { |
866 | Double = std::move(RHS.Double); |
867 | } else if (this != &RHS) { |
868 | this->~Storage(); |
869 | new (this) Storage(std::move(RHS)); |
870 | } |
871 | return *this; |
872 | } |
873 | } U; |
874 | |
875 | template <typename T> static bool usesLayout(const fltSemantics &Semantics) { |
876 | static_assert(std::is_same<T, IEEEFloat>::value || |
877 | std::is_same<T, DoubleAPFloat>::value); |
878 | if (std::is_same<T, DoubleAPFloat>::value) { |
879 | return &Semantics == &PPCDoubleDouble(); |
880 | } |
881 | return &Semantics != &PPCDoubleDouble(); |
882 | } |
883 | |
884 | IEEEFloat &getIEEE() { |
885 | if (usesLayout<IEEEFloat>(Semantics: *U.semantics)) |
886 | return U.IEEE; |
887 | if (usesLayout<DoubleAPFloat>(Semantics: *U.semantics)) |
888 | return U.Double.getFirst().U.IEEE; |
889 | llvm_unreachable("Unexpected semantics" ); |
890 | } |
891 | |
892 | const IEEEFloat &getIEEE() const { |
893 | if (usesLayout<IEEEFloat>(Semantics: *U.semantics)) |
894 | return U.IEEE; |
895 | if (usesLayout<DoubleAPFloat>(Semantics: *U.semantics)) |
896 | return U.Double.getFirst().U.IEEE; |
897 | llvm_unreachable("Unexpected semantics" ); |
898 | } |
899 | |
900 | void makeZero(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeZero(Neg)); } |
901 | |
902 | void makeInf(bool Neg) { APFLOAT_DISPATCH_ON_SEMANTICS(makeInf(Neg)); } |
903 | |
904 | void makeNaN(bool SNaN, bool Neg, const APInt *fill) { |
905 | APFLOAT_DISPATCH_ON_SEMANTICS(makeNaN(SNaN, Neg, fill)); |
906 | } |
907 | |
908 | void makeLargest(bool Neg) { |
909 | APFLOAT_DISPATCH_ON_SEMANTICS(makeLargest(Neg)); |
910 | } |
911 | |
912 | void makeSmallest(bool Neg) { |
913 | APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallest(Neg)); |
914 | } |
915 | |
916 | void makeSmallestNormalized(bool Neg) { |
917 | APFLOAT_DISPATCH_ON_SEMANTICS(makeSmallestNormalized(Neg)); |
918 | } |
919 | |
920 | explicit APFloat(IEEEFloat F, const fltSemantics &S) : U(std::move(F), S) {} |
921 | explicit APFloat(DoubleAPFloat F, const fltSemantics &S) |
922 | : U(std::move(F), S) {} |
923 | |
924 | cmpResult compareAbsoluteValue(const APFloat &RHS) const { |
925 | assert(&getSemantics() == &RHS.getSemantics() && |
926 | "Should only compare APFloats with the same semantics" ); |
927 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
928 | return U.IEEE.compareAbsoluteValue(RHS.U.IEEE); |
929 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
930 | return U.Double.compareAbsoluteValue(RHS: RHS.U.Double); |
931 | llvm_unreachable("Unexpected semantics" ); |
932 | } |
933 | |
934 | public: |
935 | APFloat(const fltSemantics &Semantics) : U(Semantics) {} |
936 | APFloat(const fltSemantics &Semantics, StringRef S); |
937 | APFloat(const fltSemantics &Semantics, integerPart I) : U(Semantics, I) {} |
938 | template <typename T, |
939 | typename = std::enable_if_t<std::is_floating_point<T>::value>> |
940 | APFloat(const fltSemantics &Semantics, T V) = delete; |
941 | // TODO: Remove this constructor. This isn't faster than the first one. |
942 | APFloat(const fltSemantics &Semantics, uninitializedTag) |
943 | : U(Semantics, uninitialized) {} |
944 | APFloat(const fltSemantics &Semantics, const APInt &I) : U(Semantics, I) {} |
945 | explicit APFloat(double d) : U(IEEEFloat(d), IEEEdouble()) {} |
946 | explicit APFloat(float f) : U(IEEEFloat(f), IEEEsingle()) {} |
947 | APFloat(const APFloat &RHS) = default; |
948 | APFloat(APFloat &&RHS) = default; |
949 | |
950 | ~APFloat() = default; |
951 | |
952 | bool needsCleanup() const { APFLOAT_DISPATCH_ON_SEMANTICS(needsCleanup()); } |
953 | |
954 | /// Factory for Positive and Negative Zero. |
955 | /// |
956 | /// \param Negative True iff the number should be negative. |
