1 | //===-- APFloat.cpp - Implement APFloat class -----------------------------===// |
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 | // This file implements a class to represent arbitrary precision floating |
10 | // point values and provide a variety of arithmetic operations on them. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/ADT/APFloat.h" |
15 | #include "llvm/ADT/APSInt.h" |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/FloatingPointMode.h" |
18 | #include "llvm/ADT/FoldingSet.h" |
19 | #include "llvm/ADT/Hashing.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/ADT/StringExtras.h" |
22 | #include "llvm/ADT/StringRef.h" |
23 | #include "llvm/Config/llvm-config.h" |
24 | #include "llvm/Support/Debug.h" |
25 | #include "llvm/Support/Error.h" |
26 | #include "llvm/Support/MathExtras.h" |
27 | #include "llvm/Support/raw_ostream.h" |
28 | #include <cstring> |
29 | #include <limits.h> |
30 | |
31 | #define APFLOAT_DISPATCH_ON_SEMANTICS(METHOD_CALL) \ |
32 | do { \ |
33 | if (usesLayout<IEEEFloat>(getSemantics())) \ |
34 | return U.IEEE.METHOD_CALL; \ |
35 | if (usesLayout<DoubleAPFloat>(getSemantics())) \ |
36 | return U.Double.METHOD_CALL; \ |
37 | llvm_unreachable("Unexpected semantics"); \ |
38 | } while (false) |
39 | |
40 | using namespace llvm; |
41 | |
42 | /// A macro used to combine two fcCategory enums into one key which can be used |
43 | /// in a switch statement to classify how the interaction of two APFloat's |
44 | /// categories affects an operation. |
45 | /// |
46 | /// TODO: If clang source code is ever allowed to use constexpr in its own |
47 | /// codebase, change this into a static inline function. |
48 | #define PackCategoriesIntoKey(_lhs, _rhs) ((_lhs) * 4 + (_rhs)) |
49 | |
50 | /* Assumed in hexadecimal significand parsing, and conversion to |
51 | hexadecimal strings. */ |
52 | static_assert(APFloatBase::integerPartWidth % 4 == 0, "Part width must be divisible by 4!" ); |
53 | |
54 | namespace llvm { |
55 | |
56 | // How the nonfinite values Inf and NaN are represented. |
57 | enum class fltNonfiniteBehavior { |
58 | // Represents standard IEEE 754 behavior. A value is nonfinite if the |
59 | // exponent field is all 1s. In such cases, a value is Inf if the |
60 | // significand bits are all zero, and NaN otherwise |
61 | IEEE754, |
62 | |
63 | // This behavior is present in the Float8ExMyFN* types (Float8E4M3FN, |
64 | // Float8E5M2FNUZ, Float8E4M3FNUZ, and Float8E4M3B11FNUZ). There is no |
65 | // representation for Inf, and operations that would ordinarily produce Inf |
66 | // produce NaN instead. |
67 | // The details of the NaN representation(s) in this form are determined by the |
68 | // `fltNanEncoding` enum. We treat all NaNs as quiet, as the available |
69 | // encodings do not distinguish between signalling and quiet NaN. |
70 | NanOnly, |
71 | |
72 | // This behavior is present in Float6E3M2FN, Float6E2M3FN, and |
73 | // Float4E2M1FN types, which do not support Inf or NaN values. |
74 | FiniteOnly, |
75 | }; |
76 | |
77 | // How NaN values are represented. This is curently only used in combination |
78 | // with fltNonfiniteBehavior::NanOnly, and using a variant other than IEEE |
79 | // while having IEEE non-finite behavior is liable to lead to unexpected |
80 | // results. |
81 | enum class fltNanEncoding { |
82 | // Represents the standard IEEE behavior where a value is NaN if its |
83 | // exponent is all 1s and the significand is non-zero. |
84 | IEEE, |
85 | |
86 | // Represents the behavior in the Float8E4M3FN floating point type where NaN |
87 | // is represented by having the exponent and mantissa set to all 1s. |
88 | // This behavior matches the FP8 E4M3 type described in |
89 | // https://arxiv.org/abs/2209.05433. We treat both signed and unsigned NaNs |
90 | // as non-signalling, although the paper does not state whether the NaN |
91 | // values are signalling or not. |
92 | AllOnes, |
93 | |
94 | // Represents the behavior in Float8E{5,4}E{2,3}FNUZ floating point types |
95 | // where NaN is represented by a sign bit of 1 and all 0s in the exponent |
96 | // and mantissa (i.e. the negative zero encoding in a IEEE float). Since |
97 | // there is only one NaN value, it is treated as quiet NaN. This matches the |
98 | // behavior described in https://arxiv.org/abs/2206.02915 . |
99 | NegativeZero, |
100 | }; |
101 | |
102 | /* Represents floating point arithmetic semantics. */ |
103 | struct fltSemantics { |
104 | /* The largest E such that 2^E is representable; this matches the |
105 | definition of IEEE 754. */ |
106 | APFloatBase::ExponentType maxExponent; |
107 | |
108 | /* The smallest E such that 2^E is a normalized number; this |
109 | matches the definition of IEEE 754. */ |
110 | APFloatBase::ExponentType minExponent; |
111 | |
112 | /* Number of bits in the significand. This includes the integer |
113 | bit. */ |
114 | unsigned int precision; |
115 | |
116 | /* Number of bits actually used in the semantics. */ |
117 | unsigned int sizeInBits; |
118 | |
119 | fltNonfiniteBehavior nonFiniteBehavior = fltNonfiniteBehavior::IEEE754; |
120 | |
121 | fltNanEncoding nanEncoding = fltNanEncoding::IEEE; |
122 | // Returns true if any number described by this semantics can be precisely |
123 | // represented by the specified semantics. Does not take into account |
124 | // the value of fltNonfiniteBehavior. |
125 | bool isRepresentableBy(const fltSemantics &S) const { |
126 | return maxExponent <= S.maxExponent && minExponent >= S.minExponent && |
127 | precision <= S.precision; |
128 | } |
129 | }; |
130 | |
131 | static constexpr fltSemantics semIEEEhalf = {.maxExponent: 15, .minExponent: -14, .precision: 11, .sizeInBits: 16}; |
132 | static constexpr fltSemantics semBFloat = {.maxExponent: 127, .minExponent: -126, .precision: 8, .sizeInBits: 16}; |
133 | static constexpr fltSemantics semIEEEsingle = {.maxExponent: 127, .minExponent: -126, .precision: 24, .sizeInBits: 32}; |
134 | static constexpr fltSemantics semIEEEdouble = {.maxExponent: 1023, .minExponent: -1022, .precision: 53, .sizeInBits: 64}; |
135 | static constexpr fltSemantics semIEEEquad = {.maxExponent: 16383, .minExponent: -16382, .precision: 113, .sizeInBits: 128}; |
136 | static constexpr fltSemantics semFloat8E5M2 = {.maxExponent: 15, .minExponent: -14, .precision: 3, .sizeInBits: 8}; |
137 | static constexpr fltSemantics semFloat8E5M2FNUZ = { |
138 | .maxExponent: 15, .minExponent: -15, .precision: 3, .sizeInBits: 8, .nonFiniteBehavior: fltNonfiniteBehavior::NanOnly, .nanEncoding: fltNanEncoding::NegativeZero}; |
139 | static constexpr fltSemantics semFloat8E4M3 = {.maxExponent: 7, .minExponent: -6, .precision: 4, .sizeInBits: 8}; |
140 | static constexpr fltSemantics semFloat8E4M3FN = { |
141 | .maxExponent: 8, .minExponent: -6, .precision: 4, .sizeInBits: 8, .nonFiniteBehavior: fltNonfiniteBehavior::NanOnly, .nanEncoding: fltNanEncoding::AllOnes}; |
142 | static constexpr fltSemantics semFloat8E4M3FNUZ = { |
143 | .maxExponent: 7, .minExponent: -7, .precision: 4, .sizeInBits: 8, .nonFiniteBehavior: fltNonfiniteBehavior::NanOnly, .nanEncoding: fltNanEncoding::NegativeZero}; |
144 | static constexpr fltSemantics semFloat8E4M3B11FNUZ = { |
145 | .maxExponent: 4, .minExponent: -10, .precision: 4, .sizeInBits: 8, .nonFiniteBehavior: fltNonfiniteBehavior::NanOnly, .nanEncoding: fltNanEncoding::NegativeZero}; |
146 | static constexpr fltSemantics semFloatTF32 = {.maxExponent: 127, .minExponent: -126, .precision: 11, .sizeInBits: 19}; |
147 | static constexpr fltSemantics semFloat6E3M2FN = { |
148 | .maxExponent: 4, .minExponent: -2, .precision: 3, .sizeInBits: 6, .nonFiniteBehavior: fltNonfiniteBehavior::FiniteOnly}; |
149 | static constexpr fltSemantics semFloat6E2M3FN = { |
150 | .maxExponent: 2, .minExponent: 0, .precision: 4, .sizeInBits: 6, .nonFiniteBehavior: fltNonfiniteBehavior::FiniteOnly}; |
151 | static constexpr fltSemantics semFloat4E2M1FN = { |
152 | .maxExponent: 2, .minExponent: 0, .precision: 2, .sizeInBits: 4, .nonFiniteBehavior: fltNonfiniteBehavior::FiniteOnly}; |
153 | static constexpr fltSemantics semX87DoubleExtended = {.maxExponent: 16383, .minExponent: -16382, .precision: 64, .sizeInBits: 80}; |
154 | static constexpr fltSemantics semBogus = {.maxExponent: 0, .minExponent: 0, .precision: 0, .sizeInBits: 0}; |
155 | |
156 | /* The IBM double-double semantics. Such a number consists of a pair of IEEE |
157 | 64-bit doubles (Hi, Lo), where |Hi| > |Lo|, and if normal, |
158 | (double)(Hi + Lo) == Hi. The numeric value it's modeling is Hi + Lo. |
159 | Therefore it has two 53-bit mantissa parts that aren't necessarily adjacent |
160 | to each other, and two 11-bit exponents. |
161 | |
162 | Note: we need to make the value different from semBogus as otherwise |
163 | an unsafe optimization may collapse both values to a single address, |
164 | and we heavily rely on them having distinct addresses. */ |
165 | static constexpr fltSemantics semPPCDoubleDouble = {.maxExponent: -1, .minExponent: 0, .precision: 0, .sizeInBits: 128}; |
166 | |
167 | /* These are legacy semantics for the fallback, inaccrurate implementation of |
168 | IBM double-double, if the accurate semPPCDoubleDouble doesn't handle the |
169 | operation. It's equivalent to having an IEEE number with consecutive 106 |
170 | bits of mantissa and 11 bits of exponent. |
171 | |
172 | It's not equivalent to IBM double-double. For example, a legit IBM |
173 | double-double, 1 + epsilon: |
174 | |
175 | 1 + epsilon = 1 + (1 >> 1076) |
176 | |
177 | is not representable by a consecutive 106 bits of mantissa. |
178 | |
179 | Currently, these semantics are used in the following way: |
180 | |
181 | semPPCDoubleDouble -> (IEEEdouble, IEEEdouble) -> |
182 | (64-bit APInt, 64-bit APInt) -> (128-bit APInt) -> |
183 | semPPCDoubleDoubleLegacy -> IEEE operations |
184 | |
185 | We use bitcastToAPInt() to get the bit representation (in APInt) of the |
186 | underlying IEEEdouble, then use the APInt constructor to construct the |
187 | legacy IEEE float. |
188 | |
189 | TODO: Implement all operations in semPPCDoubleDouble, and delete these |
190 | semantics. */ |
191 | static constexpr fltSemantics semPPCDoubleDoubleLegacy = {.maxExponent: 1023, .minExponent: -1022 + 53, |
192 | .precision: 53 + 53, .sizeInBits: 128}; |
193 | |
194 | const llvm::fltSemantics &APFloatBase::EnumToSemantics(Semantics S) { |
195 | switch (S) { |
196 | case S_IEEEhalf: |
197 | return IEEEhalf(); |
198 | case S_BFloat: |
199 | return BFloat(); |
200 | case S_IEEEsingle: |
201 | return IEEEsingle(); |
202 | case S_IEEEdouble: |
203 | return IEEEdouble(); |
204 | case S_IEEEquad: |
205 | return IEEEquad(); |
206 | case S_PPCDoubleDouble: |
207 | return PPCDoubleDouble(); |
208 | case S_Float8E5M2: |
209 | return Float8E5M2(); |
210 | case S_Float8E5M2FNUZ: |
211 | return Float8E5M2FNUZ(); |
212 | case S_Float8E4M3: |
213 | return Float8E4M3(); |
214 | case S_Float8E4M3FN: |
215 | return Float8E4M3FN(); |
216 | case S_Float8E4M3FNUZ: |
217 | return Float8E4M3FNUZ(); |
218 | case S_Float8E4M3B11FNUZ: |
219 | return Float8E4M3B11FNUZ(); |
220 | case S_FloatTF32: |
221 | return FloatTF32(); |
222 | case S_Float6E3M2FN: |
223 | return Float6E3M2FN(); |
224 | case S_Float6E2M3FN: |
225 | return Float6E2M3FN(); |
226 | case S_Float4E2M1FN: |
227 | return Float4E2M1FN(); |
228 | case S_x87DoubleExtended: |
229 | return x87DoubleExtended(); |
230 | } |
231 | llvm_unreachable("Unrecognised floating semantics" ); |
232 | } |
233 | |
234 | APFloatBase::Semantics |
235 | APFloatBase::SemanticsToEnum(const llvm::fltSemantics &Sem) { |
236 | if (&Sem == &llvm::APFloat::IEEEhalf()) |
237 | return S_IEEEhalf; |
238 | else if (&Sem == &llvm::APFloat::BFloat()) |
239 | return S_BFloat; |
240 | else if (&Sem == &llvm::APFloat::IEEEsingle()) |
241 | return S_IEEEsingle; |
242 | else if (&Sem == &llvm::APFloat::IEEEdouble()) |
243 | return S_IEEEdouble; |
244 | else if (&Sem == &llvm::APFloat::IEEEquad()) |
245 | return S_IEEEquad; |
246 | else if (&Sem == &llvm::APFloat::PPCDoubleDouble()) |
247 | return S_PPCDoubleDouble; |
248 | else if (&Sem == &llvm::APFloat::Float8E5M2()) |
249 | return S_Float8E5M2; |
250 | else if (&Sem == &llvm::APFloat::Float8E5M2FNUZ()) |
251 | return S_Float8E5M2FNUZ; |
252 | else if (&Sem == &llvm::APFloat::Float8E4M3()) |
253 | return S_Float8E4M3; |
254 | else if (&Sem == &llvm::APFloat::Float8E4M3FN()) |
255 | return S_Float8E4M3FN; |
256 | else if (&Sem == &llvm::APFloat::Float8E4M3FNUZ()) |
257 | return S_Float8E4M3FNUZ; |
258 | else if (&Sem == &llvm::APFloat::Float8E4M3B11FNUZ()) |
259 | return S_Float8E4M3B11FNUZ; |
260 | else if (&Sem == &llvm::APFloat::FloatTF32()) |
261 | return S_FloatTF32; |
262 | else if (&Sem == &llvm::APFloat::Float6E3M2FN()) |
263 | return S_Float6E3M2FN; |
264 | else if (&Sem == &llvm::APFloat::Float6E2M3FN()) |
265 | return S_Float6E2M3FN; |
266 | else if (&Sem == &llvm::APFloat::Float4E2M1FN()) |
267 | return S_Float4E2M1FN; |
268 | else if (&Sem == &llvm::APFloat::x87DoubleExtended()) |
269 | return S_x87DoubleExtended; |
270 | else |
271 | llvm_unreachable("Unknown floating semantics" ); |
272 | } |
273 | |
274 | const fltSemantics &APFloatBase::IEEEhalf() { return semIEEEhalf; } |
275 | const fltSemantics &APFloatBase::BFloat() { return semBFloat; } |
276 | const fltSemantics &APFloatBase::IEEEsingle() { return semIEEEsingle; } |
277 | const fltSemantics &APFloatBase::IEEEdouble() { return semIEEEdouble; } |
278 | const fltSemantics &APFloatBase::IEEEquad() { return semIEEEquad; } |
279 | const fltSemantics &APFloatBase::PPCDoubleDouble() { |
280 | return semPPCDoubleDouble; |
281 | } |
282 | const fltSemantics &APFloatBase::Float8E5M2() { return semFloat8E5M2; } |
283 | const fltSemantics &APFloatBase::Float8E5M2FNUZ() { return semFloat8E5M2FNUZ; } |
284 | const fltSemantics &APFloatBase::Float8E4M3() { return semFloat8E4M3; } |
285 | const fltSemantics &APFloatBase::Float8E4M3FN() { return semFloat8E4M3FN; } |
286 | const fltSemantics &APFloatBase::Float8E4M3FNUZ() { return semFloat8E4M3FNUZ; } |
287 | const fltSemantics &APFloatBase::Float8E4M3B11FNUZ() { |
288 | return semFloat8E4M3B11FNUZ; |
289 | } |
290 | const fltSemantics &APFloatBase::FloatTF32() { return semFloatTF32; } |
291 | const fltSemantics &APFloatBase::Float6E3M2FN() { return semFloat6E3M2FN; } |
292 | const fltSemantics &APFloatBase::Float6E2M3FN() { return semFloat6E2M3FN; } |
293 | const fltSemantics &APFloatBase::Float4E2M1FN() { return semFloat4E2M1FN; } |
294 | const fltSemantics &APFloatBase::x87DoubleExtended() { |
295 | return semX87DoubleExtended; |
296 | } |
297 | const fltSemantics &APFloatBase::Bogus() { return semBogus; } |
298 | |
299 | constexpr RoundingMode APFloatBase::rmNearestTiesToEven; |
300 | constexpr RoundingMode APFloatBase::rmTowardPositive; |
301 | constexpr RoundingMode APFloatBase::rmTowardNegative; |
302 | constexpr RoundingMode APFloatBase::rmTowardZero; |
303 | constexpr RoundingMode APFloatBase::rmNearestTiesToAway; |
304 | |
305 | /* A tight upper bound on number of parts required to hold the value |
306 | pow(5, power) is |
307 | |
308 | power * 815 / (351 * integerPartWidth) + 1 |
309 | |
310 | However, whilst the result may require only this many parts, |
311 | because we are multiplying two values to get it, the |
312 | multiplication may require an extra part with the excess part |
313 | being zero (consider the trivial case of 1 * 1, tcFullMultiply |
314 | requires two parts to hold the single-part result). So we add an |
315 | extra one to guarantee enough space whilst multiplying. */ |
316 | const unsigned int maxExponent = 16383; |
317 | const unsigned int maxPrecision = 113; |
318 | const unsigned int maxPowerOfFiveExponent = maxExponent + maxPrecision - 1; |
319 | const unsigned int maxPowerOfFiveParts = |
320 | 2 + |
321 | ((maxPowerOfFiveExponent * 815) / (351 * APFloatBase::integerPartWidth)); |
322 | |
323 | unsigned int APFloatBase::semanticsPrecision(const fltSemantics &semantics) { |
324 | return semantics.precision; |
325 | } |
326 | APFloatBase::ExponentType |
327 | APFloatBase::semanticsMaxExponent(const fltSemantics &semantics) { |
328 | return semantics.maxExponent; |
329 | } |
330 | APFloatBase::ExponentType |
331 | APFloatBase::semanticsMinExponent(const fltSemantics &semantics) { |
332 | return semantics.minExponent; |
333 | } |
334 | unsigned int APFloatBase::semanticsSizeInBits(const fltSemantics &semantics) { |
335 | return semantics.sizeInBits; |
336 | } |
337 | unsigned int APFloatBase::semanticsIntSizeInBits(const fltSemantics &semantics, |
338 | bool isSigned) { |
339 | // The max FP value is pow(2, MaxExponent) * (1 + MaxFraction), so we need |
340 | // at least one more bit than the MaxExponent to hold the max FP value. |
341 | unsigned int MinBitWidth = semanticsMaxExponent(semantics) + 1; |
342 | // Extra sign bit needed. |
343 | if (isSigned) |
344 | ++MinBitWidth; |
345 | return MinBitWidth; |
346 | } |
347 | |
348 | bool APFloatBase::isRepresentableAsNormalIn(const fltSemantics &Src, |
349 | const fltSemantics &Dst) { |
350 | // Exponent range must be larger. |
351 | if (Src.maxExponent >= Dst.maxExponent || Src.minExponent <= Dst.minExponent) |
352 | return false; |
353 | |
354 | // If the mantissa is long enough, the result value could still be denormal |
355 | // with a larger exponent range. |
356 | // |
357 | // FIXME: This condition is probably not accurate but also shouldn't be a |
358 | // practical concern with existing types. |
359 | return Dst.precision >= Src.precision; |
360 | } |
361 | |
362 | unsigned APFloatBase::getSizeInBits(const fltSemantics &Sem) { |
363 | return Sem.sizeInBits; |
364 | } |
365 | |
366 | static constexpr APFloatBase::ExponentType |
367 | exponentZero(const fltSemantics &semantics) { |
368 | return semantics.minExponent - 1; |
369 | } |
370 | |
371 | static constexpr APFloatBase::ExponentType |
372 | exponentInf(const fltSemantics &semantics) { |
373 | return semantics.maxExponent + 1; |
374 | } |
375 | |
376 | static constexpr APFloatBase::ExponentType |
377 | exponentNaN(const fltSemantics &semantics) { |
378 | if (semantics.nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
379 | if (semantics.nanEncoding == fltNanEncoding::NegativeZero) |
380 | return exponentZero(semantics); |
381 | return semantics.maxExponent; |
382 | } |
383 | return semantics.maxExponent + 1; |
384 | } |
385 | |
386 | /* A bunch of private, handy routines. */ |
387 | |
388 | static inline Error createError(const Twine &Err) { |
389 | return make_error<StringError>(Args: Err, Args: inconvertibleErrorCode()); |
390 | } |
391 | |
392 | static constexpr inline unsigned int partCountForBits(unsigned int bits) { |
393 | return ((bits) + APFloatBase::integerPartWidth - 1) / APFloatBase::integerPartWidth; |
394 | } |
395 | |
396 | /* Returns 0U-9U. Return values >= 10U are not digits. */ |
397 | static inline unsigned int |
398 | decDigitValue(unsigned int c) |
399 | { |
400 | return c - '0'; |
401 | } |
402 | |
403 | /* Return the value of a decimal exponent of the form |
404 | [+-]ddddddd. |
405 | |
406 | If the exponent overflows, returns a large exponent with the |
407 | appropriate sign. */ |
408 | static Expected<int> readExponent(StringRef::iterator begin, |
409 | StringRef::iterator end) { |
410 | bool isNegative; |
411 | unsigned int absExponent; |
412 | const unsigned int overlargeExponent = 24000; /* FIXME. */ |
413 | StringRef::iterator p = begin; |
414 | |
415 | // Treat no exponent as 0 to match binutils |
416 | if (p == end || ((*p == '-' || *p == '+') && (p + 1) == end)) { |
417 | return 0; |
418 | } |
419 | |
420 | isNegative = (*p == '-'); |
421 | if (*p == '-' || *p == '+') { |
422 | p++; |
423 | if (p == end) |
424 | return createError(Err: "Exponent has no digits" ); |
425 | } |
426 | |
427 | absExponent = decDigitValue(c: *p++); |
428 | if (absExponent >= 10U) |
429 | return createError(Err: "Invalid character in exponent" ); |
430 | |
431 | for (; p != end; ++p) { |
432 | unsigned int value; |
433 | |
434 | value = decDigitValue(c: *p); |
435 | if (value >= 10U) |
436 | return createError(Err: "Invalid character in exponent" ); |
437 | |
438 | absExponent = absExponent * 10U + value; |
439 | if (absExponent >= overlargeExponent) { |
440 | absExponent = overlargeExponent; |
441 | break; |
442 | } |
443 | } |
444 | |
445 | if (isNegative) |
446 | return -(int) absExponent; |
447 | else |
448 | return (int) absExponent; |
449 | } |
450 | |
451 | /* This is ugly and needs cleaning up, but I don't immediately see |
452 | how whilst remaining safe. */ |
453 | static Expected<int> totalExponent(StringRef::iterator p, |
454 | StringRef::iterator end, |
455 | int exponentAdjustment) { |
456 | int unsignedExponent; |
457 | bool negative, overflow; |
458 | int exponent = 0; |
459 | |
460 | if (p == end) |
461 | return createError(Err: "Exponent has no digits" ); |
462 | |
463 | negative = *p == '-'; |
464 | if (*p == '-' || *p == '+') { |
465 | p++; |
466 | if (p == end) |
467 | return createError(Err: "Exponent has no digits" ); |
468 | } |
469 | |
470 | unsignedExponent = 0; |
471 | overflow = false; |
472 | for (; p != end; ++p) { |
473 | unsigned int value; |
474 | |
475 | value = decDigitValue(c: *p); |
476 | if (value >= 10U) |
477 | return createError(Err: "Invalid character in exponent" ); |
478 | |
479 | unsignedExponent = unsignedExponent * 10 + value; |
480 | if (unsignedExponent > 32767) { |
481 | overflow = true; |
482 | break; |
483 | } |
484 | } |
485 | |
486 | if (exponentAdjustment > 32767 || exponentAdjustment < -32768) |
487 | overflow = true; |
488 | |
489 | if (!overflow) { |
490 | exponent = unsignedExponent; |
491 | if (negative) |
492 | exponent = -exponent; |
493 | exponent += exponentAdjustment; |
494 | if (exponent > 32767 || exponent < -32768) |
495 | overflow = true; |
496 | } |
497 | |
498 | if (overflow) |
499 | exponent = negative ? -32768: 32767; |
500 | |
501 | return exponent; |
502 | } |
503 | |
504 | static Expected<StringRef::iterator> |
505 | skipLeadingZeroesAndAnyDot(StringRef::iterator begin, StringRef::iterator end, |
506 | StringRef::iterator *dot) { |
507 | StringRef::iterator p = begin; |
508 | *dot = end; |
509 | while (p != end && *p == '0') |
510 | p++; |
511 | |
512 | if (p != end && *p == '.') { |
513 | *dot = p++; |
514 | |
515 | if (end - begin == 1) |
516 | return createError(Err: "Significand has no digits" ); |
517 | |
518 | while (p != end && *p == '0') |
519 | p++; |
520 | } |
521 | |
522 | return p; |
523 | } |
524 | |
525 | /* Given a normal decimal floating point number of the form |
526 | |
527 | dddd.dddd[eE][+-]ddd |
528 | |
529 | where the decimal point and exponent are optional, fill out the |
530 | structure D. Exponent is appropriate if the significand is |
531 | treated as an integer, and normalizedExponent if the significand |
532 | is taken to have the decimal point after a single leading |
533 | non-zero digit. |
534 | |
535 | If the value is zero, V->firstSigDigit points to a non-digit, and |
536 | the return exponent is zero. |
537 | */ |
538 | struct decimalInfo { |
539 | const char *firstSigDigit; |
540 | const char *lastSigDigit; |
541 | int exponent; |
542 | int normalizedExponent; |
543 | }; |
544 | |
545 | static Error interpretDecimal(StringRef::iterator begin, |
546 | StringRef::iterator end, decimalInfo *D) { |
547 | StringRef::iterator dot = end; |
548 | |
549 | auto PtrOrErr = skipLeadingZeroesAndAnyDot(begin, end, dot: &dot); |
550 | if (!PtrOrErr) |
551 | return PtrOrErr.takeError(); |
552 | StringRef::iterator p = *PtrOrErr; |
553 | |
554 | D->firstSigDigit = p; |
555 | D->exponent = 0; |
556 | D->normalizedExponent = 0; |
557 | |
558 | for (; p != end; ++p) { |
559 | if (*p == '.') { |
560 | if (dot != end) |
561 | return createError(Err: "String contains multiple dots" ); |
562 | dot = p++; |
563 | if (p == end) |
564 | break; |
565 | } |
566 | if (decDigitValue(c: *p) >= 10U) |
567 | break; |
568 | } |
569 | |
570 | if (p != end) { |
571 | if (*p != 'e' && *p != 'E') |
572 | return createError(Err: "Invalid character in significand" ); |
573 | if (p == begin) |
574 | return createError(Err: "Significand has no digits" ); |
575 | if (dot != end && p - begin == 1) |
576 | return createError(Err: "Significand has no digits" ); |
577 | |
578 | /* p points to the first non-digit in the string */ |
579 | auto ExpOrErr = readExponent(begin: p + 1, end); |
580 | if (!ExpOrErr) |
581 | return ExpOrErr.takeError(); |
582 | D->exponent = *ExpOrErr; |
583 | |
584 | /* Implied decimal point? */ |
585 | if (dot == end) |
586 | dot = p; |
587 | } |
588 | |
589 | /* If number is all zeroes accept any exponent. */ |
590 | if (p != D->firstSigDigit) { |
591 | /* Drop insignificant trailing zeroes. */ |
592 | if (p != begin) { |
593 | do |
594 | do |
595 | p--; |
596 | while (p != begin && *p == '0'); |
597 | while (p != begin && *p == '.'); |
598 | } |
599 | |
600 | /* Adjust the exponents for any decimal point. */ |
601 | D->exponent += static_cast<APFloat::ExponentType>((dot - p) - (dot > p)); |
602 | D->normalizedExponent = (D->exponent + |
603 | static_cast<APFloat::ExponentType>((p - D->firstSigDigit) |
604 | - (dot > D->firstSigDigit && dot < p))); |
605 | } |
606 | |
607 | D->lastSigDigit = p; |
608 | return Error::success(); |
609 | } |
610 | |
611 | /* Return the trailing fraction of a hexadecimal number. |
612 | DIGITVALUE is the first hex digit of the fraction, P points to |
613 | the next digit. */ |
614 | static Expected<lostFraction> |
615 | trailingHexadecimalFraction(StringRef::iterator p, StringRef::iterator end, |
616 | unsigned int digitValue) { |
617 | unsigned int hexDigit; |
618 | |
619 | /* If the first trailing digit isn't 0 or 8 we can work out the |
620 | fraction immediately. */ |
621 | if (digitValue > 8) |
622 | return lfMoreThanHalf; |
623 | else if (digitValue < 8 && digitValue > 0) |
624 | return lfLessThanHalf; |
625 | |
626 | // Otherwise we need to find the first non-zero digit. |
627 | while (p != end && (*p == '0' || *p == '.')) |
628 | p++; |
629 | |
630 | if (p == end) |
631 | return createError(Err: "Invalid trailing hexadecimal fraction!" ); |
632 | |
633 | hexDigit = hexDigitValue(C: *p); |
634 | |
635 | /* If we ran off the end it is exactly zero or one-half, otherwise |
636 | a little more. */ |
637 | if (hexDigit == UINT_MAX) |
638 | return digitValue == 0 ? lfExactlyZero: lfExactlyHalf; |
639 | else |
640 | return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf; |
641 | } |
642 | |
643 | /* Return the fraction lost were a bignum truncated losing the least |
644 | significant BITS bits. */ |
645 | static lostFraction |
646 | lostFractionThroughTruncation(const APFloatBase::integerPart *parts, |
647 | unsigned int partCount, |
648 | unsigned int bits) |
649 | { |
650 | unsigned int lsb; |
651 | |
652 | lsb = APInt::tcLSB(parts, n: partCount); |
653 | |
654 | /* Note this is guaranteed true if bits == 0, or LSB == UINT_MAX. */ |
655 | if (bits <= lsb) |
656 | return lfExactlyZero; |
657 | if (bits == lsb + 1) |
658 | return lfExactlyHalf; |
659 | if (bits <= partCount * APFloatBase::integerPartWidth && |
660 | APInt::tcExtractBit(parts, bit: bits - 1)) |
661 | return lfMoreThanHalf; |
662 | |
663 | return lfLessThanHalf; |
664 | } |
665 | |
666 | /* Shift DST right BITS bits noting lost fraction. */ |
667 | static lostFraction |
668 | shiftRight(APFloatBase::integerPart *dst, unsigned int parts, unsigned int bits) |
669 | { |
670 | lostFraction lost_fraction; |
671 | |
672 | lost_fraction = lostFractionThroughTruncation(parts: dst, partCount: parts, bits); |
673 | |
674 | APInt::tcShiftRight(dst, Words: parts, Count: bits); |
675 | |
676 | return lost_fraction; |
677 | } |
678 | |
679 | /* Combine the effect of two lost fractions. */ |
680 | static lostFraction |
681 | combineLostFractions(lostFraction moreSignificant, |
682 | lostFraction lessSignificant) |
683 | { |
684 | if (lessSignificant != lfExactlyZero) { |
685 | if (moreSignificant == lfExactlyZero) |
686 | moreSignificant = lfLessThanHalf; |
687 | else if (moreSignificant == lfExactlyHalf) |
688 | moreSignificant = lfMoreThanHalf; |
689 | } |
690 | |
691 | return moreSignificant; |
692 | } |
693 | |
694 | /* The error from the true value, in half-ulps, on multiplying two |
695 | floating point numbers, which differ from the value they |
696 | approximate by at most HUE1 and HUE2 half-ulps, is strictly less |
697 | than the returned value. |
698 | |
699 | See "How to Read Floating Point Numbers Accurately" by William D |
700 | Clinger. */ |
701 | static unsigned int |
702 | HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2) |
703 | { |
704 | assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8)); |
705 | |
706 | if (HUerr1 + HUerr2 == 0) |
707 | return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */ |
708 | else |
709 | return inexactMultiply + 2 * (HUerr1 + HUerr2); |
710 | } |
711 | |
712 | /* The number of ulps from the boundary (zero, or half if ISNEAREST) |
713 | when the least significant BITS are truncated. BITS cannot be |
714 | zero. */ |
715 | static APFloatBase::integerPart |
716 | ulpsFromBoundary(const APFloatBase::integerPart *parts, unsigned int bits, |
717 | bool isNearest) { |
718 | unsigned int count, partBits; |
719 | APFloatBase::integerPart part, boundary; |
720 | |
721 | assert(bits != 0); |
722 | |
723 | bits--; |
724 | count = bits / APFloatBase::integerPartWidth; |
725 | partBits = bits % APFloatBase::integerPartWidth + 1; |
726 | |
727 | part = parts[count] & (~(APFloatBase::integerPart) 0 >> (APFloatBase::integerPartWidth - partBits)); |
728 | |
729 | if (isNearest) |
730 | boundary = (APFloatBase::integerPart) 1 << (partBits - 1); |
731 | else |
732 | boundary = 0; |
733 | |
734 | if (count == 0) { |
735 | if (part - boundary <= boundary - part) |
736 | return part - boundary; |
737 | else |
738 | return boundary - part; |
739 | } |
740 | |
741 | if (part == boundary) { |
742 | while (--count) |
743 | if (parts[count]) |
744 | return ~(APFloatBase::integerPart) 0; /* A lot. */ |
745 | |
746 | return parts[0]; |
747 | } else if (part == boundary - 1) { |
748 | while (--count) |
749 | if (~parts[count]) |
750 | return ~(APFloatBase::integerPart) 0; /* A lot. */ |
751 | |
752 | return -parts[0]; |
753 | } |
754 | |
755 | return ~(APFloatBase::integerPart) 0; /* A lot. */ |
756 | } |
757 | |
758 | /* Place pow(5, power) in DST, and return the number of parts used. |
759 | DST must be at least one part larger than size of the answer. */ |
760 | static unsigned int |
761 | powerOf5(APFloatBase::integerPart *dst, unsigned int power) { |
762 | static const APFloatBase::integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125, 15625, 78125 }; |
763 | APFloatBase::integerPart pow5s[maxPowerOfFiveParts * 2 + 5]; |
764 | pow5s[0] = 78125 * 5; |
765 | |
766 | unsigned int partsCount = 1; |
767 | APFloatBase::integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5; |
768 | unsigned int result; |
769 | assert(power <= maxExponent); |
770 | |
771 | p1 = dst; |
772 | p2 = scratch; |
773 | |
774 | *p1 = firstEightPowers[power & 7]; |
775 | power >>= 3; |
776 | |
777 | result = 1; |
778 | pow5 = pow5s; |
779 | |
780 | for (unsigned int n = 0; power; power >>= 1, n++) { |
781 | /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */ |
782 | if (n != 0) { |
783 | APInt::tcFullMultiply(pow5, pow5 - partsCount, pow5 - partsCount, |
784 | partsCount, partsCount); |
785 | partsCount *= 2; |
786 | if (pow5[partsCount - 1] == 0) |
787 | partsCount--; |
788 | } |
789 | |
790 | if (power & 1) { |
791 | APFloatBase::integerPart *tmp; |
792 | |
793 | APInt::tcFullMultiply(p2, p1, pow5, result, partsCount); |
794 | result += partsCount; |
795 | if (p2[result - 1] == 0) |
796 | result--; |
797 | |
798 | /* Now result is in p1 with partsCount parts and p2 is scratch |
799 | space. */ |
800 | tmp = p1; |
801 | p1 = p2; |
802 | p2 = tmp; |
803 | } |
804 | |
805 | pow5 += partsCount; |
806 | } |
807 | |
808 | if (p1 != dst) |
809 | APInt::tcAssign(dst, p1, result); |
810 | |
811 | return result; |
812 | } |
813 | |
814 | /* Zero at the end to avoid modular arithmetic when adding one; used |
815 | when rounding up during hexadecimal output. */ |
816 | static const char hexDigitsLower[] = "0123456789abcdef0" ; |
817 | static const char hexDigitsUpper[] = "0123456789ABCDEF0" ; |
818 | static const char infinityL[] = "infinity" ; |
819 | static const char infinityU[] = "INFINITY" ; |
820 | static const char NaNL[] = "nan" ; |
821 | static const char NaNU[] = "NAN" ; |
822 | |
823 | /* Write out an integerPart in hexadecimal, starting with the most |
824 | significant nibble. Write out exactly COUNT hexdigits, return |
825 | COUNT. */ |
826 | static unsigned int |
827 | partAsHex (char *dst, APFloatBase::integerPart part, unsigned int count, |
828 | const char *hexDigitChars) |
829 | { |
830 | unsigned int result = count; |
831 | |
832 | assert(count != 0 && count <= APFloatBase::integerPartWidth / 4); |
833 | |
834 | part >>= (APFloatBase::integerPartWidth - 4 * count); |
835 | while (count--) { |
836 | dst[count] = hexDigitChars[part & 0xf]; |
837 | part >>= 4; |
838 | } |
839 | |
840 | return result; |
841 | } |
842 | |
843 | /* Write out an unsigned decimal integer. */ |
844 | static char * |
845 | writeUnsignedDecimal (char *dst, unsigned int n) |
846 | { |
847 | char buff[40], *p; |
848 | |
849 | p = buff; |
850 | do |
851 | *p++ = '0' + n % 10; |
852 | while (n /= 10); |
853 | |
854 | do |
855 | *dst++ = *--p; |
856 | while (p != buff); |
857 | |
858 | return dst; |
859 | } |
860 | |
861 | /* Write out a signed decimal integer. */ |
862 | static char * |
863 | writeSignedDecimal (char *dst, int value) |
864 | { |
865 | if (value < 0) { |
866 | *dst++ = '-'; |
867 | dst = writeUnsignedDecimal(dst, n: -(unsigned) value); |
868 | } else |
869 | dst = writeUnsignedDecimal(dst, n: value); |
870 | |
871 | return dst; |
872 | } |
873 | |
874 | namespace detail { |
875 | /* Constructors. */ |
876 | void IEEEFloat::initialize(const fltSemantics *ourSemantics) { |
877 | unsigned int count; |
878 | |
879 | semantics = ourSemantics; |
880 | count = partCount(); |
881 | if (count > 1) |
882 | significand.parts = new integerPart[count]; |
883 | } |
884 | |
885 | void IEEEFloat::freeSignificand() { |
886 | if (needsCleanup()) |
887 | delete [] significand.parts; |
888 | } |
889 | |
890 | void IEEEFloat::assign(const IEEEFloat &rhs) { |
891 | assert(semantics == rhs.semantics); |
892 | |
893 | sign = rhs.sign; |
894 | category = rhs.category; |
895 | exponent = rhs.exponent; |
896 | if (isFiniteNonZero() || category == fcNaN) |
897 | copySignificand(rhs); |
898 | } |
899 | |
900 | void IEEEFloat::copySignificand(const IEEEFloat &rhs) { |
901 | assert(isFiniteNonZero() || category == fcNaN); |
902 | assert(rhs.partCount() >= partCount()); |
903 | |
904 | APInt::tcAssign(significandParts(), rhs.significandParts(), |
905 | partCount()); |
906 | } |
907 | |
908 | /* Make this number a NaN, with an arbitrary but deterministic value |
909 | for the significand. If double or longer, this is a signalling NaN, |
910 | which may not be ideal. If float, this is QNaN(0). */ |
911 | void IEEEFloat::makeNaN(bool SNaN, bool Negative, const APInt *fill) { |
912 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
913 | llvm_unreachable("This floating point format does not support NaN" ); |
914 | |
915 | category = fcNaN; |
916 | sign = Negative; |
917 | exponent = exponentNaN(); |
918 | |
919 | integerPart *significand = significandParts(); |
920 | unsigned numParts = partCount(); |
921 | |
922 | APInt fill_storage; |
923 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
924 | // Finite-only types do not distinguish signalling and quiet NaN, so |
925 | // make them all signalling. |
926 | SNaN = false; |
927 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) { |
928 | sign = true; |
929 | fill_storage = APInt::getZero(numBits: semantics->precision - 1); |
930 | } else { |
931 | fill_storage = APInt::getAllOnes(numBits: semantics->precision - 1); |
932 | } |
933 | fill = &fill_storage; |
934 | } |
935 | |
936 | // Set the significand bits to the fill. |
937 | if (!fill || fill->getNumWords() < numParts) |
938 | APInt::tcSet(significand, 0, numParts); |
939 | if (fill) { |
940 | APInt::tcAssign(significand, fill->getRawData(), |
941 | std::min(a: fill->getNumWords(), b: numParts)); |
942 | |
943 | // Zero out the excess bits of the significand. |
944 | unsigned bitsToPreserve = semantics->precision - 1; |
945 | unsigned part = bitsToPreserve / 64; |
946 | bitsToPreserve %= 64; |
947 | significand[part] &= ((1ULL << bitsToPreserve) - 1); |
948 | for (part++; part != numParts; ++part) |
949 | significand[part] = 0; |
950 | } |
951 | |
952 | unsigned QNaNBit = semantics->precision - 2; |
953 | |
954 | if (SNaN) { |
955 | // We always have to clear the QNaN bit to make it an SNaN. |
956 | APInt::tcClearBit(significand, bit: QNaNBit); |
957 | |
958 | // If there are no bits set in the payload, we have to set |
959 | // *something* to make it a NaN instead of an infinity; |
960 | // conventionally, this is the next bit down from the QNaN bit. |
961 | if (APInt::tcIsZero(significand, numParts)) |
962 | APInt::tcSetBit(significand, bit: QNaNBit - 1); |
963 | } else if (semantics->nanEncoding == fltNanEncoding::NegativeZero) { |
964 | // The only NaN is a quiet NaN, and it has no bits sets in the significand. |
965 | // Do nothing. |
966 | } else { |
967 | // We always have to set the QNaN bit to make it a QNaN. |
968 | APInt::tcSetBit(significand, bit: QNaNBit); |
969 | } |
970 | |
971 | // For x87 extended precision, we want to make a NaN, not a |
972 | // pseudo-NaN. Maybe we should expose the ability to make |
973 | // pseudo-NaNs? |
974 | if (semantics == &semX87DoubleExtended) |
975 | APInt::tcSetBit(significand, bit: QNaNBit + 1); |
976 | } |
977 | |
978 | IEEEFloat &IEEEFloat::operator=(const IEEEFloat &rhs) { |
979 | if (this != &rhs) { |
980 | if (semantics != rhs.semantics) { |
981 | freeSignificand(); |
982 | initialize(ourSemantics: rhs.semantics); |
983 | } |
984 | assign(rhs); |
985 | } |
986 | |
987 | return *this; |
988 | } |
989 | |
990 | IEEEFloat &IEEEFloat::operator=(IEEEFloat &&rhs) { |
991 | freeSignificand(); |
992 | |
993 | semantics = rhs.semantics; |
994 | significand = rhs.significand; |
995 | exponent = rhs.exponent; |
996 | category = rhs.category; |
997 | sign = rhs.sign; |
998 | |
999 | rhs.semantics = &semBogus; |
1000 | return *this; |
1001 | } |
1002 | |
1003 | bool IEEEFloat::isDenormal() const { |
1004 | return isFiniteNonZero() && (exponent == semantics->minExponent) && |
1005 | (APInt::tcExtractBit(significandParts(), |
1006 | bit: semantics->precision - 1) == 0); |
1007 | } |
1008 | |
1009 | bool IEEEFloat::isSmallest() const { |
1010 | // The smallest number by magnitude in our format will be the smallest |
1011 | // denormal, i.e. the floating point number with exponent being minimum |
1012 | // exponent and significand bitwise equal to 1 (i.e. with MSB equal to 0). |
1013 | return isFiniteNonZero() && exponent == semantics->minExponent && |
1014 | significandMSB() == 0; |
1015 | } |
1016 | |
1017 | bool IEEEFloat::isSmallestNormalized() const { |
1018 | return getCategory() == fcNormal && exponent == semantics->minExponent && |
1019 | isSignificandAllZerosExceptMSB(); |
1020 | } |
1021 | |
1022 | bool IEEEFloat::isSignificandAllOnes() const { |
1023 | // Test if the significand excluding the integral bit is all ones. This allows |
1024 | // us to test for binade boundaries. |
1025 | const integerPart *Parts = significandParts(); |
1026 | const unsigned PartCount = partCountForBits(bits: semantics->precision); |
1027 | for (unsigned i = 0; i < PartCount - 1; i++) |
1028 | if (~Parts[i]) |
1029 | return false; |
1030 | |
1031 | // Set the unused high bits to all ones when we compare. |
1032 | const unsigned NumHighBits = |
1033 | PartCount*integerPartWidth - semantics->precision + 1; |