957 | static APFloat getZero(const fltSemantics &Sem, bool Negative = false) { |
958 | APFloat Val(Sem, uninitialized); |
959 | Val.makeZero(Neg: Negative); |
960 | return Val; |
961 | } |
962 | |
963 | /// Factory for Positive and Negative Infinity. |
964 | /// |
965 | /// \param Negative True iff the number should be negative. |
966 | static APFloat getInf(const fltSemantics &Sem, bool Negative = false) { |
967 | APFloat Val(Sem, uninitialized); |
968 | Val.makeInf(Neg: Negative); |
969 | return Val; |
970 | } |
971 | |
972 | /// Factory for NaN values. |
973 | /// |
974 | /// \param Negative - True iff the NaN generated should be negative. |
975 | /// \param payload - The unspecified fill bits for creating the NaN, 0 by |
976 | /// default. The value is truncated as necessary. |
977 | static APFloat getNaN(const fltSemantics &Sem, bool Negative = false, |
978 | uint64_t payload = 0) { |
979 | if (payload) { |
980 | APInt intPayload(64, payload); |
981 | return getQNaN(Sem, Negative, payload: &intPayload); |
982 | } else { |
983 | return getQNaN(Sem, Negative, payload: nullptr); |
984 | } |
985 | } |
986 | |
987 | /// Factory for QNaN values. |
988 | static APFloat getQNaN(const fltSemantics &Sem, bool Negative = false, |
989 | const APInt *payload = nullptr) { |
990 | APFloat Val(Sem, uninitialized); |
991 | Val.makeNaN(SNaN: false, Neg: Negative, fill: payload); |
992 | return Val; |
993 | } |
994 | |
995 | /// Factory for SNaN values. |
996 | static APFloat getSNaN(const fltSemantics &Sem, bool Negative = false, |
997 | const APInt *payload = nullptr) { |
998 | APFloat Val(Sem, uninitialized); |
999 | Val.makeNaN(SNaN: true, Neg: Negative, fill: payload); |
1000 | return Val; |
1001 | } |
1002 | |
1003 | /// Returns the largest finite number in the given semantics. |
1004 | /// |
1005 | /// \param Negative - True iff the number should be negative |
1006 | static APFloat getLargest(const fltSemantics &Sem, bool Negative = false) { |
1007 | APFloat Val(Sem, uninitialized); |
1008 | Val.makeLargest(Neg: Negative); |
1009 | return Val; |
1010 | } |
1011 | |
1012 | /// Returns the smallest (by magnitude) finite number in the given semantics. |
1013 | /// Might be denormalized, which implies a relative loss of precision. |
1014 | /// |
1015 | /// \param Negative - True iff the number should be negative |
1016 | static APFloat getSmallest(const fltSemantics &Sem, bool Negative = false) { |
1017 | APFloat Val(Sem, uninitialized); |
1018 | Val.makeSmallest(Neg: Negative); |
1019 | return Val; |
1020 | } |
1021 | |
1022 | /// Returns the smallest (by magnitude) normalized finite number in the given |
1023 | /// semantics. |
1024 | /// |
1025 | /// \param Negative - True iff the number should be negative |
1026 | static APFloat getSmallestNormalized(const fltSemantics &Sem, |
1027 | bool Negative = false) { |
1028 | APFloat Val(Sem, uninitialized); |
1029 | Val.makeSmallestNormalized(Neg: Negative); |
1030 | return Val; |
1031 | } |
1032 | |
1033 | /// Returns a float which is bitcasted from an all one value int. |
1034 | /// |
1035 | /// \param Semantics - type float semantics |
1036 | static APFloat getAllOnesValue(const fltSemantics &Semantics); |
1037 | |
1038 | /// Used to insert APFloat objects, or objects that contain APFloat objects, |
1039 | /// into FoldingSets. |
1040 | void Profile(FoldingSetNodeID &NID) const; |
1041 | |
1042 | opStatus add(const APFloat &RHS, roundingMode RM) { |
1043 | assert(&getSemantics() == &RHS.getSemantics() && |
1044 | "Should only call on two APFloats with the same semantics" ); |
1045 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1046 | return U.IEEE.add(RHS.U.IEEE, RM); |