1034 | assert(NumHighBits <= integerPartWidth && NumHighBits > 0 && |
1035 | "Can not have more high bits to fill than integerPartWidth" ); |
1036 | const integerPart HighBitFill = |
1037 | ~integerPart(0) << (integerPartWidth - NumHighBits); |
1038 | if (~(Parts[PartCount - 1] | HighBitFill)) |
1039 | return false; |
1040 | |
1041 | return true; |
1042 | } |
1043 | |
1044 | bool IEEEFloat::isSignificandAllOnesExceptLSB() const { |
1045 | // Test if the significand excluding the integral bit is all ones except for |
1046 | // the least significant bit. |
1047 | const integerPart *Parts = significandParts(); |
1048 | |
1049 | if (Parts[0] & 1) |
1050 | return false; |
1051 | |
1052 | const unsigned PartCount = partCountForBits(bits: semantics->precision); |
1053 | for (unsigned i = 0; i < PartCount - 1; i++) { |
1054 | if (~Parts[i] & ~unsigned{!i}) |
1055 | return false; |
1056 | } |
1057 | |
1058 | // Set the unused high bits to all ones when we compare. |
1059 | const unsigned NumHighBits = |
1060 | PartCount * integerPartWidth - semantics->precision + 1; |
1061 | assert(NumHighBits <= integerPartWidth && NumHighBits > 0 && |
1062 | "Can not have more high bits to fill than integerPartWidth" ); |
1063 | const integerPart HighBitFill = ~integerPart(0) |
1064 | << (integerPartWidth - NumHighBits); |
1065 | if (~(Parts[PartCount - 1] | HighBitFill | 0x1)) |
1066 | return false; |
1067 | |
1068 | return true; |
1069 | } |
1070 | |
1071 | bool IEEEFloat::isSignificandAllZeros() const { |
1072 | // Test if the significand excluding the integral bit is all zeros. This |
1073 | // allows us to test for binade boundaries. |
1074 | const integerPart *Parts = significandParts(); |
1075 | const unsigned PartCount = partCountForBits(bits: semantics->precision); |
1076 | |
1077 | for (unsigned i = 0; i < PartCount - 1; i++) |
1078 | if (Parts[i]) |
1079 | return false; |
1080 | |
1081 | // Compute how many bits are used in the final word. |
1082 | const unsigned NumHighBits = |
1083 | PartCount*integerPartWidth - semantics->precision + 1; |
1084 | assert(NumHighBits < integerPartWidth && "Can not have more high bits to " |
1085 | "clear than integerPartWidth" ); |
1086 | const integerPart HighBitMask = ~integerPart(0) >> NumHighBits; |
1087 | |
1088 | if (Parts[PartCount - 1] & HighBitMask) |
1089 | return false; |
1090 | |
1091 | return true; |
1092 | } |
1093 | |
1094 | bool IEEEFloat::isSignificandAllZerosExceptMSB() const { |
1095 | const integerPart *Parts = significandParts(); |
1096 | const unsigned PartCount = partCountForBits(bits: semantics->precision); |
1097 | |
1098 | for (unsigned i = 0; i < PartCount - 1; i++) { |
1099 | if (Parts[i]) |
1100 | return false; |
1101 | } |
1102 | |
1103 | const unsigned NumHighBits = |
1104 | PartCount * integerPartWidth - semantics->precision + 1; |
1105 | return Parts[PartCount - 1] == integerPart(1) |
1106 | << (integerPartWidth - NumHighBits); |
1107 | } |
1108 | |
1109 | bool IEEEFloat::isLargest() const { |
1110 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly && |
1111 | semantics->nanEncoding == fltNanEncoding::AllOnes) { |
1112 | // The largest number by magnitude in our format will be the floating point |
1113 | // number with maximum exponent and with significand that is all ones except |
1114 | // the LSB. |
1115 | return isFiniteNonZero() && exponent == semantics->maxExponent && |
1116 | isSignificandAllOnesExceptLSB(); |
1117 | } else { |
1118 | // The largest number by magnitude in our format will be the floating point |
1119 | // number with maximum exponent and with significand that is all ones. |
1120 | return isFiniteNonZero() && exponent == semantics->maxExponent && |
1121 | isSignificandAllOnes(); |
1122 | } |
1123 | } |
1124 | |
1125 | bool IEEEFloat::isInteger() const { |
1126 | // This could be made more efficient; I'm going for obviously correct. |
1127 | if (!isFinite()) return false; |
1128 | IEEEFloat truncated = *this; |
1129 | truncated.roundToIntegral(rmTowardZero); |
1130 | return compare(truncated) == cmpEqual; |
1131 | } |
1132 | |
1133 | bool IEEEFloat::bitwiseIsEqual(const IEEEFloat &rhs) const { |
1134 | if (this == &rhs) |
1135 | return true; |
1136 | if (semantics != rhs.semantics || |
1137 | category != rhs.category || |
1138 | sign != rhs.sign) |
1139 | return false; |
1140 | if (category==fcZero || category==fcInfinity) |
1141 | return true; |
1142 | |
1143 | if (isFiniteNonZero() && exponent != rhs.exponent) |
1144 | return false; |
1145 | |
1146 | return std::equal(significandParts(), significandParts() + partCount(), |
1147 | rhs.significandParts()); |
1148 | } |
1149 | |
1150 | IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics, integerPart value) { |
1151 | initialize(ourSemantics: &ourSemantics); |
1152 | sign = 0; |
1153 | category = fcNormal; |
1154 | zeroSignificand(); |
1155 | exponent = ourSemantics.precision - 1; |
1156 | significandParts()[0] = value; |
1157 | normalize(rmNearestTiesToEven, lfExactlyZero); |
1158 | } |
1159 | |
1160 | IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics) { |
1161 | initialize(ourSemantics: &ourSemantics); |
1162 | makeZero(Neg: false); |
1163 | } |
1164 | |
1165 | // Delegate to the previous constructor, because later copy constructor may |
1166 | // actually inspects category, which can't be garbage. |
1167 | IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics, uninitializedTag tag) |
1168 | : IEEEFloat(ourSemantics) {} |
1169 | |
1170 | IEEEFloat::IEEEFloat(const IEEEFloat &rhs) { |
1171 | initialize(ourSemantics: rhs.semantics); |
1172 | assign(rhs); |
1173 | } |
1174 | |
1175 | IEEEFloat::IEEEFloat(IEEEFloat &&rhs) : semantics(&semBogus) { |
1176 | *this = std::move(rhs); |
1177 | } |
1178 | |
1179 | IEEEFloat::~IEEEFloat() { freeSignificand(); } |
1180 | |
1181 | unsigned int IEEEFloat::partCount() const { |
1182 | return partCountForBits(bits: semantics->precision + 1); |
1183 | } |
1184 | |
1185 | const IEEEFloat::integerPart *IEEEFloat::significandParts() const { |
1186 | return const_cast<IEEEFloat *>(this)->significandParts(); |
1187 | } |
1188 | |
1189 | IEEEFloat::integerPart *IEEEFloat::significandParts() { |
1190 | if (partCount() > 1) |
1191 | return significand.parts; |
1192 | else |
1193 | return &significand.part; |
1194 | } |
1195 | |
1196 | void IEEEFloat::zeroSignificand() { |
1197 | APInt::tcSet(significandParts(), 0, partCount()); |
1198 | } |
1199 | |
1200 | /* Increment an fcNormal floating point number's significand. */ |
1201 | void IEEEFloat::incrementSignificand() { |
1202 | integerPart carry; |
1203 | |
1204 | carry = APInt::tcIncrement(dst: significandParts(), parts: partCount()); |
1205 | |
1206 | /* Our callers should never cause us to overflow. */ |
1207 | assert(carry == 0); |
1208 | (void)carry; |
1209 | } |
1210 | |
1211 | /* Add the significand of the RHS. Returns the carry flag. */ |
1212 | IEEEFloat::integerPart IEEEFloat::addSignificand(const IEEEFloat &rhs) { |
1213 | integerPart *parts; |
1214 | |
1215 | parts = significandParts(); |
1216 | |
1217 | assert(semantics == rhs.semantics); |
1218 | assert(exponent == rhs.exponent); |
1219 | |
1220 | return APInt::tcAdd(parts, rhs.significandParts(), carry: 0, partCount()); |
1221 | } |
1222 | |
1223 | /* Subtract the significand of the RHS with a borrow flag. Returns |
1224 | the borrow flag. */ |
1225 | IEEEFloat::integerPart IEEEFloat::subtractSignificand(const IEEEFloat &rhs, |
1226 | integerPart borrow) { |
1227 | integerPart *parts; |
1228 | |
1229 | parts = significandParts(); |
1230 | |
1231 | assert(semantics == rhs.semantics); |
1232 | assert(exponent == rhs.exponent); |
1233 | |
1234 | return APInt::tcSubtract(parts, rhs.significandParts(), carry: borrow, |
1235 | partCount()); |
1236 | } |
1237 | |
1238 | /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it |
1239 | on to the full-precision result of the multiplication. Returns the |
1240 | lost fraction. */ |
1241 | lostFraction IEEEFloat::multiplySignificand(const IEEEFloat &rhs, |
1242 | IEEEFloat addend) { |
1243 | unsigned int omsb; // One, not zero, based MSB. |
1244 | unsigned int partsCount, newPartsCount, precision; |
1245 | integerPart *lhsSignificand; |
1246 | integerPart scratch[4]; |
1247 | integerPart *fullSignificand; |
1248 | lostFraction lost_fraction; |
1249 | bool ignored; |
1250 | |
1251 | assert(semantics == rhs.semantics); |
1252 | |
1253 | precision = semantics->precision; |
1254 | |
1255 | // Allocate space for twice as many bits as the original significand, plus one |
1256 | // extra bit for the addition to overflow into. |
1257 | newPartsCount = partCountForBits(bits: precision * 2 + 1); |
1258 | |
1259 | if (newPartsCount > 4) |
1260 | fullSignificand = new integerPart[newPartsCount]; |
1261 | else |
1262 | fullSignificand = scratch; |
1263 | |
1264 | lhsSignificand = significandParts(); |
1265 | partsCount = partCount(); |
1266 | |
1267 | APInt::tcFullMultiply(fullSignificand, lhsSignificand, |
1268 | rhs.significandParts(), partsCount, partsCount); |
1269 | |
1270 | lost_fraction = lfExactlyZero; |
1271 | omsb = APInt::tcMSB(parts: fullSignificand, n: newPartsCount) + 1; |
1272 | exponent += rhs.exponent; |
1273 | |
1274 | // Assume the operands involved in the multiplication are single-precision |
1275 | // FP, and the two multiplicants are: |
1276 | // *this = a23 . a22 ... a0 * 2^e1 |
1277 | // rhs = b23 . b22 ... b0 * 2^e2 |
1278 | // the result of multiplication is: |
1279 | // *this = c48 c47 c46 . c45 ... c0 * 2^(e1+e2) |
1280 | // Note that there are three significant bits at the left-hand side of the |
1281 | // radix point: two for the multiplication, and an overflow bit for the |
1282 | // addition (that will always be zero at this point). Move the radix point |
1283 | // toward left by two bits, and adjust exponent accordingly. |
1284 | exponent += 2; |
1285 | |
1286 | if (addend.isNonZero()) { |
1287 | // The intermediate result of the multiplication has "2 * precision" |
1288 | // signicant bit; adjust the addend to be consistent with mul result. |
1289 | // |
1290 | Significand savedSignificand = significand; |
1291 | const fltSemantics *savedSemantics = semantics; |
1292 | fltSemantics extendedSemantics; |
1293 | opStatus status; |
1294 | unsigned int extendedPrecision; |
1295 | |
1296 | // Normalize our MSB to one below the top bit to allow for overflow. |
1297 | extendedPrecision = 2 * precision + 1; |
1298 | if (omsb != extendedPrecision - 1) { |
1299 | assert(extendedPrecision > omsb); |
1300 | APInt::tcShiftLeft(fullSignificand, Words: newPartsCount, |
1301 | Count: (extendedPrecision - 1) - omsb); |
1302 | exponent -= (extendedPrecision - 1) - omsb; |
1303 | } |
1304 | |
1305 | /* Create new semantics. */ |
1306 | extendedSemantics = *semantics; |
1307 | extendedSemantics.precision = extendedPrecision; |
1308 | |
1309 | if (newPartsCount == 1) |
1310 | significand.part = fullSignificand[0]; |
1311 | else |
1312 | significand.parts = fullSignificand; |
1313 | semantics = &extendedSemantics; |
1314 | |
1315 | // Make a copy so we can convert it to the extended semantics. |
1316 | // Note that we cannot convert the addend directly, as the extendedSemantics |
1317 | // is a local variable (which we take a reference to). |
1318 | IEEEFloat extendedAddend(addend); |
1319 | status = extendedAddend.convert(extendedSemantics, rmTowardZero, &ignored); |
1320 | assert(status == opOK); |
1321 | (void)status; |
1322 | |
1323 | // Shift the significand of the addend right by one bit. This guarantees |
1324 | // that the high bit of the significand is zero (same as fullSignificand), |
1325 | // so the addition will overflow (if it does overflow at all) into the top bit. |
1326 | lost_fraction = extendedAddend.shiftSignificandRight(1); |
1327 | assert(lost_fraction == lfExactlyZero && |
1328 | "Lost precision while shifting addend for fused-multiply-add." ); |
1329 | |
1330 | lost_fraction = addOrSubtractSignificand(extendedAddend, subtract: false); |
1331 | |
1332 | /* Restore our state. */ |
1333 | if (newPartsCount == 1) |
1334 | fullSignificand[0] = significand.part; |
1335 | significand = savedSignificand; |
1336 | semantics = savedSemantics; |
1337 | |
1338 | omsb = APInt::tcMSB(parts: fullSignificand, n: newPartsCount) + 1; |
1339 | } |
1340 | |
1341 | // Convert the result having "2 * precision" significant-bits back to the one |
1342 | // having "precision" significant-bits. First, move the radix point from |
1343 | // poision "2*precision - 1" to "precision - 1". The exponent need to be |
1344 | // adjusted by "2*precision - 1" - "precision - 1" = "precision". |
1345 | exponent -= precision + 1; |
1346 | |
1347 | // In case MSB resides at the left-hand side of radix point, shift the |
1348 | // mantissa right by some amount to make sure the MSB reside right before |
1349 | // the radix point (i.e. "MSB . rest-significant-bits"). |
1350 | // |
1351 | // Note that the result is not normalized when "omsb < precision". So, the |
1352 | // caller needs to call IEEEFloat::normalize() if normalized value is |
1353 | // expected. |
1354 | if (omsb > precision) { |
1355 | unsigned int bits, significantParts; |
1356 | lostFraction lf; |
1357 | |
1358 | bits = omsb - precision; |
1359 | significantParts = partCountForBits(bits: omsb); |
1360 | lf = shiftRight(dst: fullSignificand, parts: significantParts, bits); |
1361 | lost_fraction = combineLostFractions(moreSignificant: lf, lessSignificant: lost_fraction); |
1362 | exponent += bits; |
1363 | } |
1364 | |
1365 | APInt::tcAssign(lhsSignificand, fullSignificand, partsCount); |
1366 | |
1367 | if (newPartsCount > 4) |
1368 | delete [] fullSignificand; |
1369 | |
1370 | return lost_fraction; |
1371 | } |
1372 | |
1373 | lostFraction IEEEFloat::multiplySignificand(const IEEEFloat &rhs) { |
1374 | return multiplySignificand(rhs, addend: IEEEFloat(*semantics)); |
1375 | } |
1376 | |
1377 | /* Multiply the significands of LHS and RHS to DST. */ |
1378 | lostFraction IEEEFloat::divideSignificand(const IEEEFloat &rhs) { |
1379 | unsigned int bit, i, partsCount; |
1380 | const integerPart *rhsSignificand; |
1381 | integerPart *lhsSignificand, *dividend, *divisor; |
1382 | integerPart scratch[4]; |
1383 | lostFraction lost_fraction; |
1384 | |
1385 | assert(semantics == rhs.semantics); |
1386 | |
1387 | lhsSignificand = significandParts(); |
1388 | rhsSignificand = rhs.significandParts(); |
1389 | partsCount = partCount(); |
1390 | |
1391 | if (partsCount > 2) |
1392 | dividend = new integerPart[partsCount * 2]; |
1393 | else |
1394 | dividend = scratch; |
1395 | |
1396 | divisor = dividend + partsCount; |
1397 | |
1398 | /* Copy the dividend and divisor as they will be modified in-place. */ |
1399 | for (i = 0; i < partsCount; i++) { |
1400 | dividend[i] = lhsSignificand[i]; |
1401 | divisor[i] = rhsSignificand[i]; |
1402 | lhsSignificand[i] = 0; |
1403 | } |
1404 | |
1405 | exponent -= rhs.exponent; |
1406 | |
1407 | unsigned int precision = semantics->precision; |
1408 | |
1409 | /* Normalize the divisor. */ |
1410 | bit = precision - APInt::tcMSB(parts: divisor, n: partsCount) - 1; |
1411 | if (bit) { |
1412 | exponent += bit; |
1413 | APInt::tcShiftLeft(divisor, Words: partsCount, Count: bit); |
1414 | } |
1415 | |
1416 | /* Normalize the dividend. */ |
1417 | bit = precision - APInt::tcMSB(parts: dividend, n: partsCount) - 1; |
1418 | if (bit) { |
1419 | exponent -= bit; |
1420 | APInt::tcShiftLeft(dividend, Words: partsCount, Count: bit); |
1421 | } |
1422 | |
1423 | /* Ensure the dividend >= divisor initially for the loop below. |
1424 | Incidentally, this means that the division loop below is |
1425 | guaranteed to set the integer bit to one. */ |
1426 | if (APInt::tcCompare(dividend, divisor, partsCount) < 0) { |
1427 | exponent--; |
1428 | APInt::tcShiftLeft(dividend, Words: partsCount, Count: 1); |
1429 | assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0); |
1430 | } |
1431 | |
1432 | /* Long division. */ |
1433 | for (bit = precision; bit; bit -= 1) { |
1434 | if (APInt::tcCompare(dividend, divisor, partsCount) >= 0) { |
1435 | APInt::tcSubtract(dividend, divisor, carry: 0, partsCount); |
1436 | APInt::tcSetBit(lhsSignificand, bit: bit - 1); |
1437 | } |
1438 | |
1439 | APInt::tcShiftLeft(dividend, Words: partsCount, Count: 1); |
1440 | } |
1441 | |
1442 | /* Figure out the lost fraction. */ |
1443 | int cmp = APInt::tcCompare(dividend, divisor, partsCount); |
1444 | |
1445 | if (cmp > 0) |
1446 | lost_fraction = lfMoreThanHalf; |
1447 | else if (cmp == 0) |
1448 | lost_fraction = lfExactlyHalf; |
1449 | else if (APInt::tcIsZero(dividend, partsCount)) |
1450 | lost_fraction = lfExactlyZero; |
1451 | else |
1452 | lost_fraction = lfLessThanHalf; |
1453 | |
1454 | if (partsCount > 2) |
1455 | delete [] dividend; |
1456 | |
1457 | return lost_fraction; |
1458 | } |
1459 | |
1460 | unsigned int IEEEFloat::significandMSB() const { |
1461 | return APInt::tcMSB(parts: significandParts(), n: partCount()); |
1462 | } |
1463 | |
1464 | unsigned int IEEEFloat::significandLSB() const { |
1465 | return APInt::tcLSB(significandParts(), n: partCount()); |
1466 | } |
1467 | |
1468 | /* Note that a zero result is NOT normalized to fcZero. */ |
1469 | lostFraction IEEEFloat::shiftSignificandRight(unsigned int bits) { |
1470 | /* Our exponent should not overflow. */ |
1471 | assert((ExponentType) (exponent + bits) >= exponent); |
1472 | |
1473 | exponent += bits; |
1474 | |
1475 | return shiftRight(dst: significandParts(), parts: partCount(), bits); |
1476 | } |
1477 | |
1478 | /* Shift the significand left BITS bits, subtract BITS from its exponent. */ |
1479 | void IEEEFloat::shiftSignificandLeft(unsigned int bits) { |
1480 | assert(bits < semantics->precision); |
1481 | |
1482 | if (bits) { |
1483 | unsigned int partsCount = partCount(); |
1484 | |
1485 | APInt::tcShiftLeft(significandParts(), Words: partsCount, Count: bits); |
1486 | exponent -= bits; |
1487 | |
1488 | assert(!APInt::tcIsZero(significandParts(), partsCount)); |
1489 | } |
1490 | } |
1491 | |
1492 | IEEEFloat::cmpResult |
1493 | IEEEFloat::compareAbsoluteValue(const IEEEFloat &rhs) const { |
1494 | int compare; |
1495 | |
1496 | assert(semantics == rhs.semantics); |
1497 | assert(isFiniteNonZero()); |
1498 | assert(rhs.isFiniteNonZero()); |
1499 | |
1500 | compare = exponent - rhs.exponent; |
1501 | |
1502 | /* If exponents are equal, do an unsigned bignum comparison of the |
1503 | significands. */ |
1504 | if (compare == 0) |
1505 | compare = APInt::tcCompare(significandParts(), rhs.significandParts(), |
1506 | partCount()); |
1507 | |
1508 | if (compare > 0) |
1509 | return cmpGreaterThan; |
1510 | else if (compare < 0) |
1511 | return cmpLessThan; |
1512 | else |
1513 | return cmpEqual; |
1514 | } |
1515 | |
1516 | /* Set the least significant BITS bits of a bignum, clear the |
1517 | rest. */ |
1518 | static void tcSetLeastSignificantBits(APInt::WordType *dst, unsigned parts, |
1519 | unsigned bits) { |
1520 | unsigned i = 0; |
1521 | while (bits > APInt::APINT_BITS_PER_WORD) { |
1522 | dst[i++] = ~(APInt::WordType)0; |
1523 | bits -= APInt::APINT_BITS_PER_WORD; |
1524 | } |
1525 | |
1526 | if (bits) |
1527 | dst[i++] = ~(APInt::WordType)0 >> (APInt::APINT_BITS_PER_WORD - bits); |
1528 | |
1529 | while (i < parts) |
1530 | dst[i++] = 0; |
1531 | } |
1532 | |
1533 | /* Handle overflow. Sign is preserved. We either become infinity or |
1534 | the largest finite number. */ |
1535 | IEEEFloat::opStatus IEEEFloat::handleOverflow(roundingMode rounding_mode) { |
1536 | if (semantics->nonFiniteBehavior != fltNonfiniteBehavior::FiniteOnly) { |
1537 | /* Infinity? */ |
1538 | if (rounding_mode == rmNearestTiesToEven || |
1539 | rounding_mode == rmNearestTiesToAway || |
1540 | (rounding_mode == rmTowardPositive && !sign) || |
1541 | (rounding_mode == rmTowardNegative && sign)) { |
1542 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) |
1543 | makeNaN(SNaN: false, Negative: sign); |
1544 | else |
1545 | category = fcInfinity; |
1546 | return static_cast<opStatus>(opOverflow | opInexact); |
1547 | } |
1548 | } |
1549 | |
1550 | /* Otherwise we become the largest finite number. */ |
1551 | category = fcNormal; |
1552 | exponent = semantics->maxExponent; |
1553 | tcSetLeastSignificantBits(dst: significandParts(), parts: partCount(), |
1554 | bits: semantics->precision); |
1555 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly && |
1556 | semantics->nanEncoding == fltNanEncoding::AllOnes) |
1557 | APInt::tcClearBit(significandParts(), bit: 0); |
1558 | |
1559 | return opInexact; |
1560 | } |
1561 | |
1562 | /* Returns TRUE if, when truncating the current number, with BIT the |
1563 | new LSB, with the given lost fraction and rounding mode, the result |
1564 | would need to be rounded away from zero (i.e., by increasing the |
1565 | signficand). This routine must work for fcZero of both signs, and |
1566 | fcNormal numbers. */ |
1567 | bool IEEEFloat::roundAwayFromZero(roundingMode rounding_mode, |
1568 | lostFraction lost_fraction, |
1569 | unsigned int bit) const { |
1570 | /* NaNs and infinities should not have lost fractions. */ |
1571 | assert(isFiniteNonZero() || category == fcZero); |
1572 | |
1573 | /* Current callers never pass this so we don't handle it. */ |
1574 | assert(lost_fraction != lfExactlyZero); |
1575 | |
1576 | switch (rounding_mode) { |
1577 | case rmNearestTiesToAway: |
1578 | return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf; |
1579 | |
1580 | case rmNearestTiesToEven: |
1581 | if (lost_fraction == lfMoreThanHalf) |
1582 | return true; |
1583 | |
1584 | /* Our zeroes don't have a significand to test. */ |
1585 | if (lost_fraction == lfExactlyHalf && category != fcZero) |
1586 | return APInt::tcExtractBit(significandParts(), bit); |
1587 | |
1588 | return false; |
1589 | |
1590 | case rmTowardZero: |
1591 | return false; |
1592 | |
1593 | case rmTowardPositive: |
1594 | return !sign; |
1595 | |
1596 | case rmTowardNegative: |
1597 | return sign; |
1598 | |
1599 | default: |
1600 | break; |
1601 | } |
1602 | llvm_unreachable("Invalid rounding mode found" ); |
1603 | } |
1604 | |
1605 | IEEEFloat::opStatus IEEEFloat::normalize(roundingMode rounding_mode, |
1606 | lostFraction lost_fraction) { |
1607 | unsigned int omsb; /* One, not zero, based MSB. */ |
1608 | int exponentChange; |
1609 | |
1610 | if (!isFiniteNonZero()) |
1611 | return opOK; |
1612 | |
1613 | /* Before rounding normalize the exponent of fcNormal numbers. */ |
1614 | omsb = significandMSB() + 1; |
1615 | |
1616 | if (omsb) { |
1617 | /* OMSB is numbered from 1. We want to place it in the integer |
1618 | bit numbered PRECISION if possible, with a compensating change in |
1619 | the exponent. */ |
1620 | exponentChange = omsb - semantics->precision; |
1621 | |
1622 | /* If the resulting exponent is too high, overflow according to |
1623 | the rounding mode. */ |
1624 | if (exponent + exponentChange > semantics->maxExponent) |
1625 | return handleOverflow(rounding_mode); |
1626 | |
1627 | /* Subnormal numbers have exponent minExponent, and their MSB |
1628 | is forced based on that. */ |
1629 | if (exponent + exponentChange < semantics->minExponent) |
1630 | exponentChange = semantics->minExponent - exponent; |
1631 | |
1632 | /* Shifting left is easy as we don't lose precision. */ |
1633 | if (exponentChange < 0) { |
1634 | assert(lost_fraction == lfExactlyZero); |
1635 | |
1636 | shiftSignificandLeft(bits: -exponentChange); |
1637 | |
1638 | return opOK; |
1639 | } |
1640 | |
1641 | if (exponentChange > 0) { |
1642 | lostFraction lf; |
1643 | |
1644 | /* Shift right and capture any new lost fraction. */ |
1645 | lf = shiftSignificandRight(bits: exponentChange); |
1646 | |
1647 | lost_fraction = combineLostFractions(moreSignificant: lf, lessSignificant: lost_fraction); |
1648 | |
1649 | /* Keep OMSB up-to-date. */ |
1650 | if (omsb > (unsigned) exponentChange) |
1651 | omsb -= exponentChange; |
1652 | else |
1653 | omsb = 0; |
1654 | } |
1655 | } |
1656 | |
1657 | // The all-ones values is an overflow if NaN is all ones. If NaN is |
1658 | // represented by negative zero, then it is a valid finite value. |
1659 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly && |
1660 | semantics->nanEncoding == fltNanEncoding::AllOnes && |
1661 | exponent == semantics->maxExponent && isSignificandAllOnes()) |
1662 | return handleOverflow(rounding_mode); |
1663 | |
1664 | /* Now round the number according to rounding_mode given the lost |
1665 | fraction. */ |
1666 | |
1667 | /* As specified in IEEE 754, since we do not trap we do not report |
1668 | underflow for exact results. */ |
1669 | if (lost_fraction == lfExactlyZero) { |
1670 | /* Canonicalize zeroes. */ |
1671 | if (omsb == 0) { |
1672 | category = fcZero; |
1673 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
1674 | sign = false; |
1675 | } |
1676 | |
1677 | return opOK; |
1678 | } |
1679 | |
1680 | /* Increment the significand if we're rounding away from zero. */ |
1681 | if (roundAwayFromZero(rounding_mode, lost_fraction, bit: 0)) { |
1682 | if (omsb == 0) |
1683 | exponent = semantics->minExponent; |
1684 | |
1685 | incrementSignificand(); |
1686 | omsb = significandMSB() + 1; |
1687 | |
1688 | /* Did the significand increment overflow? */ |
1689 | if (omsb == (unsigned) semantics->precision + 1) { |
1690 | /* Renormalize by incrementing the exponent and shifting our |
1691 | significand right one. However if we already have the |
1692 | maximum exponent we overflow to infinity. */ |
1693 | if (exponent == semantics->maxExponent) |
1694 | // Invoke overflow handling with a rounding mode that will guarantee |
1695 | // that the result gets turned into the correct infinity representation. |
1696 | // This is needed instead of just setting the category to infinity to |
1697 | // account for 8-bit floating point types that have no inf, only NaN. |
1698 | return handleOverflow(rounding_mode: sign ? rmTowardNegative : rmTowardPositive); |
1699 | |
1700 | shiftSignificandRight(bits: 1); |
1701 | |
1702 | return opInexact; |
1703 | } |
1704 | |
1705 | // The all-ones values is an overflow if NaN is all ones. If NaN is |
1706 | // represented by negative zero, then it is a valid finite value. |
1707 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly && |
1708 | semantics->nanEncoding == fltNanEncoding::AllOnes && |
1709 | exponent == semantics->maxExponent && isSignificandAllOnes()) |
1710 | return handleOverflow(rounding_mode); |
1711 | } |
1712 | |
1713 | /* The normal case - we were and are not denormal, and any |
1714 | significand increment above didn't overflow. */ |
1715 | if (omsb == semantics->precision) |
1716 | return opInexact; |
1717 | |
1718 | /* We have a non-zero denormal. */ |
1719 | assert(omsb < semantics->precision); |
1720 | |
1721 | /* Canonicalize zeroes. */ |
1722 | if (omsb == 0) { |
1723 | category = fcZero; |
1724 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
1725 | sign = false; |
1726 | } |
1727 | |
1728 | /* The fcZero case is a denormal that underflowed to zero. */ |
1729 | return (opStatus) (opUnderflow | opInexact); |
1730 | } |
1731 | |
1732 | IEEEFloat::opStatus IEEEFloat::addOrSubtractSpecials(const IEEEFloat &rhs, |
1733 | bool subtract) { |
1734 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1735 | default: |
1736 | llvm_unreachable(nullptr); |
1737 | |
1738 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1739 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1740 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1741 | assign(rhs); |
1742 | [[fallthrough]]; |
1743 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1744 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1745 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1746 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1747 | if (isSignaling()) { |
1748 | makeQuiet(); |
1749 | return opInvalidOp; |
1750 | } |
1751 | return rhs.isSignaling() ? opInvalidOp : opOK; |
1752 | |
1753 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1754 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1755 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1756 | return opOK; |
1757 | |
1758 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1759 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1760 | category = fcInfinity; |
1761 | sign = rhs.sign ^ subtract; |
1762 | return opOK; |
1763 | |
1764 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1765 | assign(rhs); |
1766 | sign = rhs.sign ^ subtract; |
1767 | return opOK; |
1768 | |
1769 | case PackCategoriesIntoKey(fcZero, fcZero): |
1770 | /* Sign depends on rounding mode; handled by caller. */ |
1771 | return opOK; |
1772 | |
1773 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1774 | /* Differently signed infinities can only be validly |
1775 | subtracted. */ |
1776 | if (((sign ^ rhs.sign)!=0) != subtract) { |
1777 | makeNaN(); |
1778 | return opInvalidOp; |
1779 | } |
1780 | |
1781 | return opOK; |
1782 | |
1783 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1784 | return opDivByZero; |
1785 | } |
1786 | } |
1787 | |
1788 | /* Add or subtract two normal numbers. */ |
1789 | lostFraction IEEEFloat::addOrSubtractSignificand(const IEEEFloat &rhs, |
1790 | bool subtract) { |
1791 | integerPart carry; |
1792 | lostFraction lost_fraction; |
1793 | int bits; |
1794 | |
1795 | /* Determine if the operation on the absolute values is effectively |
1796 | an addition or subtraction. */ |
1797 | subtract ^= static_cast<bool>(sign ^ rhs.sign); |
1798 | |
1799 | /* Are we bigger exponent-wise than the RHS? */ |
1800 | bits = exponent - rhs.exponent; |
1801 | |
1802 | /* Subtraction is more subtle than one might naively expect. */ |
1803 | if (subtract) { |
1804 | IEEEFloat temp_rhs(rhs); |
1805 | |
1806 | if (bits == 0) |
1807 | lost_fraction = lfExactlyZero; |
1808 | else if (bits > 0) { |
1809 | lost_fraction = temp_rhs.shiftSignificandRight(bits: bits - 1); |
1810 | shiftSignificandLeft(bits: 1); |
1811 | } else { |
1812 | lost_fraction = shiftSignificandRight(bits: -bits - 1); |
1813 | temp_rhs.shiftSignificandLeft(bits: 1); |
1814 | } |
1815 | |
1816 | // Should we reverse the subtraction. |
1817 | if (compareAbsoluteValue(rhs: temp_rhs) == cmpLessThan) { |
1818 | carry = temp_rhs.subtractSignificand |
1819 | (rhs: *this, borrow: lost_fraction != lfExactlyZero); |
1820 | copySignificand(rhs: temp_rhs); |
1821 | sign = !sign; |
1822 | } else { |
1823 | carry = subtractSignificand |
1824 | (rhs: temp_rhs, borrow: lost_fraction != lfExactlyZero); |
1825 | } |
1826 | |
1827 | /* Invert the lost fraction - it was on the RHS and |
1828 | subtracted. */ |
1829 | if (lost_fraction == lfLessThanHalf) |
1830 | lost_fraction = lfMoreThanHalf; |
1831 | else if (lost_fraction == lfMoreThanHalf) |
1832 | lost_fraction = lfLessThanHalf; |
1833 | |
1834 | /* The code above is intended to ensure that no borrow is |
1835 | necessary. */ |
1836 | assert(!carry); |
1837 | (void)carry; |
1838 | } else { |
1839 | if (bits > 0) { |
1840 | IEEEFloat temp_rhs(rhs); |
1841 | |
1842 | lost_fraction = temp_rhs.shiftSignificandRight(bits); |
1843 | carry = addSignificand(rhs: temp_rhs); |
1844 | } else { |
1845 | lost_fraction = shiftSignificandRight(bits: -bits); |
1846 | carry = addSignificand(rhs); |
1847 | } |
1848 | |
1849 | /* We have a guard bit; generating a carry cannot happen. */ |
1850 | assert(!carry); |
1851 | (void)carry; |
1852 | } |
1853 | |
1854 | return lost_fraction; |
1855 | } |
1856 | |
1857 | IEEEFloat::opStatus IEEEFloat::multiplySpecials(const IEEEFloat &rhs) { |
1858 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1859 | default: |
1860 | llvm_unreachable(nullptr); |
1861 | |
1862 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1863 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1864 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1865 | assign(rhs); |
1866 | sign = false; |
1867 | [[fallthrough]]; |
1868 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1869 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1870 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1871 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1872 | sign ^= rhs.sign; // restore the original sign |
1873 | if (isSignaling()) { |
1874 | makeQuiet(); |
1875 | return opInvalidOp; |
1876 | } |
1877 | return rhs.isSignaling() ? opInvalidOp : opOK; |
1878 | |
1879 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1880 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1881 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1882 | category = fcInfinity; |
1883 | return opOK; |
1884 | |
1885 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1886 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1887 | case PackCategoriesIntoKey(fcZero, fcZero): |
1888 | category = fcZero; |
1889 | return opOK; |
1890 | |
1891 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1892 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1893 | makeNaN(); |
1894 | return opInvalidOp; |
1895 | |
1896 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1897 | return opOK; |
1898 | } |
1899 | } |
1900 | |
1901 | IEEEFloat::opStatus IEEEFloat::divideSpecials(const IEEEFloat &rhs) { |
1902 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1903 | default: |
1904 | llvm_unreachable(nullptr); |
1905 | |
1906 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1907 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1908 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1909 | assign(rhs); |
1910 | sign = false; |
1911 | [[fallthrough]]; |
1912 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1913 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1914 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1915 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1916 | sign ^= rhs.sign; // restore the original sign |
1917 | if (isSignaling()) { |
1918 | makeQuiet(); |
1919 | return opInvalidOp; |
1920 | } |
1921 | return rhs.isSignaling() ? opInvalidOp : opOK; |
1922 | |
1923 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1924 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1925 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1926 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1927 | return opOK; |
1928 | |
1929 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1930 | category = fcZero; |
1931 | return opOK; |
1932 | |
1933 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1934 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) |
1935 | makeNaN(SNaN: false, Negative: sign); |
1936 | else |
1937 | category = fcInfinity; |
1938 | return opDivByZero; |
1939 | |
1940 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1941 | case PackCategoriesIntoKey(fcZero, fcZero): |
1942 | makeNaN(); |
1943 | return opInvalidOp; |
1944 | |
1945 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1946 | return opOK; |
1947 | } |
1948 | } |
1949 | |
1950 | IEEEFloat::opStatus IEEEFloat::modSpecials(const IEEEFloat &rhs) { |
1951 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1952 | default: |
1953 | llvm_unreachable(nullptr); |
1954 | |
1955 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1956 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1957 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1958 | assign(rhs); |
1959 | [[fallthrough]]; |
1960 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1961 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1962 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1963 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1964 | if (isSignaling()) { |
1965 | makeQuiet(); |
1966 | return opInvalidOp; |
1967 | } |
1968 | return rhs.isSignaling() ? opInvalidOp : opOK; |
1969 | |
1970 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1971 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1972 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1973 | return opOK; |
1974 | |
1975 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1976 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1977 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1978 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1979 | case PackCategoriesIntoKey(fcZero, fcZero): |
1980 | makeNaN(); |
1981 | return opInvalidOp; |
1982 | |
1983 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1984 | return opOK; |
1985 | } |
1986 | } |
1987 | |
1988 | IEEEFloat::opStatus IEEEFloat::remainderSpecials(const IEEEFloat &rhs) { |
1989 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1990 | default: |
1991 | llvm_unreachable(nullptr); |
1992 | |
1993 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1994 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1995 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1996 | assign(rhs); |
1997 | [[fallthrough]]; |
1998 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1999 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
2000 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
2001 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
2002 | if (isSignaling()) { |
2003 | makeQuiet(); |
2004 | return opInvalidOp; |
2005 | } |
2006 | return rhs.isSignaling() ? opInvalidOp : opOK; |
2007 | |
2008 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
2009 | case PackCategoriesIntoKey(fcZero, fcNormal): |
2010 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
2011 | return opOK; |
2012 | |
2013 | case PackCategoriesIntoKey(fcNormal, fcZero): |
2014 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
2015 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
2016 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
2017 | case PackCategoriesIntoKey(fcZero, fcZero): |
2018 | makeNaN(); |
2019 | return opInvalidOp; |
2020 | |
2021 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
2022 | return opDivByZero; // fake status, indicating this is not a special case |
2023 | } |
2024 | } |
2025 | |
2026 | /* Change sign. */ |
2027 | void IEEEFloat::changeSign() { |
2028 | // With NaN-as-negative-zero, neither NaN or negative zero can change |
2029 | // their signs. |
2030 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero && |
2031 | (isZero() || isNaN())) |
2032 | return; |
2033 | /* Look mummy, this one's easy. */ |
2034 | sign = !sign; |
2035 | } |
2036 | |
2037 | /* Normalized addition or subtraction. */ |
2038 | IEEEFloat::opStatus IEEEFloat::addOrSubtract(const IEEEFloat &rhs, |
2039 | roundingMode rounding_mode, |
2040 | bool subtract) { |
2041 | opStatus fs; |
2042 | |
2043 | fs = addOrSubtractSpecials(rhs, subtract); |
2044 | |
2045 | /* This return code means it was not a simple case. */ |
2046 | if (fs == opDivByZero) { |
2047 | lostFraction lost_fraction; |
2048 | |
2049 | lost_fraction = addOrSubtractSignificand(rhs, subtract); |
2050 | fs = normalize(rounding_mode, lost_fraction); |
2051 | |
2052 | /* Can only be zero if we lost no fraction. */ |
2053 | assert(category != fcZero || lost_fraction == lfExactlyZero); |
2054 | } |
2055 | |
2056 | /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a |
2057 | positive zero unless rounding to minus infinity, except that |
2058 | adding two like-signed zeroes gives that zero. */ |
2059 | if (category == fcZero) { |
2060 | if (rhs.category != fcZero || (sign == rhs.sign) == subtract) |
2061 | sign = (rounding_mode == rmTowardNegative); |
2062 | // NaN-in-negative-zero means zeros need to be normalized to +0. |
2063 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2064 | sign = false; |
2065 | } |
2066 | |
2067 | return fs; |
2068 | } |
2069 | |
2070 | /* Normalized addition. */ |
2071 | IEEEFloat::opStatus IEEEFloat::add(const IEEEFloat &rhs, |
2072 | roundingMode rounding_mode) { |
2073 | return addOrSubtract(rhs, rounding_mode, subtract: false); |
2074 | } |
2075 | |
2076 | /* Normalized subtraction. */ |
2077 | IEEEFloat::opStatus IEEEFloat::subtract(const IEEEFloat &rhs, |
2078 | roundingMode rounding_mode) { |
2079 | return addOrSubtract(rhs, rounding_mode, subtract: true); |
2080 | } |
2081 | |
2082 | /* Normalized multiply. */ |
2083 | IEEEFloat::opStatus IEEEFloat::multiply(const IEEEFloat &rhs, |