1047 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1048 | return U.Double.add(RHS: RHS.U.Double, RM); |
1049 | llvm_unreachable("Unexpected semantics" ); |
1050 | } |
1051 | opStatus subtract(const APFloat &RHS, roundingMode RM) { |
1052 | assert(&getSemantics() == &RHS.getSemantics() && |
1053 | "Should only call on two APFloats with the same semantics" ); |
1054 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1055 | return U.IEEE.subtract(RHS.U.IEEE, RM); |
1056 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1057 | return U.Double.subtract(RHS: RHS.U.Double, RM); |
1058 | llvm_unreachable("Unexpected semantics" ); |
1059 | } |
1060 | opStatus multiply(const APFloat &RHS, roundingMode RM) { |
1061 | assert(&getSemantics() == &RHS.getSemantics() && |
1062 | "Should only call on two APFloats with the same semantics" ); |
1063 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1064 | return U.IEEE.multiply(RHS.U.IEEE, RM); |
1065 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1066 | return U.Double.multiply(RHS: RHS.U.Double, RM); |
1067 | llvm_unreachable("Unexpected semantics" ); |
1068 | } |
1069 | opStatus divide(const APFloat &RHS, roundingMode RM) { |
1070 | assert(&getSemantics() == &RHS.getSemantics() && |
1071 | "Should only call on two APFloats with the same semantics" ); |
1072 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1073 | return U.IEEE.divide(RHS.U.IEEE, RM); |
1074 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1075 | return U.Double.divide(RHS: RHS.U.Double, RM); |
1076 | llvm_unreachable("Unexpected semantics" ); |
1077 | } |
1078 | opStatus remainder(const APFloat &RHS) { |
1079 | assert(&getSemantics() == &RHS.getSemantics() && |
1080 | "Should only call on two APFloats with the same semantics" ); |
1081 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1082 | return U.IEEE.remainder(RHS.U.IEEE); |
1083 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1084 | return U.Double.remainder(RHS: RHS.U.Double); |
1085 | llvm_unreachable("Unexpected semantics" ); |
1086 | } |
1087 | opStatus mod(const APFloat &RHS) { |
1088 | assert(&getSemantics() == &RHS.getSemantics() && |
1089 | "Should only call on two APFloats with the same semantics" ); |
1090 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1091 | return U.IEEE.mod(RHS.U.IEEE); |
1092 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1093 | return U.Double.mod(RHS: RHS.U.Double); |
1094 | llvm_unreachable("Unexpected semantics" ); |
1095 | } |
1096 | opStatus fusedMultiplyAdd(const APFloat &Multiplicand, const APFloat &Addend, |
1097 | roundingMode RM) { |
1098 | assert(&getSemantics() == &Multiplicand.getSemantics() && |
1099 | "Should only call on APFloats with the same semantics" ); |
1100 | assert(&getSemantics() == &Addend.getSemantics() && |
1101 | "Should only call on APFloats with the same semantics" ); |
1102 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1103 | return U.IEEE.fusedMultiplyAdd(Multiplicand.U.IEEE, Addend.U.IEEE, RM); |
1104 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1105 | return U.Double.fusedMultiplyAdd(Multiplicand: Multiplicand.U.Double, Addend: Addend.U.Double, |
1106 | RM); |
1107 | llvm_unreachable("Unexpected semantics" ); |
1108 | } |
1109 | opStatus roundToIntegral(roundingMode RM) { |
1110 | APFLOAT_DISPATCH_ON_SEMANTICS(roundToIntegral(RM)); |
1111 | } |
1112 | |
1113 | // TODO: bool parameters are not readable and a source of bugs. |
1114 | // Do something. |
1115 | opStatus next(bool nextDown) { |
1116 | APFLOAT_DISPATCH_ON_SEMANTICS(next(nextDown)); |
1117 | } |
1118 | |
1119 | /// Negate an APFloat. |
1120 | APFloat operator-() const { |
1121 | APFloat Result(*this); |
1122 | Result.changeSign(); |
1123 | return Result; |