2084 | roundingMode rounding_mode) { |
2085 | opStatus fs; |
2086 | |
2087 | sign ^= rhs.sign; |
2088 | fs = multiplySpecials(rhs); |
2089 | |
2090 | if (isZero() && semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2091 | sign = false; |
2092 | if (isFiniteNonZero()) { |
2093 | lostFraction lost_fraction = multiplySignificand(rhs); |
2094 | fs = normalize(rounding_mode, lost_fraction); |
2095 | if (lost_fraction != lfExactlyZero) |
2096 | fs = (opStatus) (fs | opInexact); |
2097 | } |
2098 | |
2099 | return fs; |
2100 | } |
2101 | |
2102 | /* Normalized divide. */ |
2103 | IEEEFloat::opStatus IEEEFloat::divide(const IEEEFloat &rhs, |
2104 | roundingMode rounding_mode) { |
2105 | opStatus fs; |
2106 | |
2107 | sign ^= rhs.sign; |
2108 | fs = divideSpecials(rhs); |
2109 | |
2110 | if (isZero() && semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2111 | sign = false; |
2112 | if (isFiniteNonZero()) { |
2113 | lostFraction lost_fraction = divideSignificand(rhs); |
2114 | fs = normalize(rounding_mode, lost_fraction); |
2115 | if (lost_fraction != lfExactlyZero) |
2116 | fs = (opStatus) (fs | opInexact); |
2117 | } |
2118 | |
2119 | return fs; |
2120 | } |
2121 | |
2122 | /* Normalized remainder. */ |
2123 | IEEEFloat::opStatus IEEEFloat::remainder(const IEEEFloat &rhs) { |
2124 | opStatus fs; |
2125 | unsigned int origSign = sign; |
2126 | |
2127 | // First handle the special cases. |
2128 | fs = remainderSpecials(rhs); |
2129 | if (fs != opDivByZero) |
2130 | return fs; |
2131 | |
2132 | fs = opOK; |
2133 | |
2134 | // Make sure the current value is less than twice the denom. If the addition |
2135 | // did not succeed (an overflow has happened), which means that the finite |
2136 | // value we currently posses must be less than twice the denom (as we are |
2137 | // using the same semantics). |
2138 | IEEEFloat P2 = rhs; |
2139 | if (P2.add(rhs, rounding_mode: rmNearestTiesToEven) == opOK) { |
2140 | fs = mod(P2); |
2141 | assert(fs == opOK); |
2142 | } |
2143 | |
2144 | // Lets work with absolute numbers. |
2145 | IEEEFloat P = rhs; |
2146 | P.sign = false; |
2147 | sign = false; |
2148 | |
2149 | // |
2150 | // To calculate the remainder we use the following scheme. |
2151 | // |
2152 | // The remainder is defained as follows: |
2153 | // |
2154 | // remainder = numer - rquot * denom = x - r * p |
2155 | // |
2156 | // Where r is the result of: x/p, rounded toward the nearest integral value |
2157 | // (with halfway cases rounded toward the even number). |
2158 | // |
2159 | // Currently, (after x mod 2p): |
2160 | // r is the number of 2p's present inside x, which is inherently, an even |
2161 | // number of p's. |
2162 | // |
2163 | // We may split the remaining calculation into 4 options: |
2164 | // - if x < 0.5p then we round to the nearest number with is 0, and are done. |
2165 | // - if x == 0.5p then we round to the nearest even number which is 0, and we |
2166 | // are done as well. |
2167 | // - if 0.5p < x < p then we round to nearest number which is 1, and we have |
2168 | // to subtract 1p at least once. |
2169 | // - if x >= p then we must subtract p at least once, as x must be a |
2170 | // remainder. |
2171 | // |
2172 | // By now, we were done, or we added 1 to r, which in turn, now an odd number. |
2173 | // |
2174 | // We can now split the remaining calculation to the following 3 options: |
2175 | // - if x < 0.5p then we round to the nearest number with is 0, and are done. |
2176 | // - if x == 0.5p then we round to the nearest even number. As r is odd, we |
2177 | // must round up to the next even number. so we must subtract p once more. |
2178 | // - if x > 0.5p (and inherently x < p) then we must round r up to the next |
2179 | // integral, and subtract p once more. |
2180 | // |
2181 | |
2182 | // Extend the semantics to prevent an overflow/underflow or inexact result. |
2183 | bool losesInfo; |
2184 | fltSemantics extendedSemantics = *semantics; |
2185 | extendedSemantics.maxExponent++; |
2186 | extendedSemantics.minExponent--; |
2187 | extendedSemantics.precision += 2; |
2188 | |
2189 | IEEEFloat VEx = *this; |
2190 | fs = VEx.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo); |
2191 | assert(fs == opOK && !losesInfo); |
2192 | IEEEFloat PEx = P; |
2193 | fs = PEx.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo); |
2194 | assert(fs == opOK && !losesInfo); |
2195 | |
2196 | // It is simpler to work with 2x instead of 0.5p, and we do not need to lose |
2197 | // any fraction. |
2198 | fs = VEx.add(rhs: VEx, rounding_mode: rmNearestTiesToEven); |
2199 | assert(fs == opOK); |
2200 | |
2201 | if (VEx.compare(PEx) == cmpGreaterThan) { |
2202 | fs = subtract(rhs: P, rounding_mode: rmNearestTiesToEven); |
2203 | assert(fs == opOK); |
2204 | |
2205 | // Make VEx = this.add(this), but because we have different semantics, we do |
2206 | // not want to `convert` again, so we just subtract PEx twice (which equals |
2207 | // to the desired value). |
2208 | fs = VEx.subtract(rhs: PEx, rounding_mode: rmNearestTiesToEven); |
2209 | assert(fs == opOK); |
2210 | fs = VEx.subtract(rhs: PEx, rounding_mode: rmNearestTiesToEven); |
2211 | assert(fs == opOK); |
2212 | |
2213 | cmpResult result = VEx.compare(PEx); |
2214 | if (result == cmpGreaterThan || result == cmpEqual) { |
2215 | fs = subtract(rhs: P, rounding_mode: rmNearestTiesToEven); |
2216 | assert(fs == opOK); |
2217 | } |
2218 | } |
2219 | |
2220 | if (isZero()) { |
2221 | sign = origSign; // IEEE754 requires this |
2222 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2223 | // But some 8-bit floats only have positive 0. |
2224 | sign = false; |
2225 | } |
2226 | |
2227 | else |
2228 | sign ^= origSign; |
2229 | return fs; |
2230 | } |
2231 | |
2232 | /* Normalized llvm frem (C fmod). */ |
2233 | IEEEFloat::opStatus IEEEFloat::mod(const IEEEFloat &rhs) { |
2234 | opStatus fs; |
2235 | fs = modSpecials(rhs); |
2236 | unsigned int origSign = sign; |
2237 | |
2238 | while (isFiniteNonZero() && rhs.isFiniteNonZero() && |
2239 | compareAbsoluteValue(rhs) != cmpLessThan) { |
2240 | int Exp = ilogb(Arg: *this) - ilogb(Arg: rhs); |
2241 | IEEEFloat V = scalbn(X: rhs, Exp, rmNearestTiesToEven); |
2242 | // V can overflow to NaN with fltNonfiniteBehavior::NanOnly, so explicitly |
2243 | // check for it. |
2244 | if (V.isNaN() || compareAbsoluteValue(rhs: V) == cmpLessThan) |
2245 | V = scalbn(X: rhs, Exp: Exp - 1, rmNearestTiesToEven); |
2246 | V.sign = sign; |
2247 | |
2248 | fs = subtract(rhs: V, rounding_mode: rmNearestTiesToEven); |
2249 | assert(fs==opOK); |
2250 | } |
2251 | if (isZero()) { |
2252 | sign = origSign; // fmod requires this |
2253 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2254 | sign = false; |
2255 | } |
2256 | return fs; |
2257 | } |
2258 | |
2259 | /* Normalized fused-multiply-add. */ |
2260 | IEEEFloat::opStatus IEEEFloat::fusedMultiplyAdd(const IEEEFloat &multiplicand, |
2261 | const IEEEFloat &addend, |
2262 | roundingMode rounding_mode) { |
2263 | opStatus fs; |
2264 | |
2265 | /* Post-multiplication sign, before addition. */ |
2266 | sign ^= multiplicand.sign; |
2267 | |
2268 | /* If and only if all arguments are normal do we need to do an |
2269 | extended-precision calculation. */ |
2270 | if (isFiniteNonZero() && |
2271 | multiplicand.isFiniteNonZero() && |
2272 | addend.isFinite()) { |
2273 | lostFraction lost_fraction; |
2274 | |
2275 | lost_fraction = multiplySignificand(rhs: multiplicand, addend); |
2276 | fs = normalize(rounding_mode, lost_fraction); |
2277 | if (lost_fraction != lfExactlyZero) |
2278 | fs = (opStatus) (fs | opInexact); |
2279 | |
2280 | /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a |
2281 | positive zero unless rounding to minus infinity, except that |
2282 | adding two like-signed zeroes gives that zero. */ |
2283 | if (category == fcZero && !(fs & opUnderflow) && sign != addend.sign) { |
2284 | sign = (rounding_mode == rmTowardNegative); |
2285 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
2286 | sign = false; |
2287 | } |
2288 | } else { |
2289 | fs = multiplySpecials(rhs: multiplicand); |
2290 | |
2291 | /* FS can only be opOK or opInvalidOp. There is no more work |
2292 | to do in the latter case. The IEEE-754R standard says it is |
2293 | implementation-defined in this case whether, if ADDEND is a |
2294 | quiet NaN, we raise invalid op; this implementation does so. |
2295 | |
2296 | If we need to do the addition we can do so with normal |
2297 | precision. */ |
2298 | if (fs == opOK) |
2299 | fs = addOrSubtract(rhs: addend, rounding_mode, subtract: false); |
2300 | } |
2301 | |
2302 | return fs; |
2303 | } |
2304 | |
2305 | /* Rounding-mode correct round to integral value. */ |
2306 | IEEEFloat::opStatus IEEEFloat::roundToIntegral(roundingMode rounding_mode) { |
2307 | opStatus fs; |
2308 | |
2309 | if (isInfinity()) |
2310 | // [IEEE Std 754-2008 6.1]: |
2311 | // The behavior of infinity in floating-point arithmetic is derived from the |
2312 | // limiting cases of real arithmetic with operands of arbitrarily |
2313 | // large magnitude, when such a limit exists. |
2314 | // ... |
2315 | // Operations on infinite operands are usually exact and therefore signal no |
2316 | // exceptions ... |
2317 | return opOK; |
2318 | |
2319 | if (isNaN()) { |
2320 | if (isSignaling()) { |
2321 | // [IEEE Std 754-2008 6.2]: |
2322 | // Under default exception handling, any operation signaling an invalid |
2323 | // operation exception and for which a floating-point result is to be |
2324 | // delivered shall deliver a quiet NaN. |
2325 | makeQuiet(); |
2326 | // [IEEE Std 754-2008 6.2]: |
2327 | // Signaling NaNs shall be reserved operands that, under default exception |
2328 | // handling, signal the invalid operation exception(see 7.2) for every |
2329 | // general-computational and signaling-computational operation except for |
2330 | // the conversions described in 5.12. |
2331 | return opInvalidOp; |
2332 | } else { |
2333 | // [IEEE Std 754-2008 6.2]: |
2334 | // For an operation with quiet NaN inputs, other than maximum and minimum |
2335 | // operations, if a floating-point result is to be delivered the result |
2336 | // shall be a quiet NaN which should be one of the input NaNs. |
2337 | // ... |
2338 | // Every general-computational and quiet-computational operation involving |
2339 | // one or more input NaNs, none of them signaling, shall signal no |
2340 | // exception, except fusedMultiplyAdd might signal the invalid operation |
2341 | // exception(see 7.2). |
2342 | return opOK; |
2343 | } |
2344 | } |
2345 | |
2346 | if (isZero()) { |
2347 | // [IEEE Std 754-2008 6.3]: |
2348 | // ... the sign of the result of conversions, the quantize operation, the |
2349 | // roundToIntegral operations, and the roundToIntegralExact(see 5.3.1) is |
2350 | // the sign of the first or only operand. |
2351 | return opOK; |
2352 | } |
2353 | |
2354 | // If the exponent is large enough, we know that this value is already |
2355 | // integral, and the arithmetic below would potentially cause it to saturate |
2356 | // to +/-Inf. Bail out early instead. |
2357 | if (exponent+1 >= (int)semanticsPrecision(semantics: *semantics)) |
2358 | return opOK; |
2359 | |
2360 | // The algorithm here is quite simple: we add 2^(p-1), where p is the |
2361 | // precision of our format, and then subtract it back off again. The choice |
2362 | // of rounding modes for the addition/subtraction determines the rounding mode |
2363 | // for our integral rounding as well. |
2364 | // NOTE: When the input value is negative, we do subtraction followed by |
2365 | // addition instead. |
2366 | APInt IntegerConstant(NextPowerOf2(A: semanticsPrecision(semantics: *semantics)), 1); |
2367 | IntegerConstant <<= semanticsPrecision(semantics: *semantics)-1; |
2368 | IEEEFloat MagicConstant(*semantics); |
2369 | fs = MagicConstant.convertFromAPInt(IntegerConstant, false, |
2370 | rmNearestTiesToEven); |
2371 | assert(fs == opOK); |
2372 | MagicConstant.sign = sign; |
2373 | |
2374 | // Preserve the input sign so that we can handle the case of zero result |
2375 | // correctly. |
2376 | bool inputSign = isNegative(); |
2377 | |
2378 | fs = add(rhs: MagicConstant, rounding_mode); |
2379 | |
2380 | // Current value and 'MagicConstant' are both integers, so the result of the |
2381 | // subtraction is always exact according to Sterbenz' lemma. |
2382 | subtract(rhs: MagicConstant, rounding_mode); |
2383 | |
2384 | // Restore the input sign. |
2385 | if (inputSign != isNegative()) |
2386 | changeSign(); |
2387 | |
2388 | return fs; |
2389 | } |
2390 | |
2391 | |
2392 | /* Comparison requires normalized numbers. */ |
2393 | IEEEFloat::cmpResult IEEEFloat::compare(const IEEEFloat &rhs) const { |
2394 | cmpResult result; |
2395 | |
2396 | assert(semantics == rhs.semantics); |
2397 | |
2398 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
2399 | default: |
2400 | llvm_unreachable(nullptr); |
2401 | |
2402 | case PackCategoriesIntoKey(fcNaN, fcZero): |
2403 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
2404 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
2405 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
2406 | case PackCategoriesIntoKey(fcZero, fcNaN): |
2407 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
2408 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
2409 | return cmpUnordered; |
2410 | |
2411 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
2412 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
2413 | case PackCategoriesIntoKey(fcNormal, fcZero): |
2414 | if (sign) |
2415 | return cmpLessThan; |
2416 | else |
2417 | return cmpGreaterThan; |
2418 | |
2419 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
2420 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
2421 | case PackCategoriesIntoKey(fcZero, fcNormal): |
2422 | if (rhs.sign) |
2423 | return cmpGreaterThan; |
2424 | else |
2425 | return cmpLessThan; |
2426 | |
2427 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
2428 | if (sign == rhs.sign) |
2429 | return cmpEqual; |
2430 | else if (sign) |
2431 | return cmpLessThan; |
2432 | else |
2433 | return cmpGreaterThan; |
2434 | |
2435 | case PackCategoriesIntoKey(fcZero, fcZero): |
2436 | return cmpEqual; |
2437 | |
2438 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
2439 | break; |
2440 | } |
2441 | |
2442 | /* Two normal numbers. Do they have the same sign? */ |
2443 | if (sign != rhs.sign) { |
2444 | if (sign) |
2445 | result = cmpLessThan; |
2446 | else |
2447 | result = cmpGreaterThan; |
2448 | } else { |
2449 | /* Compare absolute values; invert result if negative. */ |
2450 | result = compareAbsoluteValue(rhs); |
2451 | |
2452 | if (sign) { |
2453 | if (result == cmpLessThan) |
2454 | result = cmpGreaterThan; |
2455 | else if (result == cmpGreaterThan) |
2456 | result = cmpLessThan; |
2457 | } |
2458 | } |
2459 | |
2460 | return result; |
2461 | } |
2462 | |
2463 | /// IEEEFloat::convert - convert a value of one floating point type to another. |
2464 | /// The return value corresponds to the IEEE754 exceptions. *losesInfo |
2465 | /// records whether the transformation lost information, i.e. whether |
2466 | /// converting the result back to the original type will produce the |
2467 | /// original value (this is almost the same as return value==fsOK, but there |
2468 | /// are edge cases where this is not so). |
2469 | |
2470 | IEEEFloat::opStatus IEEEFloat::convert(const fltSemantics &toSemantics, |
2471 | roundingMode rounding_mode, |
2472 | bool *losesInfo) { |
2473 | lostFraction lostFraction; |
2474 | unsigned int newPartCount, oldPartCount; |
2475 | opStatus fs; |
2476 | int shift; |
2477 | const fltSemantics &fromSemantics = *semantics; |
2478 | bool is_signaling = isSignaling(); |
2479 | |
2480 | lostFraction = lfExactlyZero; |
2481 | newPartCount = partCountForBits(bits: toSemantics.precision + 1); |
2482 | oldPartCount = partCount(); |
2483 | shift = toSemantics.precision - fromSemantics.precision; |
2484 | |
2485 | bool X86SpecialNan = false; |
2486 | if (&fromSemantics == &semX87DoubleExtended && |
2487 | &toSemantics != &semX87DoubleExtended && category == fcNaN && |
2488 | (!(*significandParts() & 0x8000000000000000ULL) || |
2489 | !(*significandParts() & 0x4000000000000000ULL))) { |
2490 | // x86 has some unusual NaNs which cannot be represented in any other |
2491 | // format; note them here. |
2492 | X86SpecialNan = true; |
2493 | } |
2494 | |
2495 | // If this is a truncation of a denormal number, and the target semantics |
2496 | // has larger exponent range than the source semantics (this can happen |
2497 | // when truncating from PowerPC double-double to double format), the |
2498 | // right shift could lose result mantissa bits. Adjust exponent instead |
2499 | // of performing excessive shift. |
2500 | // Also do a similar trick in case shifting denormal would produce zero |
2501 | // significand as this case isn't handled correctly by normalize. |
2502 | if (shift < 0 && isFiniteNonZero()) { |
2503 | int omsb = significandMSB() + 1; |
2504 | int exponentChange = omsb - fromSemantics.precision; |
2505 | if (exponent + exponentChange < toSemantics.minExponent) |
2506 | exponentChange = toSemantics.minExponent - exponent; |
2507 | if (exponentChange < shift) |
2508 | exponentChange = shift; |
2509 | if (exponentChange < 0) { |
2510 | shift -= exponentChange; |
2511 | exponent += exponentChange; |
2512 | } else if (omsb <= -shift) { |
2513 | exponentChange = omsb + shift - 1; // leave at least one bit set |
2514 | shift -= exponentChange; |
2515 | exponent += exponentChange; |
2516 | } |
2517 | } |
2518 | |
2519 | // If this is a truncation, perform the shift before we narrow the storage. |
2520 | if (shift < 0 && (isFiniteNonZero() || |
2521 | (category == fcNaN && semantics->nonFiniteBehavior != |
2522 | fltNonfiniteBehavior::NanOnly))) |
2523 | lostFraction = shiftRight(dst: significandParts(), parts: oldPartCount, bits: -shift); |
2524 | |
2525 | // Fix the storage so it can hold to new value. |
2526 | if (newPartCount > oldPartCount) { |
2527 | // The new type requires more storage; make it available. |
2528 | integerPart *newParts; |
2529 | newParts = new integerPart[newPartCount]; |
2530 | APInt::tcSet(newParts, 0, newPartCount); |
2531 | if (isFiniteNonZero() || category==fcNaN) |
2532 | APInt::tcAssign(newParts, significandParts(), oldPartCount); |
2533 | freeSignificand(); |
2534 | significand.parts = newParts; |
2535 | } else if (newPartCount == 1 && oldPartCount != 1) { |
2536 | // Switch to built-in storage for a single part. |
2537 | integerPart newPart = 0; |
2538 | if (isFiniteNonZero() || category==fcNaN) |
2539 | newPart = significandParts()[0]; |
2540 | freeSignificand(); |
2541 | significand.part = newPart; |
2542 | } |
2543 | |
2544 | // Now that we have the right storage, switch the semantics. |
2545 | semantics = &toSemantics; |
2546 | |
2547 | // If this is an extension, perform the shift now that the storage is |
2548 | // available. |
2549 | if (shift > 0 && (isFiniteNonZero() || category==fcNaN)) |
2550 | APInt::tcShiftLeft(significandParts(), Words: newPartCount, Count: shift); |
2551 | |
2552 | if (isFiniteNonZero()) { |
2553 | fs = normalize(rounding_mode, lost_fraction: lostFraction); |
2554 | *losesInfo = (fs != opOK); |
2555 | } else if (category == fcNaN) { |
2556 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
2557 | *losesInfo = |
2558 | fromSemantics.nonFiniteBehavior != fltNonfiniteBehavior::NanOnly; |
2559 | makeNaN(SNaN: false, Negative: sign); |
2560 | return is_signaling ? opInvalidOp : opOK; |
2561 | } |
2562 | |
2563 | // If NaN is negative zero, we need to create a new NaN to avoid converting |
2564 | // NaN to -Inf. |
2565 | if (fromSemantics.nanEncoding == fltNanEncoding::NegativeZero && |
2566 | semantics->nanEncoding != fltNanEncoding::NegativeZero) |
2567 | makeNaN(SNaN: false, Negative: false); |
2568 | |
2569 | *losesInfo = lostFraction != lfExactlyZero || X86SpecialNan; |
2570 | |
2571 | // For x87 extended precision, we want to make a NaN, not a special NaN if |
2572 | // the input wasn't special either. |
2573 | if (!X86SpecialNan && semantics == &semX87DoubleExtended) |
2574 | APInt::tcSetBit(significandParts(), bit: semantics->precision - 1); |
2575 | |
2576 | // Convert of sNaN creates qNaN and raises an exception (invalid op). |
2577 | // This also guarantees that a sNaN does not become Inf on a truncation |
2578 | // that loses all payload bits. |
2579 | if (is_signaling) { |
2580 | makeQuiet(); |
2581 | fs = opInvalidOp; |
2582 | } else { |
2583 | fs = opOK; |
2584 | } |
2585 | } else if (category == fcInfinity && |
2586 | semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
2587 | makeNaN(SNaN: false, Negative: sign); |
2588 | *losesInfo = true; |
2589 | fs = opInexact; |
2590 | } else if (category == fcZero && |
2591 | semantics->nanEncoding == fltNanEncoding::NegativeZero) { |
2592 | // Negative zero loses info, but positive zero doesn't. |
2593 | *losesInfo = |
2594 | fromSemantics.nanEncoding != fltNanEncoding::NegativeZero && sign; |
2595 | fs = *losesInfo ? opInexact : opOK; |
2596 | // NaN is negative zero means -0 -> +0, which can lose information |
2597 | sign = false; |
2598 | } else { |
2599 | *losesInfo = false; |
2600 | fs = opOK; |
2601 | } |
2602 | |
2603 | return fs; |
2604 | } |
2605 | |
2606 | /* Convert a floating point number to an integer according to the |
2607 | rounding mode. If the rounded integer value is out of range this |
2608 | returns an invalid operation exception and the contents of the |
2609 | destination parts are unspecified. If the rounded value is in |
2610 | range but the floating point number is not the exact integer, the C |
2611 | standard doesn't require an inexact exception to be raised. IEEE |
2612 | 854 does require it so we do that. |
2613 | |
2614 | Note that for conversions to integer type the C standard requires |
2615 | round-to-zero to always be used. */ |
2616 | IEEEFloat::opStatus IEEEFloat::convertToSignExtendedInteger( |
2617 | MutableArrayRef<integerPart> parts, unsigned int width, bool isSigned, |
2618 | roundingMode rounding_mode, bool *isExact) const { |
2619 | lostFraction lost_fraction; |
2620 | const integerPart *src; |
2621 | unsigned int dstPartsCount, truncatedBits; |
2622 | |
2623 | *isExact = false; |
2624 | |
2625 | /* Handle the three special cases first. */ |
2626 | if (category == fcInfinity || category == fcNaN) |
2627 | return opInvalidOp; |
2628 | |
2629 | dstPartsCount = partCountForBits(bits: width); |
2630 | assert(dstPartsCount <= parts.size() && "Integer too big" ); |
2631 | |
2632 | if (category == fcZero) { |
2633 | APInt::tcSet(parts.data(), 0, dstPartsCount); |
2634 | // Negative zero can't be represented as an int. |
2635 | *isExact = !sign; |
2636 | return opOK; |
2637 | } |
2638 | |
2639 | src = significandParts(); |
2640 | |
2641 | /* Step 1: place our absolute value, with any fraction truncated, in |
2642 | the destination. */ |
2643 | if (exponent < 0) { |
2644 | /* Our absolute value is less than one; truncate everything. */ |
2645 | APInt::tcSet(parts.data(), 0, dstPartsCount); |
2646 | /* For exponent -1 the integer bit represents .5, look at that. |
2647 | For smaller exponents leftmost truncated bit is 0. */ |
2648 | truncatedBits = semantics->precision -1U - exponent; |
2649 | } else { |
2650 | /* We want the most significant (exponent + 1) bits; the rest are |
2651 | truncated. */ |
2652 | unsigned int bits = exponent + 1U; |
2653 | |
2654 | /* Hopelessly large in magnitude? */ |
2655 | if (bits > width) |
2656 | return opInvalidOp; |
2657 | |
2658 | if (bits < semantics->precision) { |
2659 | /* We truncate (semantics->precision - bits) bits. */ |
2660 | truncatedBits = semantics->precision - bits; |
2661 | APInt::tcExtract(parts.data(), dstCount: dstPartsCount, src, srcBits: bits, srcLSB: truncatedBits); |
2662 | } else { |
2663 | /* We want at least as many bits as are available. */ |
2664 | APInt::tcExtract(parts.data(), dstCount: dstPartsCount, src, srcBits: semantics->precision, |
2665 | srcLSB: 0); |
2666 | APInt::tcShiftLeft(parts.data(), Words: dstPartsCount, |
2667 | Count: bits - semantics->precision); |
2668 | truncatedBits = 0; |
2669 | } |
2670 | } |
2671 | |
2672 | /* Step 2: work out any lost fraction, and increment the absolute |
2673 | value if we would round away from zero. */ |
2674 | if (truncatedBits) { |
2675 | lost_fraction = lostFractionThroughTruncation(parts: src, partCount: partCount(), |
2676 | bits: truncatedBits); |
2677 | if (lost_fraction != lfExactlyZero && |
2678 | roundAwayFromZero(rounding_mode, lost_fraction, bit: truncatedBits)) { |
2679 | if (APInt::tcIncrement(dst: parts.data(), parts: dstPartsCount)) |
2680 | return opInvalidOp; /* Overflow. */ |
2681 | } |
2682 | } else { |
2683 | lost_fraction = lfExactlyZero; |
2684 | } |
2685 | |
2686 | /* Step 3: check if we fit in the destination. */ |
2687 | unsigned int omsb = APInt::tcMSB(parts: parts.data(), n: dstPartsCount) + 1; |
2688 | |
2689 | if (sign) { |
2690 | if (!isSigned) { |
2691 | /* Negative numbers cannot be represented as unsigned. */ |
2692 | if (omsb != 0) |
2693 | return opInvalidOp; |
2694 | } else { |
2695 | /* It takes omsb bits to represent the unsigned integer value. |
2696 | We lose a bit for the sign, but care is needed as the |
2697 | maximally negative integer is a special case. */ |
2698 | if (omsb == width && |
2699 | APInt::tcLSB(parts.data(), n: dstPartsCount) + 1 != omsb) |
2700 | return opInvalidOp; |
2701 | |
2702 | /* This case can happen because of rounding. */ |
2703 | if (omsb > width) |
2704 | return opInvalidOp; |
2705 | } |
2706 | |
2707 | APInt::tcNegate (parts.data(), dstPartsCount); |
2708 | } else { |
2709 | if (omsb >= width + !isSigned) |
2710 | return opInvalidOp; |
2711 | } |
2712 | |
2713 | if (lost_fraction == lfExactlyZero) { |
2714 | *isExact = true; |
2715 | return opOK; |
2716 | } else |
2717 | return opInexact; |
2718 | } |
2719 | |
2720 | /* Same as convertToSignExtendedInteger, except we provide |
2721 | deterministic values in case of an invalid operation exception, |
2722 | namely zero for NaNs and the minimal or maximal value respectively |
2723 | for underflow or overflow. |
2724 | The *isExact output tells whether the result is exact, in the sense |
2725 | that converting it back to the original floating point type produces |
2726 | the original value. This is almost equivalent to result==opOK, |
2727 | except for negative zeroes. |
2728 | */ |
2729 | IEEEFloat::opStatus |
2730 | IEEEFloat::convertToInteger(MutableArrayRef<integerPart> parts, |
2731 | unsigned int width, bool isSigned, |
2732 | roundingMode rounding_mode, bool *isExact) const { |
2733 | opStatus fs; |
2734 | |
2735 | fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode, |
2736 | isExact); |
2737 | |
2738 | if (fs == opInvalidOp) { |
2739 | unsigned int bits, dstPartsCount; |
2740 | |
2741 | dstPartsCount = partCountForBits(bits: width); |
2742 | assert(dstPartsCount <= parts.size() && "Integer too big" ); |
2743 | |
2744 | if (category == fcNaN) |
2745 | bits = 0; |
2746 | else if (sign) |
2747 | bits = isSigned; |
2748 | else |
2749 | bits = width - isSigned; |
2750 | |
2751 | tcSetLeastSignificantBits(dst: parts.data(), parts: dstPartsCount, bits); |
2752 | if (sign && isSigned) |
2753 | APInt::tcShiftLeft(parts.data(), Words: dstPartsCount, Count: width - 1); |
2754 | } |
2755 | |
2756 | return fs; |
2757 | } |
2758 | |
2759 | /* Convert an unsigned integer SRC to a floating point number, |
2760 | rounding according to ROUNDING_MODE. The sign of the floating |
2761 | point number is not modified. */ |
2762 | IEEEFloat::opStatus IEEEFloat::convertFromUnsignedParts( |
2763 | const integerPart *src, unsigned int srcCount, roundingMode rounding_mode) { |
2764 | unsigned int omsb, precision, dstCount; |
2765 | integerPart *dst; |
2766 | lostFraction lost_fraction; |
2767 | |
2768 | category = fcNormal; |
2769 | omsb = APInt::tcMSB(parts: src, n: srcCount) + 1; |
2770 | dst = significandParts(); |
2771 | dstCount = partCount(); |
2772 | precision = semantics->precision; |
2773 | |
2774 | /* We want the most significant PRECISION bits of SRC. There may not |
2775 | be that many; extract what we can. */ |
2776 | if (precision <= omsb) { |
2777 | exponent = omsb - 1; |
2778 | lost_fraction = lostFractionThroughTruncation(parts: src, partCount: srcCount, |
2779 | bits: omsb - precision); |
2780 | APInt::tcExtract(dst, dstCount, src, srcBits: precision, srcLSB: omsb - precision); |
2781 | } else { |
2782 | exponent = precision - 1; |
2783 | lost_fraction = lfExactlyZero; |
2784 | APInt::tcExtract(dst, dstCount, src, srcBits: omsb, srcLSB: 0); |
2785 | } |
2786 | |
2787 | return normalize(rounding_mode, lost_fraction); |
2788 | } |
2789 | |
2790 | IEEEFloat::opStatus IEEEFloat::convertFromAPInt(const APInt &Val, bool isSigned, |
2791 | roundingMode rounding_mode) { |
2792 | unsigned int partCount = Val.getNumWords(); |
2793 | APInt api = Val; |
2794 | |
2795 | sign = false; |
2796 | if (isSigned && api.isNegative()) { |
2797 | sign = true; |
2798 | api = -api; |
2799 | } |
2800 | |
2801 | return convertFromUnsignedParts(src: api.getRawData(), srcCount: partCount, rounding_mode); |
2802 | } |
2803 | |
2804 | /* Convert a two's complement integer SRC to a floating point number, |
2805 | rounding according to ROUNDING_MODE. ISSIGNED is true if the |
2806 | integer is signed, in which case it must be sign-extended. */ |
2807 | IEEEFloat::opStatus |
2808 | IEEEFloat::convertFromSignExtendedInteger(const integerPart *src, |
2809 | unsigned int srcCount, bool isSigned, |
2810 | roundingMode rounding_mode) { |
2811 | opStatus status; |
2812 | |
2813 | if (isSigned && |
2814 | APInt::tcExtractBit(src, bit: srcCount * integerPartWidth - 1)) { |
2815 | integerPart *copy; |
2816 | |
2817 | /* If we're signed and negative negate a copy. */ |
2818 | sign = true; |
2819 | copy = new integerPart[srcCount]; |
2820 | APInt::tcAssign(copy, src, srcCount); |
2821 | APInt::tcNegate(copy, srcCount); |
2822 | status = convertFromUnsignedParts(src: copy, srcCount, rounding_mode); |
2823 | delete [] copy; |
2824 | } else { |
2825 | sign = false; |
2826 | status = convertFromUnsignedParts(src, srcCount, rounding_mode); |
2827 | } |
2828 | |
2829 | return status; |
2830 | } |
2831 | |
2832 | /* FIXME: should this just take a const APInt reference? */ |
2833 | IEEEFloat::opStatus |
2834 | IEEEFloat::convertFromZeroExtendedInteger(const integerPart *parts, |
2835 | unsigned int width, bool isSigned, |
2836 | roundingMode rounding_mode) { |
2837 | unsigned int partCount = partCountForBits(bits: width); |
2838 | APInt api = APInt(width, ArrayRef(parts, partCount)); |
2839 | |
2840 | sign = false; |
2841 | if (isSigned && APInt::tcExtractBit(parts, bit: width - 1)) { |
2842 | sign = true; |
2843 | api = -api; |
2844 | } |
2845 | |
2846 | return convertFromUnsignedParts(src: api.getRawData(), srcCount: partCount, rounding_mode); |
2847 | } |
2848 | |
2849 | Expected<IEEEFloat::opStatus> |
2850 | IEEEFloat::convertFromHexadecimalString(StringRef s, |
2851 | roundingMode rounding_mode) { |
2852 | lostFraction lost_fraction = lfExactlyZero; |
2853 | |
2854 | category = fcNormal; |
2855 | zeroSignificand(); |
2856 | exponent = 0; |
2857 | |
2858 | integerPart *significand = significandParts(); |
2859 | unsigned partsCount = partCount(); |
2860 | unsigned bitPos = partsCount * integerPartWidth; |
2861 | bool computedTrailingFraction = false; |
2862 | |
2863 | // Skip leading zeroes and any (hexa)decimal point. |
2864 | StringRef::iterator begin = s.begin(); |
2865 | StringRef::iterator end = s.end(); |
2866 | StringRef::iterator dot; |
2867 | auto PtrOrErr = skipLeadingZeroesAndAnyDot(begin, end, dot: &dot); |
2868 | if (!PtrOrErr) |
2869 | return PtrOrErr.takeError(); |
2870 | StringRef::iterator p = *PtrOrErr; |
2871 | StringRef::iterator firstSignificantDigit = p; |
2872 | |
2873 | while (p != end) { |
2874 | integerPart hex_value; |
2875 | |
2876 | if (*p == '.') { |
2877 | if (dot != end) |
2878 | return createError(Err: "String contains multiple dots" ); |
2879 | dot = p++; |
2880 | continue; |
2881 | } |
2882 | |
2883 | hex_value = hexDigitValue(C: *p); |
2884 | if (hex_value == UINT_MAX) |
2885 | break; |
2886 | |
2887 | p++; |
2888 | |
2889 | // Store the number while we have space. |
2890 | if (bitPos) { |
2891 | bitPos -= 4; |
2892 | hex_value <<= bitPos % integerPartWidth; |
2893 | significand[bitPos / integerPartWidth] |= hex_value; |
2894 | } else if (!computedTrailingFraction) { |
2895 | auto FractOrErr = trailingHexadecimalFraction(p, end, digitValue: hex_value); |
2896 | if (!FractOrErr) |
2897 | return FractOrErr.takeError(); |
2898 | lost_fraction = *FractOrErr; |
2899 | computedTrailingFraction = true; |
2900 | } |
2901 | } |
2902 | |
2903 | /* Hex floats require an exponent but not a hexadecimal point. */ |
2904 | if (p == end) |
2905 | return createError(Err: "Hex strings require an exponent" ); |
2906 | if (*p != 'p' && *p != 'P') |
2907 | return createError(Err: "Invalid character in significand" ); |
2908 | if (p == begin) |
2909 | return createError(Err: "Significand has no digits" ); |
2910 | if (dot != end && p - begin == 1) |
2911 | return createError(Err: "Significand has no digits" ); |
2912 | |
2913 | /* Ignore the exponent if we are zero. */ |
2914 | if (p != firstSignificantDigit) { |
2915 | int expAdjustment; |
2916 | |
2917 | /* Implicit hexadecimal point? */ |
2918 | if (dot == end) |
2919 | dot = p; |
2920 | |
2921 | /* Calculate the exponent adjustment implicit in the number of |
2922 | significant digits. */ |
2923 | expAdjustment = static_cast<int>(dot - firstSignificantDigit); |
2924 | if (expAdjustment < 0) |
2925 | expAdjustment++; |
2926 | expAdjustment = expAdjustment * 4 - 1; |
2927 | |
2928 | /* Adjust for writing the significand starting at the most |
2929 | significant nibble. */ |
2930 | expAdjustment += semantics->precision; |
2931 | expAdjustment -= partsCount * integerPartWidth; |
2932 | |
2933 | /* Adjust for the given exponent. */ |
2934 | auto ExpOrErr = totalExponent(p: p + 1, end, exponentAdjustment: expAdjustment); |
2935 | if (!ExpOrErr) |
2936 | return ExpOrErr.takeError(); |
2937 | exponent = *ExpOrErr; |
2938 | } |
2939 | |
2940 | return normalize(rounding_mode, lost_fraction); |
2941 | } |
2942 | |
2943 | IEEEFloat::opStatus |
2944 | IEEEFloat::roundSignificandWithExponent(const integerPart *decSigParts, |
2945 | unsigned sigPartCount, int exp, |
2946 | roundingMode rounding_mode) { |
2947 | unsigned int parts, pow5PartCount; |
2948 | fltSemantics calcSemantics = { .maxExponent: 32767, .minExponent: -32767, .precision: 0, .sizeInBits: 0 }; |
2949 | integerPart pow5Parts[maxPowerOfFiveParts]; |
2950 | bool isNearest; |
2951 | |
2952 | isNearest = (rounding_mode == rmNearestTiesToEven || |
2953 | rounding_mode == rmNearestTiesToAway); |
2954 | |
2955 | parts = partCountForBits(bits: semantics->precision + 11); |
2956 | |
2957 | /* Calculate pow(5, abs(exp)). */ |
2958 | pow5PartCount = powerOf5(dst: pow5Parts, power: exp >= 0 ? exp: -exp); |
2959 | |
2960 | for (;; parts *= 2) { |
2961 | opStatus sigStatus, powStatus; |
2962 | unsigned int excessPrecision, truncatedBits; |
2963 | |
2964 | calcSemantics.precision = parts * integerPartWidth - 1; |
2965 | excessPrecision = calcSemantics.precision - semantics->precision; |
2966 | truncatedBits = excessPrecision; |
2967 | |
2968 | IEEEFloat decSig(calcSemantics, uninitialized); |
2969 | decSig.makeZero(Neg: sign); |
2970 | IEEEFloat pow5(calcSemantics); |
2971 | |
2972 | sigStatus = decSig.convertFromUnsignedParts(src: decSigParts, srcCount: sigPartCount, |
2973 | rounding_mode: rmNearestTiesToEven); |
2974 | powStatus = pow5.convertFromUnsignedParts(src: pow5Parts, srcCount: pow5PartCount, |
2975 | rounding_mode: rmNearestTiesToEven); |
2976 | /* Add exp, as 10^n = 5^n * 2^n. */ |
2977 | decSig.exponent += exp; |
2978 | |
2979 | lostFraction calcLostFraction; |
2980 | integerPart HUerr, HUdistance; |
2981 | unsigned int powHUerr; |
2982 | |
2983 | if (exp >= 0) { |
2984 | /* multiplySignificand leaves the precision-th bit set to 1. */ |
2985 | calcLostFraction = decSig.multiplySignificand(rhs: pow5); |
2986 | powHUerr = powStatus != opOK; |
2987 | } else { |
2988 | calcLostFraction = decSig.divideSignificand(rhs: pow5); |
2989 | /* Denormal numbers have less precision. */ |
2990 | if (decSig.exponent < semantics->minExponent) { |
2991 | excessPrecision += (semantics->minExponent - decSig.exponent); |
2992 | truncatedBits = excessPrecision; |
2993 | if (excessPrecision > calcSemantics.precision) |
2994 | excessPrecision = calcSemantics.precision; |
2995 | } |
2996 | /* Extra half-ulp lost in reciprocal of exponent. */ |
2997 | powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2; |
2998 | } |
2999 | |
3000 | /* Both multiplySignificand and divideSignificand return the |
3001 | result with the integer bit set. */ |
3002 | assert(APInt::tcExtractBit |
3003 | (decSig.significandParts(), calcSemantics.precision - 1) == 1); |
3004 | |
3005 | HUerr = HUerrBound(inexactMultiply: calcLostFraction != lfExactlyZero, HUerr1: sigStatus != opOK, |
3006 | HUerr2: powHUerr); |
3007 | HUdistance = 2 * ulpsFromBoundary(parts: decSig.significandParts(), |
3008 | bits: excessPrecision, isNearest); |
3009 | |
3010 | /* Are we guaranteed to round correctly if we truncate? */ |
3011 | if (HUdistance >= HUerr) { |
3012 | APInt::tcExtract(significandParts(), dstCount: partCount(), decSig.significandParts(), |
3013 | srcBits: calcSemantics.precision - excessPrecision, |
3014 | srcLSB: excessPrecision); |
3015 | /* Take the exponent of decSig. If we tcExtract-ed less bits |
3016 | above we must adjust our exponent to compensate for the |
3017 | implicit right shift. */ |
3018 | exponent = (decSig.exponent + semantics->precision |
3019 | - (calcSemantics.precision - excessPrecision)); |
3020 | calcLostFraction = lostFractionThroughTruncation(parts: decSig.significandParts(), |
3021 | partCount: decSig.partCount(), |
3022 | bits: truncatedBits); |
3023 | return normalize(rounding_mode, lost_fraction: calcLostFraction); |
3024 | } |
3025 | } |
3026 | } |
3027 | |
3028 | Expected<IEEEFloat::opStatus> |
3029 | IEEEFloat::convertFromDecimalString(StringRef str, roundingMode rounding_mode) { |
3030 | decimalInfo D; |
3031 | opStatus fs; |
3032 | |
3033 | /* Scan the text. */ |
3034 | StringRef::iterator p = str.begin(); |
3035 | if (Error Err = interpretDecimal(begin: p, end: str.end(), D: &D)) |
3036 | return std::move(Err); |
3037 | |
3038 | /* Handle the quick cases. First the case of no significant digits, |
3039 | i.e. zero, and then exponents that are obviously too large or too |
3040 | small. Writing L for log 10 / log 2, a number d.ddddd*10^exp |
3041 | definitely overflows if |
3042 | |
3043 | (exp - 1) * L >= maxExponent |
3044 | |
3045 | and definitely underflows to zero where |
3046 | |
3047 | (exp + 1) * L <= minExponent - precision |
3048 | |
3049 | With integer arithmetic the tightest bounds for L are |
3050 | |
3051 | 93/28 < L < 196/59 [ numerator <= 256 ] |
3052 | 42039/12655 < L < 28738/8651 [ numerator <= 65536 ] |
3053 | */ |
3054 | |
3055 | // Test if we have a zero number allowing for strings with no null terminators |
3056 | // and zero decimals with non-zero exponents. |
3057 | // |
3058 | // We computed firstSigDigit by ignoring all zeros and dots. Thus if |
3059 | // D->firstSigDigit equals str.end(), every digit must be a zero and there can |
3060 | // be at most one dot. On the other hand, if we have a zero with a non-zero |
3061 | // exponent, then we know that D.firstSigDigit will be non-numeric. |
3062 | if (D.firstSigDigit == str.end() || decDigitValue(c: *D.firstSigDigit) >= 10U) { |
3063 | category = fcZero; |
3064 | fs = opOK; |
3065 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
3066 | sign = false; |
3067 | |
3068 | /* Check whether the normalized exponent is high enough to overflow |