1124 | } |
1125 | |
1126 | /// Add two APFloats, rounding ties to the nearest even. |
1127 | /// No error checking. |
1128 | APFloat operator+(const APFloat &RHS) const { |
1129 | APFloat Result(*this); |
1130 | (void)Result.add(RHS, RM: rmNearestTiesToEven); |
1131 | return Result; |
1132 | } |
1133 | |
1134 | /// Subtract two APFloats, rounding ties to the nearest even. |
1135 | /// No error checking. |
1136 | APFloat operator-(const APFloat &RHS) const { |
1137 | APFloat Result(*this); |
1138 | (void)Result.subtract(RHS, RM: rmNearestTiesToEven); |
1139 | return Result; |
1140 | } |
1141 | |
1142 | /// Multiply two APFloats, rounding ties to the nearest even. |
1143 | /// No error checking. |
1144 | APFloat operator*(const APFloat &RHS) const { |
1145 | APFloat Result(*this); |
1146 | (void)Result.multiply(RHS, RM: rmNearestTiesToEven); |
1147 | return Result; |
1148 | } |
1149 | |
1150 | /// Divide the first APFloat by the second, rounding ties to the nearest even. |
1151 | /// No error checking. |
1152 | APFloat operator/(const APFloat &RHS) const { |
1153 | APFloat Result(*this); |
1154 | (void)Result.divide(RHS, RM: rmNearestTiesToEven); |
1155 | return Result; |
1156 | } |
1157 | |
1158 | void changeSign() { APFLOAT_DISPATCH_ON_SEMANTICS(changeSign()); } |
1159 | void clearSign() { |
1160 | if (isNegative()) |
1161 | changeSign(); |
1162 | } |
1163 | void copySign(const APFloat &RHS) { |
1164 | if (isNegative() != RHS.isNegative()) |
1165 | changeSign(); |
1166 | } |
1167 | |
1168 | /// A static helper to produce a copy of an APFloat value with its sign |
1169 | /// copied from some other APFloat. |
1170 | static APFloat copySign(APFloat Value, const APFloat &Sign) { |
1171 | Value.copySign(RHS: Sign); |
1172 | return Value; |
1173 | } |
1174 | |
1175 | /// Assuming this is an IEEE-754 NaN value, quiet its signaling bit. |
1176 | /// This preserves the sign and payload bits. |
1177 | APFloat makeQuiet() const { |
1178 | APFloat Result(*this); |
1179 | Result.getIEEE().makeQuiet(); |
1180 | return Result; |
1181 | } |
1182 | |
1183 | opStatus convert(const fltSemantics &ToSemantics, roundingMode RM, |
1184 | bool *losesInfo); |
1185 | opStatus convertToInteger(MutableArrayRef<integerPart> Input, |
1186 | unsigned int Width, bool IsSigned, roundingMode RM, |
1187 | bool *IsExact) const { |
1188 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1189 | convertToInteger(Input, Width, IsSigned, RM, IsExact)); |
1190 | } |
1191 | opStatus convertToInteger(APSInt &Result, roundingMode RM, |
1192 | bool *IsExact) const; |
1193 | opStatus convertFromAPInt(const APInt &Input, bool IsSigned, |
1194 | roundingMode RM) { |
1195 | APFLOAT_DISPATCH_ON_SEMANTICS(convertFromAPInt(Input, IsSigned, RM)); |
1196 | } |
1197 | opStatus convertFromSignExtendedInteger(const integerPart *Input, |
1198 | unsigned int InputSize, bool IsSigned, |
1199 | roundingMode RM) { |
1200 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1201 | convertFromSignExtendedInteger(Input, InputSize, IsSigned, RM)); |
1202 | } |
1203 | opStatus convertFromZeroExtendedInteger(const integerPart *Input, |
1204 | unsigned int InputSize, bool IsSigned, |
1205 | roundingMode RM) { |
1206 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1207 | convertFromZeroExtendedInteger(Input, InputSize, IsSigned, RM)); |
1208 | } |
1209 | Expected<opStatus> convertFromString(StringRef, roundingMode); |
1210 | APInt bitcastToAPInt() const { |
1211 | APFLOAT_DISPATCH_ON_SEMANTICS(bitcastToAPInt()); |
1212 | } |
1213 | |
1214 | /// Converts this APFloat to host double value. |
1215 | /// |
1216 | /// \pre The APFloat must be built using semantics, that can be represented by |