3069 | max during the log-rebasing in the max-exponent check below. */ |
3070 | } else if (D.normalizedExponent - 1 > INT_MAX / 42039) { |
3071 | fs = handleOverflow(rounding_mode); |
3072 | |
3073 | /* If it wasn't, then it also wasn't high enough to overflow max |
3074 | during the log-rebasing in the min-exponent check. Check that it |
3075 | won't overflow min in either check, then perform the min-exponent |
3076 | check. */ |
3077 | } else if (D.normalizedExponent - 1 < INT_MIN / 42039 || |
3078 | (D.normalizedExponent + 1) * 28738 <= |
3079 | 8651 * (semantics->minExponent - (int) semantics->precision)) { |
3080 | /* Underflow to zero and round. */ |
3081 | category = fcNormal; |
3082 | zeroSignificand(); |
3083 | fs = normalize(rounding_mode, lost_fraction: lfLessThanHalf); |
3084 | |
3085 | /* We can finally safely perform the max-exponent check. */ |
3086 | } else if ((D.normalizedExponent - 1) * 42039 |
3087 | >= 12655 * semantics->maxExponent) { |
3088 | /* Overflow and round. */ |
3089 | fs = handleOverflow(rounding_mode); |
3090 | } else { |
3091 | integerPart *decSignificand; |
3092 | unsigned int partCount; |
3093 | |
3094 | /* A tight upper bound on number of bits required to hold an |
3095 | N-digit decimal integer is N * 196 / 59. Allocate enough space |
3096 | to hold the full significand, and an extra part required by |
3097 | tcMultiplyPart. */ |
3098 | partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1; |
3099 | partCount = partCountForBits(bits: 1 + 196 * partCount / 59); |
3100 | decSignificand = new integerPart[partCount + 1]; |
3101 | partCount = 0; |
3102 | |
3103 | /* Convert to binary efficiently - we do almost all multiplication |
3104 | in an integerPart. When this would overflow do we do a single |
3105 | bignum multiplication, and then revert again to multiplication |
3106 | in an integerPart. */ |
3107 | do { |
3108 | integerPart decValue, val, multiplier; |
3109 | |
3110 | val = 0; |
3111 | multiplier = 1; |
3112 | |
3113 | do { |
3114 | if (*p == '.') { |
3115 | p++; |
3116 | if (p == str.end()) { |
3117 | break; |
3118 | } |
3119 | } |
3120 | decValue = decDigitValue(c: *p++); |
3121 | if (decValue >= 10U) { |
3122 | delete[] decSignificand; |
3123 | return createError(Err: "Invalid character in significand" ); |
3124 | } |
3125 | multiplier *= 10; |
3126 | val = val * 10 + decValue; |
3127 | /* The maximum number that can be multiplied by ten with any |
3128 | digit added without overflowing an integerPart. */ |
3129 | } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10); |
3130 | |
3131 | /* Multiply out the current part. */ |
3132 | APInt::tcMultiplyPart(dst: decSignificand, src: decSignificand, multiplier, carry: val, |
3133 | srcParts: partCount, dstParts: partCount + 1, add: false); |
3134 | |
3135 | /* If we used another part (likely but not guaranteed), increase |
3136 | the count. */ |
3137 | if (decSignificand[partCount]) |
3138 | partCount++; |
3139 | } while (p <= D.lastSigDigit); |
3140 | |
3141 | category = fcNormal; |
3142 | fs = roundSignificandWithExponent(decSigParts: decSignificand, sigPartCount: partCount, |
3143 | exp: D.exponent, rounding_mode); |
3144 | |
3145 | delete [] decSignificand; |
3146 | } |
3147 | |
3148 | return fs; |
3149 | } |
3150 | |
3151 | bool IEEEFloat::convertFromStringSpecials(StringRef str) { |
3152 | const size_t MIN_NAME_SIZE = 3; |
3153 | |
3154 | if (str.size() < MIN_NAME_SIZE) |
3155 | return false; |
3156 | |
3157 | if (str == "inf" || str == "INFINITY" || str == "+Inf" ) { |
3158 | makeInf(Neg: false); |
3159 | return true; |
3160 | } |
3161 | |
3162 | bool IsNegative = str.front() == '-'; |
3163 | if (IsNegative) { |
3164 | str = str.drop_front(); |
3165 | if (str.size() < MIN_NAME_SIZE) |
3166 | return false; |
3167 | |
3168 | if (str == "inf" || str == "INFINITY" || str == "Inf" ) { |
3169 | makeInf(Neg: true); |
3170 | return true; |
3171 | } |
3172 | } |
3173 | |
3174 | // If we have a 's' (or 'S') prefix, then this is a Signaling NaN. |
3175 | bool IsSignaling = str.front() == 's' || str.front() == 'S'; |
3176 | if (IsSignaling) { |
3177 | str = str.drop_front(); |
3178 | if (str.size() < MIN_NAME_SIZE) |
3179 | return false; |
3180 | } |
3181 | |
3182 | if (str.starts_with(Prefix: "nan" ) || str.starts_with(Prefix: "NaN" )) { |
3183 | str = str.drop_front(N: 3); |
3184 | |
3185 | // A NaN without payload. |
3186 | if (str.empty()) { |
3187 | makeNaN(SNaN: IsSignaling, Negative: IsNegative); |
3188 | return true; |
3189 | } |
3190 | |
3191 | // Allow the payload to be inside parentheses. |
3192 | if (str.front() == '(') { |
3193 | // Parentheses should be balanced (and not empty). |
3194 | if (str.size() <= 2 || str.back() != ')') |
3195 | return false; |
3196 | |
3197 | str = str.slice(Start: 1, End: str.size() - 1); |
3198 | } |
3199 | |
3200 | // Determine the payload number's radix. |
3201 | unsigned Radix = 10; |
3202 | if (str[0] == '0') { |
3203 | if (str.size() > 1 && tolower(c: str[1]) == 'x') { |
3204 | str = str.drop_front(N: 2); |
3205 | Radix = 16; |
3206 | } else |
3207 | Radix = 8; |
3208 | } |
3209 | |
3210 | // Parse the payload and make the NaN. |
3211 | APInt Payload; |
3212 | if (!str.getAsInteger(Radix, Result&: Payload)) { |
3213 | makeNaN(SNaN: IsSignaling, Negative: IsNegative, fill: &Payload); |
3214 | return true; |
3215 | } |
3216 | } |
3217 | |
3218 | return false; |
3219 | } |
3220 | |
3221 | Expected<IEEEFloat::opStatus> |
3222 | IEEEFloat::convertFromString(StringRef str, roundingMode rounding_mode) { |
3223 | if (str.empty()) |
3224 | return createError(Err: "Invalid string length" ); |
3225 | |
3226 | // Handle special cases. |
3227 | if (convertFromStringSpecials(str)) |
3228 | return opOK; |
3229 | |
3230 | /* Handle a leading minus sign. */ |
3231 | StringRef::iterator p = str.begin(); |
3232 | size_t slen = str.size(); |
3233 | sign = *p == '-' ? 1 : 0; |
3234 | if (*p == '-' || *p == '+') { |
3235 | p++; |
3236 | slen--; |
3237 | if (!slen) |
3238 | return createError(Err: "String has no digits" ); |
3239 | } |
3240 | |
3241 | if (slen >= 2 && p[0] == '0' && (p[1] == 'x' || p[1] == 'X')) { |
3242 | if (slen == 2) |
3243 | return createError(Err: "Invalid string" ); |
3244 | return convertFromHexadecimalString(s: StringRef(p + 2, slen - 2), |
3245 | rounding_mode); |
3246 | } |
3247 | |
3248 | return convertFromDecimalString(str: StringRef(p, slen), rounding_mode); |
3249 | } |
3250 | |
3251 | /* Write out a hexadecimal representation of the floating point value |
3252 | to DST, which must be of sufficient size, in the C99 form |
3253 | [-]0xh.hhhhp[+-]d. Return the number of characters written, |
3254 | excluding the terminating NUL. |
3255 | |
3256 | If UPPERCASE, the output is in upper case, otherwise in lower case. |
3257 | |
3258 | HEXDIGITS digits appear altogether, rounding the value if |
3259 | necessary. If HEXDIGITS is 0, the minimal precision to display the |
3260 | number precisely is used instead. If nothing would appear after |
3261 | the decimal point it is suppressed. |
3262 | |
3263 | The decimal exponent is always printed and has at least one digit. |
3264 | Zero values display an exponent of zero. Infinities and NaNs |
3265 | appear as "infinity" or "nan" respectively. |
3266 | |
3267 | The above rules are as specified by C99. There is ambiguity about |
3268 | what the leading hexadecimal digit should be. This implementation |
3269 | uses whatever is necessary so that the exponent is displayed as |
3270 | stored. This implies the exponent will fall within the IEEE format |
3271 | range, and the leading hexadecimal digit will be 0 (for denormals), |
3272 | 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with |
3273 | any other digits zero). |
3274 | */ |
3275 | unsigned int IEEEFloat::convertToHexString(char *dst, unsigned int hexDigits, |
3276 | bool upperCase, |
3277 | roundingMode rounding_mode) const { |
3278 | char *p; |
3279 | |
3280 | p = dst; |
3281 | if (sign) |
3282 | *dst++ = '-'; |
3283 | |
3284 | switch (category) { |
3285 | case fcInfinity: |
3286 | memcpy (dest: dst, src: upperCase ? infinityU: infinityL, n: sizeof infinityU - 1); |
3287 | dst += sizeof infinityL - 1; |
3288 | break; |
3289 | |
3290 | case fcNaN: |
3291 | memcpy (dest: dst, src: upperCase ? NaNU: NaNL, n: sizeof NaNU - 1); |
3292 | dst += sizeof NaNU - 1; |
3293 | break; |
3294 | |
3295 | case fcZero: |
3296 | *dst++ = '0'; |
3297 | *dst++ = upperCase ? 'X': 'x'; |
3298 | *dst++ = '0'; |
3299 | if (hexDigits > 1) { |
3300 | *dst++ = '.'; |
3301 | memset (s: dst, c: '0', n: hexDigits - 1); |
3302 | dst += hexDigits - 1; |
3303 | } |
3304 | *dst++ = upperCase ? 'P': 'p'; |
3305 | *dst++ = '0'; |
3306 | break; |
3307 | |
3308 | case fcNormal: |
3309 | dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode); |
3310 | break; |
3311 | } |
3312 | |
3313 | *dst = 0; |
3314 | |
3315 | return static_cast<unsigned int>(dst - p); |
3316 | } |
3317 | |
3318 | /* Does the hard work of outputting the correctly rounded hexadecimal |
3319 | form of a normal floating point number with the specified number of |
3320 | hexadecimal digits. If HEXDIGITS is zero the minimum number of |
3321 | digits necessary to print the value precisely is output. */ |
3322 | char *IEEEFloat::convertNormalToHexString(char *dst, unsigned int hexDigits, |
3323 | bool upperCase, |
3324 | roundingMode rounding_mode) const { |
3325 | unsigned int count, valueBits, shift, partsCount, outputDigits; |
3326 | const char *hexDigitChars; |
3327 | const integerPart *significand; |
3328 | char *p; |
3329 | bool roundUp; |
3330 | |
3331 | *dst++ = '0'; |
3332 | *dst++ = upperCase ? 'X': 'x'; |
3333 | |
3334 | roundUp = false; |
3335 | hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower; |
3336 | |
3337 | significand = significandParts(); |
3338 | partsCount = partCount(); |
3339 | |
3340 | /* +3 because the first digit only uses the single integer bit, so |
3341 | we have 3 virtual zero most-significant-bits. */ |
3342 | valueBits = semantics->precision + 3; |
3343 | shift = integerPartWidth - valueBits % integerPartWidth; |
3344 | |
3345 | /* The natural number of digits required ignoring trailing |
3346 | insignificant zeroes. */ |
3347 | outputDigits = (valueBits - significandLSB () + 3) / 4; |
3348 | |
3349 | /* hexDigits of zero means use the required number for the |
3350 | precision. Otherwise, see if we are truncating. If we are, |
3351 | find out if we need to round away from zero. */ |
3352 | if (hexDigits) { |
3353 | if (hexDigits < outputDigits) { |
3354 | /* We are dropping non-zero bits, so need to check how to round. |
3355 | "bits" is the number of dropped bits. */ |
3356 | unsigned int bits; |
3357 | lostFraction fraction; |
3358 | |
3359 | bits = valueBits - hexDigits * 4; |
3360 | fraction = lostFractionThroughTruncation (parts: significand, partCount: partsCount, bits); |
3361 | roundUp = roundAwayFromZero(rounding_mode, lost_fraction: fraction, bit: bits); |
3362 | } |
3363 | outputDigits = hexDigits; |
3364 | } |
3365 | |
3366 | /* Write the digits consecutively, and start writing in the location |
3367 | of the hexadecimal point. We move the most significant digit |
3368 | left and add the hexadecimal point later. */ |
3369 | p = ++dst; |
3370 | |
3371 | count = (valueBits + integerPartWidth - 1) / integerPartWidth; |
3372 | |
3373 | while (outputDigits && count) { |
3374 | integerPart part; |
3375 | |
3376 | /* Put the most significant integerPartWidth bits in "part". */ |
3377 | if (--count == partsCount) |
3378 | part = 0; /* An imaginary higher zero part. */ |
3379 | else |
3380 | part = significand[count] << shift; |
3381 | |
3382 | if (count && shift) |
3383 | part |= significand[count - 1] >> (integerPartWidth - shift); |
3384 | |
3385 | /* Convert as much of "part" to hexdigits as we can. */ |
3386 | unsigned int curDigits = integerPartWidth / 4; |
3387 | |
3388 | if (curDigits > outputDigits) |
3389 | curDigits = outputDigits; |
3390 | dst += partAsHex (dst, part, count: curDigits, hexDigitChars); |
3391 | outputDigits -= curDigits; |
3392 | } |
3393 | |
3394 | if (roundUp) { |
3395 | char *q = dst; |
3396 | |
3397 | /* Note that hexDigitChars has a trailing '0'. */ |
3398 | do { |
3399 | q--; |
3400 | *q = hexDigitChars[hexDigitValue (C: *q) + 1]; |
3401 | } while (*q == '0'); |
3402 | assert(q >= p); |
3403 | } else { |
3404 | /* Add trailing zeroes. */ |
3405 | memset (s: dst, c: '0', n: outputDigits); |
3406 | dst += outputDigits; |
3407 | } |
3408 | |
3409 | /* Move the most significant digit to before the point, and if there |
3410 | is something after the decimal point add it. This must come |
3411 | after rounding above. */ |
3412 | p[-1] = p[0]; |
3413 | if (dst -1 == p) |
3414 | dst--; |
3415 | else |
3416 | p[0] = '.'; |
3417 | |
3418 | /* Finally output the exponent. */ |
3419 | *dst++ = upperCase ? 'P': 'p'; |
3420 | |
3421 | return writeSignedDecimal (dst, value: exponent); |
3422 | } |
3423 | |
3424 | hash_code hash_value(const IEEEFloat &Arg) { |
3425 | if (!Arg.isFiniteNonZero()) |
3426 | return hash_combine(args: (uint8_t)Arg.category, |
3427 | // NaN has no sign, fix it at zero. |
3428 | args: Arg.isNaN() ? (uint8_t)0 : (uint8_t)Arg.sign, |
3429 | args: Arg.semantics->precision); |
3430 | |
3431 | // Normal floats need their exponent and significand hashed. |
3432 | return hash_combine(args: (uint8_t)Arg.category, args: (uint8_t)Arg.sign, |
3433 | args: Arg.semantics->precision, args: Arg.exponent, |
3434 | args: hash_combine_range( |
3435 | first: Arg.significandParts(), |
3436 | last: Arg.significandParts() + Arg.partCount())); |
3437 | } |
3438 | |
3439 | // Conversion from APFloat to/from host float/double. It may eventually be |
3440 | // possible to eliminate these and have everybody deal with APFloats, but that |
3441 | // will take a while. This approach will not easily extend to long double. |
3442 | // Current implementation requires integerPartWidth==64, which is correct at |
3443 | // the moment but could be made more general. |
3444 | |
3445 | // Denormals have exponent minExponent in APFloat, but minExponent-1 in |
3446 | // the actual IEEE respresentations. We compensate for that here. |
3447 | |
3448 | APInt IEEEFloat::convertF80LongDoubleAPFloatToAPInt() const { |
3449 | assert(semantics == (const llvm::fltSemantics*)&semX87DoubleExtended); |
3450 | assert(partCount()==2); |
3451 | |
3452 | uint64_t myexponent, mysignificand; |
3453 | |
3454 | if (isFiniteNonZero()) { |
3455 | myexponent = exponent+16383; //bias |
3456 | mysignificand = significandParts()[0]; |
3457 | if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL)) |
3458 | myexponent = 0; // denormal |
3459 | } else if (category==fcZero) { |
3460 | myexponent = 0; |
3461 | mysignificand = 0; |
3462 | } else if (category==fcInfinity) { |
3463 | myexponent = 0x7fff; |
3464 | mysignificand = 0x8000000000000000ULL; |
3465 | } else { |
3466 | assert(category == fcNaN && "Unknown category" ); |
3467 | myexponent = 0x7fff; |
3468 | mysignificand = significandParts()[0]; |
3469 | } |
3470 | |
3471 | uint64_t words[2]; |
3472 | words[0] = mysignificand; |
3473 | words[1] = ((uint64_t)(sign & 1) << 15) | |
3474 | (myexponent & 0x7fffLL); |
3475 | return APInt(80, words); |
3476 | } |
3477 | |
3478 | APInt IEEEFloat::convertPPCDoubleDoubleAPFloatToAPInt() const { |
3479 | assert(semantics == (const llvm::fltSemantics *)&semPPCDoubleDoubleLegacy); |
3480 | assert(partCount()==2); |
3481 | |
3482 | uint64_t words[2]; |
3483 | opStatus fs; |
3484 | bool losesInfo; |
3485 | |
3486 | // Convert number to double. To avoid spurious underflows, we re- |
3487 | // normalize against the "double" minExponent first, and only *then* |
3488 | // truncate the mantissa. The result of that second conversion |
3489 | // may be inexact, but should never underflow. |
3490 | // Declare fltSemantics before APFloat that uses it (and |
3491 | // saves pointer to it) to ensure correct destruction order. |
3492 | fltSemantics extendedSemantics = *semantics; |
3493 | extendedSemantics.minExponent = semIEEEdouble.minExponent; |
3494 | IEEEFloat extended(*this); |
3495 | fs = extended.convert(toSemantics: extendedSemantics, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3496 | assert(fs == opOK && !losesInfo); |
3497 | (void)fs; |
3498 | |
3499 | IEEEFloat u(extended); |
3500 | fs = u.convert(toSemantics: semIEEEdouble, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3501 | assert(fs == opOK || fs == opInexact); |
3502 | (void)fs; |
3503 | words[0] = *u.convertDoubleAPFloatToAPInt().getRawData(); |
3504 | |
3505 | // If conversion was exact or resulted in a special case, we're done; |
3506 | // just set the second double to zero. Otherwise, re-convert back to |
3507 | // the extended format and compute the difference. This now should |
3508 | // convert exactly to double. |
3509 | if (u.isFiniteNonZero() && losesInfo) { |
3510 | fs = u.convert(toSemantics: extendedSemantics, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3511 | assert(fs == opOK && !losesInfo); |
3512 | (void)fs; |
3513 | |
3514 | IEEEFloat v(extended); |
3515 | v.subtract(rhs: u, rounding_mode: rmNearestTiesToEven); |
3516 | fs = v.convert(toSemantics: semIEEEdouble, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3517 | assert(fs == opOK && !losesInfo); |
3518 | (void)fs; |
3519 | words[1] = *v.convertDoubleAPFloatToAPInt().getRawData(); |
3520 | } else { |
3521 | words[1] = 0; |
3522 | } |
3523 | |
3524 | return APInt(128, words); |
3525 | } |
3526 | |
3527 | template <const fltSemantics &S> |
3528 | APInt IEEEFloat::convertIEEEFloatToAPInt() const { |
3529 | assert(semantics == &S); |
3530 | |
3531 | constexpr int bias = -(S.minExponent - 1); |
3532 | constexpr unsigned int trailing_significand_bits = S.precision - 1; |
3533 | constexpr int integer_bit_part = trailing_significand_bits / integerPartWidth; |
3534 | constexpr integerPart integer_bit = |
3535 | integerPart{1} << (trailing_significand_bits % integerPartWidth); |
3536 | constexpr uint64_t significand_mask = integer_bit - 1; |
3537 | constexpr unsigned int exponent_bits = |
3538 | S.sizeInBits - 1 - trailing_significand_bits; |
3539 | static_assert(exponent_bits < 64); |
3540 | constexpr uint64_t exponent_mask = (uint64_t{1} << exponent_bits) - 1; |
3541 | |
3542 | uint64_t myexponent; |
3543 | std::array<integerPart, partCountForBits(bits: trailing_significand_bits)> |
3544 | mysignificand; |
3545 | |
3546 | if (isFiniteNonZero()) { |
3547 | myexponent = exponent + bias; |
3548 | std::copy_n(significandParts(), mysignificand.size(), |
3549 | mysignificand.begin()); |
3550 | if (myexponent == 1 && |
3551 | !(significandParts()[integer_bit_part] & integer_bit)) |
3552 | myexponent = 0; // denormal |
3553 | } else if (category == fcZero) { |
3554 | myexponent = ::exponentZero(semantics: S) + bias; |
3555 | mysignificand.fill(0); |
3556 | } else if (category == fcInfinity) { |
3557 | if (S.nonFiniteBehavior == fltNonfiniteBehavior::NanOnly || |
3558 | S.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
3559 | llvm_unreachable("semantics don't support inf!" ); |
3560 | myexponent = ::exponentInf(semantics: S) + bias; |
3561 | mysignificand.fill(0); |
3562 | } else { |
3563 | assert(category == fcNaN && "Unknown category!" ); |
3564 | if (S.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
3565 | llvm_unreachable("semantics don't support NaN!" ); |
3566 | myexponent = ::exponentNaN(semantics: S) + bias; |
3567 | std::copy_n(significandParts(), mysignificand.size(), |
3568 | mysignificand.begin()); |
3569 | } |
3570 | std::array<uint64_t, (S.sizeInBits + 63) / 64> words; |
3571 | auto words_iter = |
3572 | std::copy_n(mysignificand.begin(), mysignificand.size(), words.begin()); |
3573 | if constexpr (significand_mask != 0) { |
3574 | // Clear the integer bit. |
3575 | words[mysignificand.size() - 1] &= significand_mask; |
3576 | } |
3577 | std::fill(words_iter, words.end(), uint64_t{0}); |
3578 | constexpr size_t last_word = words.size() - 1; |
3579 | uint64_t shifted_sign = static_cast<uint64_t>(sign & 1) |