1217 | /// the host double type without loss of precision. It can be IEEEdouble and |
1218 | /// shorter semantics, like IEEEsingle and others. |
1219 | double convertToDouble() const; |
1220 | |
1221 | /// Converts this APFloat to host float value. |
1222 | /// |
1223 | /// \pre The APFloat must be built using semantics, that can be represented by |
1224 | /// the host float type without loss of precision. It can be IEEEsingle and |
1225 | /// shorter semantics, like IEEEhalf. |
1226 | float convertToFloat() const; |
1227 | |
1228 | bool operator==(const APFloat &RHS) const { return compare(RHS) == cmpEqual; } |
1229 | |
1230 | bool operator!=(const APFloat &RHS) const { return compare(RHS) != cmpEqual; } |
1231 | |
1232 | bool operator<(const APFloat &RHS) const { |
1233 | return compare(RHS) == cmpLessThan; |
1234 | } |
1235 | |
1236 | bool operator>(const APFloat &RHS) const { |
1237 | return compare(RHS) == cmpGreaterThan; |
1238 | } |
1239 | |
1240 | bool operator<=(const APFloat &RHS) const { |
1241 | cmpResult Res = compare(RHS); |
1242 | return Res == cmpLessThan || Res == cmpEqual; |
1243 | } |
1244 | |
1245 | bool operator>=(const APFloat &RHS) const { |
1246 | cmpResult Res = compare(RHS); |
1247 | return Res == cmpGreaterThan || Res == cmpEqual; |
1248 | } |
1249 | |
1250 | cmpResult compare(const APFloat &RHS) const { |
1251 | assert(&getSemantics() == &RHS.getSemantics() && |
1252 | "Should only compare APFloats with the same semantics" ); |
1253 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1254 | return U.IEEE.compare(RHS.U.IEEE); |
1255 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1256 | return U.Double.compare(RHS: RHS.U.Double); |
1257 | llvm_unreachable("Unexpected semantics" ); |
1258 | } |
1259 | |
1260 | bool bitwiseIsEqual(const APFloat &RHS) const { |
1261 | if (&getSemantics() != &RHS.getSemantics()) |
1262 | return false; |
1263 | if (usesLayout<IEEEFloat>(Semantics: getSemantics())) |
1264 | return U.IEEE.bitwiseIsEqual(RHS.U.IEEE); |
1265 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics())) |
1266 | return U.Double.bitwiseIsEqual(RHS: RHS.U.Double); |
1267 | llvm_unreachable("Unexpected semantics" ); |
1268 | } |
1269 | |
1270 | /// We don't rely on operator== working on double values, as |
1271 | /// it returns true for things that are clearly not equal, like -0.0 and 0.0. |
1272 | /// As such, this method can be used to do an exact bit-for-bit comparison of |
1273 | /// two floating point values. |
1274 | /// |
1275 | /// We leave the version with the double argument here because it's just so |
1276 | /// convenient to write "2.0" and the like. Without this function we'd |
1277 | /// have to duplicate its logic everywhere it's called. |
1278 | bool isExactlyValue(double V) const { |
1279 | bool ignored; |
1280 | APFloat Tmp(V); |
1281 | Tmp.convert(ToSemantics: getSemantics(), RM: APFloat::rmNearestTiesToEven, losesInfo: &ignored); |
1282 | return bitwiseIsEqual(RHS: Tmp); |
1283 | } |
1284 | |
1285 | unsigned int convertToHexString(char *DST, unsigned int HexDigits, |
1286 | bool UpperCase, roundingMode RM) const { |
1287 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1288 | convertToHexString(DST, HexDigits, UpperCase, RM)); |
1289 | } |
1290 | |
1291 | bool isZero() const { return getCategory() == fcZero; } |
1292 | bool isInfinity() const { return getCategory() == fcInfinity; } |
1293 | bool isNaN() const { return getCategory() == fcNaN; } |
1294 | |
1295 | bool isNegative() const { return getIEEE().isNegative(); } |
1296 | bool isDenormal() const { APFLOAT_DISPATCH_ON_SEMANTICS(isDenormal()); } |
1297 | bool isSignaling() const { return getIEEE().isSignaling(); } |
1298 | |
1299 | bool isNormal() const { return !isDenormal() && isFiniteNonZero(); } |
1300 | bool isFinite() const { return !isNaN() && !isInfinity(); } |
1301 | |
1302 | fltCategory getCategory() const { return getIEEE().getCategory(); } |
1303 | const fltSemantics &getSemantics() const { return *U.semantics; } |
1304 | bool isNonZero() const { return !isZero(); } |
1305 | bool isFiniteNonZero() const { return isFinite() && !isZero(); } |
1306 | bool isPosZero() const { return isZero() && !isNegative(); } |
1307 | bool isNegZero() const { return isZero() && isNegative(); } |
1308 | bool isPosInfinity() const { return isInfinity() && !isNegative(); } |
1309 | bool isNegInfinity() const { return isInfinity() && isNegative(); } |
1310 | bool isSmallest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isSmallest()); } |
1311 | bool isLargest() const { APFLOAT_DISPATCH_ON_SEMANTICS(isLargest()); } |
1312 | bool isInteger() const { APFLOAT_DISPATCH_ON_SEMANTICS(isInteger()); } |
1313 | bool isIEEE() const { return usesLayout<IEEEFloat>(Semantics: getSemantics()); } |
1314 | |
1315 | bool isSmallestNormalized() const { |
1316 | APFLOAT_DISPATCH_ON_SEMANTICS(isSmallestNormalized()); |
1317 | } |
1318 | |
1319 | /// Return the FPClassTest which will return true for the value. |
1320 | FPClassTest classify() const; |
1321 | |
1322 | APFloat &operator=(const APFloat &RHS) = default; |
1323 | APFloat &operator=(APFloat &&RHS) = default; |
1324 | |
1325 | void toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision = 0, |
1326 | unsigned FormatMaxPadding = 3, bool TruncateZero = true) const { |
1327 | APFLOAT_DISPATCH_ON_SEMANTICS( |
1328 | toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero)); |
1329 | } |
1330 | |
1331 | void print(raw_ostream &) const; |
1332 | void dump() const; |
1333 | |
1334 | bool getExactInverse(APFloat *inv) const { |
1335 | APFLOAT_DISPATCH_ON_SEMANTICS(getExactInverse(inv)); |
1336 | } |
1337 | |
1338 | LLVM_READONLY |
1339 | int getExactLog2Abs() const { |
1340 | APFLOAT_DISPATCH_ON_SEMANTICS(getExactLog2Abs()); |
1341 | } |
1342 | |
1343 | LLVM_READONLY |
1344 | int getExactLog2() const { |
1345 | APFLOAT_DISPATCH_ON_SEMANTICS(getExactLog2()); |
1346 | } |
1347 | |
1348 | friend hash_code hash_value(const APFloat &Arg); |
1349 | friend int ilogb(const APFloat &Arg) { return ilogb(Arg: Arg.getIEEE()); } |
1350 | friend APFloat scalbn(APFloat X, int Exp, roundingMode RM); |
1351 | friend APFloat frexp(const APFloat &X, int &Exp, roundingMode RM); |
1352 | friend IEEEFloat; |
1353 | friend DoubleAPFloat; |
1354 | }; |
1355 | |
1356 | /// See friend declarations above. |
1357 | /// |
1358 | /// These additional declarations are required in order to compile LLVM with IBM |
1359 | /// xlC compiler. |
1360 | hash_code hash_value(const APFloat &Arg); |
1361 | inline APFloat scalbn(APFloat X, int Exp, APFloat::roundingMode RM) { |
1362 | if (APFloat::usesLayout<detail::IEEEFloat>(Semantics: X.getSemantics())) |
1363 | return APFloat(scalbn(X: X.U.IEEE, Exp, RM), X.getSemantics()); |
1364 | if (APFloat::usesLayout<detail::DoubleAPFloat>(Semantics: X.getSemantics())) |
1365 | return APFloat(scalbn(Arg: X.U.Double, Exp, RM), X.getSemantics()); |
1366 | llvm_unreachable("Unexpected semantics" ); |
1367 | } |
1368 | |
1369 | /// Equivalent of C standard library function. |
1370 | /// |
1371 | /// While the C standard says Exp is an unspecified value for infinity and nan, |
1372 | /// this returns INT_MAX for infinities, and INT_MIN for NaNs. |
1373 | inline APFloat frexp(const APFloat &X, int &Exp, APFloat::roundingMode RM) { |
1374 | if (APFloat::usesLayout<detail::IEEEFloat>(Semantics: X.getSemantics())) |
1375 | return APFloat(frexp(Val: X.U.IEEE, Exp, RM), X.getSemantics()); |
1376 | if (APFloat::usesLayout<detail::DoubleAPFloat>(Semantics: X.getSemantics())) |
1377 | return APFloat(frexp(X: X.U.Double, Exp, RM), X.getSemantics()); |
1378 | llvm_unreachable("Unexpected semantics" ); |
1379 | } |
1380 | /// Returns the absolute value of the argument. |
1381 | inline APFloat abs(APFloat X) { |
1382 | X.clearSign(); |
1383 | return X; |
1384 | } |
1385 | |
1386 | /// Returns the negated value of the argument. |
1387 | inline APFloat neg(APFloat X) { |
1388 | X.changeSign(); |
1389 | return X; |
1390 | } |
1391 | |
1392 | /// Implements IEEE-754 2019 minimumNumber semantics. Returns the smaller of the |
1393 | /// 2 arguments if both are not NaN. If either argument is a NaN, returns the |
1394 | /// other argument. -0 is treated as ordered less than +0. |
1395 | LLVM_READONLY |
1396 | inline APFloat minnum(const APFloat &A, const APFloat &B) { |
1397 | if (A.isNaN()) |
1398 | return B; |
1399 | if (B.isNaN()) |
1400 | return A; |
1401 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1402 | return A.isNegative() ? A : B; |
1403 | return B < A ? B : A; |
1404 | } |
1405 | |
1406 | /// Implements IEEE-754 2019 maximumNumber semantics. Returns the larger of the |
1407 | /// 2 arguments if both are not NaN. If either argument is a NaN, returns the |
1408 | /// other argument. +0 is treated as ordered greater than -0. |
1409 | LLVM_READONLY |
1410 | inline APFloat maxnum(const APFloat &A, const APFloat &B) { |
1411 | if (A.isNaN()) |
1412 | return B; |
1413 | if (B.isNaN()) |
1414 | return A; |
1415 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1416 | return A.isNegative() ? B : A; |
1417 | return A < B ? B : A; |
1418 | } |
1419 | |
1420 | /// Implements IEEE 754-2019 minimum semantics. Returns the smaller of 2 |
1421 | /// arguments, propagating NaNs and treating -0 as less than +0. |
1422 | LLVM_READONLY |
1423 | inline APFloat minimum(const APFloat &A, const APFloat &B) { |
1424 | if (A.isNaN()) |
1425 | return A; |
1426 | if (B.isNaN()) |
1427 | return B; |
1428 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1429 | return A.isNegative() ? A : B; |
1430 | return B < A ? B : A; |
1431 | } |
1432 | |
1433 | /// Implements IEEE 754-2019 maximum semantics. Returns the larger of 2 |
1434 | /// arguments, propagating NaNs and treating -0 as less than +0. |
1435 | LLVM_READONLY |
1436 | inline APFloat maximum(const APFloat &A, const APFloat &B) { |
1437 | if (A.isNaN()) |
1438 | return A; |
1439 | if (B.isNaN()) |
1440 | return B; |
1441 | if (A.isZero() && B.isZero() && (A.isNegative() != B.isNegative())) |
1442 | return A.isNegative() ? B : A; |
1443 | return A < B ? B : A; |
1444 | } |
1445 | |
1446 | // We want the following functions to be available in the header for inlining. |
1447 | // We cannot define them inline in the class definition of `DoubleAPFloat` |
1448 | // because doing so would instantiate `std::unique_ptr<APFloat[]>` before |
1449 | // `APFloat` is defined, and that would be undefined behavior. |
1450 | namespace detail { |
1451 | |
1452 | DoubleAPFloat &DoubleAPFloat::operator=(DoubleAPFloat &&RHS) { |
1453 | if (this != &RHS) { |
1454 | this->~DoubleAPFloat(); |
1455 | new (this) DoubleAPFloat(std::move(RHS)); |
1456 | } |
1457 | return *this; |
1458 | } |
1459 | |
1460 | APFloat &DoubleAPFloat::getFirst() { return Floats[0]; } |
1461 | const APFloat &DoubleAPFloat::getFirst() const { return Floats[0]; } |
1462 | APFloat &DoubleAPFloat::getSecond() { return Floats[1]; } |
1463 | const APFloat &DoubleAPFloat::getSecond() const { return Floats[1]; } |
1464 | |
1465 | } // namespace detail |
1466 | |
1467 | } // namespace llvm |
1468 | |
1469 | #undef APFLOAT_DISPATCH_ON_SEMANTICS |
1470 | #endif // LLVM_ADT_APFLOAT_H |
1471 | |