3580 | << ((S.sizeInBits - 1) % 64); |
3581 | words[last_word] |= shifted_sign; |
3582 | uint64_t shifted_exponent = (myexponent & exponent_mask) |
3583 | << (trailing_significand_bits % 64); |
3584 | words[last_word] |= shifted_exponent; |
3585 | if constexpr (last_word == 0) { |
3586 | return APInt(S.sizeInBits, words[0]); |
3587 | } |
3588 | return APInt(S.sizeInBits, words); |
3589 | } |
3590 | |
3591 | APInt IEEEFloat::convertQuadrupleAPFloatToAPInt() const { |
3592 | assert(partCount() == 2); |
3593 | return convertIEEEFloatToAPInt<semIEEEquad>(); |
3594 | } |
3595 | |
3596 | APInt IEEEFloat::convertDoubleAPFloatToAPInt() const { |
3597 | assert(partCount()==1); |
3598 | return convertIEEEFloatToAPInt<semIEEEdouble>(); |
3599 | } |
3600 | |
3601 | APInt IEEEFloat::convertFloatAPFloatToAPInt() const { |
3602 | assert(partCount()==1); |
3603 | return convertIEEEFloatToAPInt<semIEEEsingle>(); |
3604 | } |
3605 | |
3606 | APInt IEEEFloat::convertBFloatAPFloatToAPInt() const { |
3607 | assert(partCount() == 1); |
3608 | return convertIEEEFloatToAPInt<semBFloat>(); |
3609 | } |
3610 | |
3611 | APInt IEEEFloat::convertHalfAPFloatToAPInt() const { |
3612 | assert(partCount()==1); |
3613 | return convertIEEEFloatToAPInt<semIEEEhalf>(); |
3614 | } |
3615 | |
3616 | APInt IEEEFloat::convertFloat8E5M2APFloatToAPInt() const { |
3617 | assert(partCount() == 1); |
3618 | return convertIEEEFloatToAPInt<semFloat8E5M2>(); |
3619 | } |
3620 | |
3621 | APInt IEEEFloat::convertFloat8E5M2FNUZAPFloatToAPInt() const { |
3622 | assert(partCount() == 1); |
3623 | return convertIEEEFloatToAPInt<semFloat8E5M2FNUZ>(); |
3624 | } |
3625 | |
3626 | APInt IEEEFloat::convertFloat8E4M3APFloatToAPInt() const { |
3627 | assert(partCount() == 1); |
3628 | return convertIEEEFloatToAPInt<semFloat8E4M3>(); |
3629 | } |
3630 | |
3631 | APInt IEEEFloat::convertFloat8E4M3FNAPFloatToAPInt() const { |
3632 | assert(partCount() == 1); |
3633 | return convertIEEEFloatToAPInt<semFloat8E4M3FN>(); |
3634 | } |
3635 | |
3636 | APInt IEEEFloat::convertFloat8E4M3FNUZAPFloatToAPInt() const { |
3637 | assert(partCount() == 1); |
3638 | return convertIEEEFloatToAPInt<semFloat8E4M3FNUZ>(); |
3639 | } |
3640 | |
3641 | APInt IEEEFloat::convertFloat8E4M3B11FNUZAPFloatToAPInt() const { |
3642 | assert(partCount() == 1); |
3643 | return convertIEEEFloatToAPInt<semFloat8E4M3B11FNUZ>(); |
3644 | } |
3645 | |
3646 | APInt IEEEFloat::convertFloatTF32APFloatToAPInt() const { |
3647 | assert(partCount() == 1); |
3648 | return convertIEEEFloatToAPInt<semFloatTF32>(); |
3649 | } |
3650 | |
3651 | APInt IEEEFloat::convertFloat6E3M2FNAPFloatToAPInt() const { |
3652 | assert(partCount() == 1); |
3653 | return convertIEEEFloatToAPInt<semFloat6E3M2FN>(); |
3654 | } |
3655 | |
3656 | APInt IEEEFloat::convertFloat6E2M3FNAPFloatToAPInt() const { |
3657 | assert(partCount() == 1); |
3658 | return convertIEEEFloatToAPInt<semFloat6E2M3FN>(); |
3659 | } |
3660 | |
3661 | APInt IEEEFloat::convertFloat4E2M1FNAPFloatToAPInt() const { |
3662 | assert(partCount() == 1); |
3663 | return convertIEEEFloatToAPInt<semFloat4E2M1FN>(); |
3664 | } |
3665 | |
3666 | // This function creates an APInt that is just a bit map of the floating |
3667 | // point constant as it would appear in memory. It is not a conversion, |
3668 | // and treating the result as a normal integer is unlikely to be useful. |
3669 | |
3670 | APInt IEEEFloat::bitcastToAPInt() const { |
3671 | if (semantics == (const llvm::fltSemantics*)&semIEEEhalf) |
3672 | return convertHalfAPFloatToAPInt(); |
3673 | |
3674 | if (semantics == (const llvm::fltSemantics *)&semBFloat) |
3675 | return convertBFloatAPFloatToAPInt(); |
3676 | |
3677 | if (semantics == (const llvm::fltSemantics*)&semIEEEsingle) |
3678 | return convertFloatAPFloatToAPInt(); |
3679 | |
3680 | if (semantics == (const llvm::fltSemantics*)&semIEEEdouble) |
3681 | return convertDoubleAPFloatToAPInt(); |
3682 | |
3683 | if (semantics == (const llvm::fltSemantics*)&semIEEEquad) |
3684 | return convertQuadrupleAPFloatToAPInt(); |
3685 | |
3686 | if (semantics == (const llvm::fltSemantics *)&semPPCDoubleDoubleLegacy) |
3687 | return convertPPCDoubleDoubleAPFloatToAPInt(); |
3688 | |
3689 | if (semantics == (const llvm::fltSemantics *)&semFloat8E5M2) |
3690 | return convertFloat8E5M2APFloatToAPInt(); |
3691 | |
3692 | if (semantics == (const llvm::fltSemantics *)&semFloat8E5M2FNUZ) |
3693 | return convertFloat8E5M2FNUZAPFloatToAPInt(); |
3694 | |
3695 | if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3) |
3696 | return convertFloat8E4M3APFloatToAPInt(); |
3697 | |
3698 | if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3FN) |
3699 | return convertFloat8E4M3FNAPFloatToAPInt(); |
3700 | |
3701 | if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3FNUZ) |
3702 | return convertFloat8E4M3FNUZAPFloatToAPInt(); |
3703 | |
3704 | if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3B11FNUZ) |
3705 | return convertFloat8E4M3B11FNUZAPFloatToAPInt(); |
3706 | |
3707 | if (semantics == (const llvm::fltSemantics *)&semFloatTF32) |
3708 | return convertFloatTF32APFloatToAPInt(); |
3709 | |
3710 | if (semantics == (const llvm::fltSemantics *)&semFloat6E3M2FN) |
3711 | return convertFloat6E3M2FNAPFloatToAPInt(); |
3712 | |
3713 | if (semantics == (const llvm::fltSemantics *)&semFloat6E2M3FN) |
3714 | return convertFloat6E2M3FNAPFloatToAPInt(); |
3715 | |
3716 | if (semantics == (const llvm::fltSemantics *)&semFloat4E2M1FN) |
3717 | return convertFloat4E2M1FNAPFloatToAPInt(); |
3718 | |
3719 | assert(semantics == (const llvm::fltSemantics*)&semX87DoubleExtended && |
3720 | "unknown format!" ); |
3721 | return convertF80LongDoubleAPFloatToAPInt(); |
3722 | } |
3723 | |
3724 | float IEEEFloat::convertToFloat() const { |
3725 | assert(semantics == (const llvm::fltSemantics*)&semIEEEsingle && |
3726 | "Float semantics are not IEEEsingle" ); |
3727 | APInt api = bitcastToAPInt(); |
3728 | return api.bitsToFloat(); |
3729 | } |
3730 | |
3731 | double IEEEFloat::convertToDouble() const { |
3732 | assert(semantics == (const llvm::fltSemantics*)&semIEEEdouble && |
3733 | "Float semantics are not IEEEdouble" ); |
3734 | APInt api = bitcastToAPInt(); |
3735 | return api.bitsToDouble(); |
3736 | } |
3737 | |
3738 | #ifdef HAS_IEE754_FLOAT128 |
3739 | float128 IEEEFloat::convertToQuad() const { |
3740 | assert(semantics == (const llvm::fltSemantics *)&semIEEEquad && |
3741 | "Float semantics are not IEEEquads" ); |
3742 | APInt api = bitcastToAPInt(); |
3743 | return api.bitsToQuad(); |
3744 | } |
3745 | #endif |
3746 | |
3747 | /// Integer bit is explicit in this format. Intel hardware (387 and later) |
3748 | /// does not support these bit patterns: |
3749 | /// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity") |
3750 | /// exponent = all 1's, integer bit 0, significand nonzero ("pseudoNaN") |
3751 | /// exponent!=0 nor all 1's, integer bit 0 ("unnormal") |
3752 | /// exponent = 0, integer bit 1 ("pseudodenormal") |
3753 | /// At the moment, the first three are treated as NaNs, the last one as Normal. |
3754 | void IEEEFloat::initFromF80LongDoubleAPInt(const APInt &api) { |
3755 | uint64_t i1 = api.getRawData()[0]; |
3756 | uint64_t i2 = api.getRawData()[1]; |
3757 | uint64_t myexponent = (i2 & 0x7fff); |
3758 | uint64_t mysignificand = i1; |
3759 | uint8_t myintegerbit = mysignificand >> 63; |
3760 | |
3761 | initialize(ourSemantics: &semX87DoubleExtended); |
3762 | assert(partCount()==2); |
3763 | |
3764 | sign = static_cast<unsigned int>(i2>>15); |
3765 | if (myexponent == 0 && mysignificand == 0) { |
3766 | makeZero(Neg: sign); |
3767 | } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) { |
3768 | makeInf(Neg: sign); |
3769 | } else if ((myexponent == 0x7fff && mysignificand != 0x8000000000000000ULL) || |
3770 | (myexponent != 0x7fff && myexponent != 0 && myintegerbit == 0)) { |
3771 | category = fcNaN; |
3772 | exponent = exponentNaN(); |
3773 | significandParts()[0] = mysignificand; |
3774 | significandParts()[1] = 0; |
3775 | } else { |
3776 | category = fcNormal; |
3777 | exponent = myexponent - 16383; |
3778 | significandParts()[0] = mysignificand; |
3779 | significandParts()[1] = 0; |
3780 | if (myexponent==0) // denormal |
3781 | exponent = -16382; |
3782 | } |
3783 | } |
3784 | |
3785 | void IEEEFloat::initFromPPCDoubleDoubleAPInt(const APInt &api) { |
3786 | uint64_t i1 = api.getRawData()[0]; |
3787 | uint64_t i2 = api.getRawData()[1]; |
3788 | opStatus fs; |
3789 | bool losesInfo; |
3790 | |
3791 | // Get the first double and convert to our format. |
3792 | initFromDoubleAPInt(api: APInt(64, i1)); |
3793 | fs = convert(toSemantics: semPPCDoubleDoubleLegacy, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3794 | assert(fs == opOK && !losesInfo); |
3795 | (void)fs; |
3796 | |
3797 | // Unless we have a special case, add in second double. |
3798 | if (isFiniteNonZero()) { |
3799 | IEEEFloat v(semIEEEdouble, APInt(64, i2)); |
3800 | fs = v.convert(toSemantics: semPPCDoubleDoubleLegacy, rounding_mode: rmNearestTiesToEven, losesInfo: &losesInfo); |
3801 | assert(fs == opOK && !losesInfo); |
3802 | (void)fs; |
3803 | |
3804 | add(rhs: v, rounding_mode: rmNearestTiesToEven); |
3805 | } |
3806 | } |
3807 | |
3808 | template <const fltSemantics &S> |
3809 | void IEEEFloat::initFromIEEEAPInt(const APInt &api) { |
3810 | assert(api.getBitWidth() == S.sizeInBits); |
3811 | constexpr integerPart integer_bit = integerPart{1} |
3812 | << ((S.precision - 1) % integerPartWidth); |
3813 | constexpr uint64_t significand_mask = integer_bit - 1; |
3814 | constexpr unsigned int trailing_significand_bits = S.precision - 1; |
3815 | constexpr unsigned int stored_significand_parts = |
3816 | partCountForBits(bits: trailing_significand_bits); |
3817 | constexpr unsigned int exponent_bits = |
3818 | S.sizeInBits - 1 - trailing_significand_bits; |
3819 | static_assert(exponent_bits < 64); |
3820 | constexpr uint64_t exponent_mask = (uint64_t{1} << exponent_bits) - 1; |
3821 | constexpr int bias = -(S.minExponent - 1); |
3822 | |
3823 | // Copy the bits of the significand. We need to clear out the exponent and |
3824 | // sign bit in the last word. |
3825 | std::array<integerPart, stored_significand_parts> mysignificand; |
3826 | std::copy_n(api.getRawData(), mysignificand.size(), mysignificand.begin()); |
3827 | if constexpr (significand_mask != 0) { |
3828 | mysignificand[mysignificand.size() - 1] &= significand_mask; |
3829 | } |
3830 | |
3831 | // We assume the last word holds the sign bit, the exponent, and potentially |
3832 | // some of the trailing significand field. |
3833 | uint64_t last_word = api.getRawData()[api.getNumWords() - 1]; |
3834 | uint64_t myexponent = |
3835 | (last_word >> (trailing_significand_bits % 64)) & exponent_mask; |
3836 | |
3837 | initialize(ourSemantics: &S); |
3838 | assert(partCount() == mysignificand.size()); |
3839 | |
3840 | sign = static_cast<unsigned int>(last_word >> ((S.sizeInBits - 1) % 64)); |
3841 | |
3842 | bool all_zero_significand = |
3843 | llvm::all_of(mysignificand, [](integerPart bits) { return bits == 0; }); |
3844 | |
3845 | bool is_zero = myexponent == 0 && all_zero_significand; |
3846 | |
3847 | if constexpr (S.nonFiniteBehavior == fltNonfiniteBehavior::IEEE754) { |
3848 | if (myexponent - bias == ::exponentInf(semantics: S) && all_zero_significand) { |
3849 | makeInf(Neg: sign); |
3850 | return; |
3851 | } |
3852 | } |
3853 | |
3854 | bool is_nan = false; |
3855 | |
3856 | if constexpr (S.nanEncoding == fltNanEncoding::IEEE) { |
3857 | is_nan = myexponent - bias == ::exponentNaN(semantics: S) && !all_zero_significand; |
3858 | } else if constexpr (S.nanEncoding == fltNanEncoding::AllOnes) { |
3859 | bool all_ones_significand = |
3860 | std::all_of(mysignificand.begin(), mysignificand.end() - 1, |
3861 | [](integerPart bits) { return bits == ~integerPart{0}; }) && |
3862 | (!significand_mask || |
3863 | mysignificand[mysignificand.size() - 1] == significand_mask); |
3864 | is_nan = myexponent - bias == ::exponentNaN(semantics: S) && all_ones_significand; |
3865 | } else if constexpr (S.nanEncoding == fltNanEncoding::NegativeZero) { |
3866 | is_nan = is_zero && sign; |
3867 | } |
3868 | |
3869 | if (is_nan) { |
3870 | category = fcNaN; |
3871 | exponent = ::exponentNaN(semantics: S); |
3872 | std::copy_n(mysignificand.begin(), mysignificand.size(), |
3873 | significandParts()); |
3874 | return; |
3875 | } |
3876 | |
3877 | if (is_zero) { |
3878 | makeZero(Neg: sign); |
3879 | return; |
3880 | } |
3881 | |
3882 | category = fcNormal; |
3883 | exponent = myexponent - bias; |
3884 | std::copy_n(mysignificand.begin(), mysignificand.size(), significandParts()); |
3885 | if (myexponent == 0) // denormal |
3886 | exponent = S.minExponent; |
3887 | else |
3888 | significandParts()[mysignificand.size()-1] |= integer_bit; // integer bit |
3889 | } |
3890 | |
3891 | void IEEEFloat::initFromQuadrupleAPInt(const APInt &api) { |
3892 | initFromIEEEAPInt<semIEEEquad>(api); |
3893 | } |
3894 | |
3895 | void IEEEFloat::initFromDoubleAPInt(const APInt &api) { |
3896 | initFromIEEEAPInt<semIEEEdouble>(api); |
3897 | } |
3898 | |
3899 | void IEEEFloat::initFromFloatAPInt(const APInt &api) { |
3900 | initFromIEEEAPInt<semIEEEsingle>(api); |
3901 | } |
3902 | |
3903 | void IEEEFloat::initFromBFloatAPInt(const APInt &api) { |
3904 | initFromIEEEAPInt<semBFloat>(api); |
3905 | } |
3906 | |
3907 | void IEEEFloat::initFromHalfAPInt(const APInt &api) { |
3908 | initFromIEEEAPInt<semIEEEhalf>(api); |
3909 | } |
3910 | |
3911 | void IEEEFloat::initFromFloat8E5M2APInt(const APInt &api) { |
3912 | initFromIEEEAPInt<semFloat8E5M2>(api); |
3913 | } |
3914 | |
3915 | void IEEEFloat::initFromFloat8E5M2FNUZAPInt(const APInt &api) { |
3916 | initFromIEEEAPInt<semFloat8E5M2FNUZ>(api); |
3917 | } |
3918 | |
3919 | void IEEEFloat::initFromFloat8E4M3APInt(const APInt &api) { |
3920 | initFromIEEEAPInt<semFloat8E4M3>(api); |
3921 | } |
3922 | |
3923 | void IEEEFloat::initFromFloat8E4M3FNAPInt(const APInt &api) { |
3924 | initFromIEEEAPInt<semFloat8E4M3FN>(api); |
3925 | } |
3926 | |
3927 | void IEEEFloat::initFromFloat8E4M3FNUZAPInt(const APInt &api) { |
3928 | initFromIEEEAPInt<semFloat8E4M3FNUZ>(api); |
3929 | } |
3930 | |
3931 | void IEEEFloat::initFromFloat8E4M3B11FNUZAPInt(const APInt &api) { |
3932 | initFromIEEEAPInt<semFloat8E4M3B11FNUZ>(api); |
3933 | } |
3934 | |
3935 | void IEEEFloat::initFromFloatTF32APInt(const APInt &api) { |
3936 | initFromIEEEAPInt<semFloatTF32>(api); |
3937 | } |
3938 | |
3939 | void IEEEFloat::initFromFloat6E3M2FNAPInt(const APInt &api) { |
3940 | initFromIEEEAPInt<semFloat6E3M2FN>(api); |
3941 | } |
3942 | |
3943 | void IEEEFloat::initFromFloat6E2M3FNAPInt(const APInt &api) { |
3944 | initFromIEEEAPInt<semFloat6E2M3FN>(api); |
3945 | } |
3946 | |
3947 | void IEEEFloat::initFromFloat4E2M1FNAPInt(const APInt &api) { |
3948 | initFromIEEEAPInt<semFloat4E2M1FN>(api); |
3949 | } |
3950 | |
3951 | /// Treat api as containing the bits of a floating point number. |
3952 | void IEEEFloat::initFromAPInt(const fltSemantics *Sem, const APInt &api) { |
3953 | assert(api.getBitWidth() == Sem->sizeInBits); |
3954 | if (Sem == &semIEEEhalf) |
3955 | return initFromHalfAPInt(api); |
3956 | if (Sem == &semBFloat) |
3957 | return initFromBFloatAPInt(api); |
3958 | if (Sem == &semIEEEsingle) |
3959 | return initFromFloatAPInt(api); |
3960 | if (Sem == &semIEEEdouble) |
3961 | return initFromDoubleAPInt(api); |
3962 | if (Sem == &semX87DoubleExtended) |
3963 | return initFromF80LongDoubleAPInt(api); |
3964 | if (Sem == &semIEEEquad) |
3965 | return initFromQuadrupleAPInt(api); |
3966 | if (Sem == &semPPCDoubleDoubleLegacy) |
3967 | return initFromPPCDoubleDoubleAPInt(api); |
3968 | if (Sem == &semFloat8E5M2) |
3969 | return initFromFloat8E5M2APInt(api); |
3970 | if (Sem == &semFloat8E5M2FNUZ) |
3971 | return initFromFloat8E5M2FNUZAPInt(api); |
3972 | if (Sem == &semFloat8E4M3) |
3973 | return initFromFloat8E4M3APInt(api); |
3974 | if (Sem == &semFloat8E4M3FN) |
3975 | return initFromFloat8E4M3FNAPInt(api); |
3976 | if (Sem == &semFloat8E4M3FNUZ) |
3977 | return initFromFloat8E4M3FNUZAPInt(api); |
3978 | if (Sem == &semFloat8E4M3B11FNUZ) |
3979 | return initFromFloat8E4M3B11FNUZAPInt(api); |
3980 | if (Sem == &semFloatTF32) |
3981 | return initFromFloatTF32APInt(api); |
3982 | if (Sem == &semFloat6E3M2FN) |
3983 | return initFromFloat6E3M2FNAPInt(api); |
3984 | if (Sem == &semFloat6E2M3FN) |
3985 | return initFromFloat6E2M3FNAPInt(api); |
3986 | if (Sem == &semFloat4E2M1FN) |
3987 | return initFromFloat4E2M1FNAPInt(api); |
3988 | |
3989 | llvm_unreachable(nullptr); |
3990 | } |
3991 | |
3992 | /// Make this number the largest magnitude normal number in the given |
3993 | /// semantics. |
3994 | void IEEEFloat::makeLargest(bool Negative) { |
3995 | // We want (in interchange format): |
3996 | // sign = {Negative} |
3997 | // exponent = 1..10 |
3998 | // significand = 1..1 |
3999 | category = fcNormal; |
4000 | sign = Negative; |
4001 | exponent = semantics->maxExponent; |
4002 | |
4003 | // Use memset to set all but the highest integerPart to all ones. |
4004 | integerPart *significand = significandParts(); |
4005 | unsigned PartCount = partCount(); |
4006 | memset(s: significand, c: 0xFF, n: sizeof(integerPart)*(PartCount - 1)); |
4007 | |
4008 | // Set the high integerPart especially setting all unused top bits for |
4009 | // internal consistency. |
4010 | const unsigned NumUnusedHighBits = |
4011 | PartCount*integerPartWidth - semantics->precision; |
4012 | significand[PartCount - 1] = (NumUnusedHighBits < integerPartWidth) |
4013 | ? (~integerPart(0) >> NumUnusedHighBits) |
4014 | : 0; |
4015 | |
4016 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly && |
4017 | semantics->nanEncoding == fltNanEncoding::AllOnes) |
4018 | significand[0] &= ~integerPart(1); |
4019 | } |
4020 | |
4021 | /// Make this number the smallest magnitude denormal number in the given |
4022 | /// semantics. |
4023 | void IEEEFloat::makeSmallest(bool Negative) { |
4024 | // We want (in interchange format): |
4025 | // sign = {Negative} |
4026 | // exponent = 0..0 |
4027 | // significand = 0..01 |
4028 | category = fcNormal; |
4029 | sign = Negative; |
4030 | exponent = semantics->minExponent; |
4031 | APInt::tcSet(significandParts(), 1, partCount()); |
4032 | } |
4033 | |
4034 | void IEEEFloat::makeSmallestNormalized(bool Negative) { |
4035 | // We want (in interchange format): |
4036 | // sign = {Negative} |
4037 | // exponent = 0..0 |
4038 | // significand = 10..0 |
4039 | |
4040 | category = fcNormal; |
4041 | zeroSignificand(); |
4042 | sign = Negative; |
4043 | exponent = semantics->minExponent; |
4044 | APInt::tcSetBit(significandParts(), bit: semantics->precision - 1); |
4045 | } |
4046 | |
4047 | IEEEFloat::IEEEFloat(const fltSemantics &Sem, const APInt &API) { |
4048 | initFromAPInt(Sem: &Sem, api: API); |
4049 | } |
4050 | |
4051 | IEEEFloat::IEEEFloat(float f) { |
4052 | initFromAPInt(Sem: &semIEEEsingle, api: APInt::floatToBits(V: f)); |
4053 | } |
4054 | |
4055 | IEEEFloat::IEEEFloat(double d) { |
4056 | initFromAPInt(Sem: &semIEEEdouble, api: APInt::doubleToBits(V: d)); |
4057 | } |
4058 | |
4059 | namespace { |
4060 | void append(SmallVectorImpl<char> &Buffer, StringRef Str) { |
4061 | Buffer.append(in_start: Str.begin(), in_end: Str.end()); |
4062 | } |
4063 | |
4064 | /// Removes data from the given significand until it is no more |
4065 | /// precise than is required for the desired precision. |
4066 | void AdjustToPrecision(APInt &significand, |
4067 | int &exp, unsigned FormatPrecision) { |
4068 | unsigned bits = significand.getActiveBits(); |
4069 | |
4070 | // 196/59 is a very slight overestimate of lg_2(10). |
4071 | unsigned bitsRequired = (FormatPrecision * 196 + 58) / 59; |
4072 | |
4073 | if (bits <= bitsRequired) return; |
4074 | |
4075 | unsigned tensRemovable = (bits - bitsRequired) * 59 / 196; |
4076 | if (!tensRemovable) return; |
4077 | |
4078 | exp += tensRemovable; |
4079 | |
4080 | APInt divisor(significand.getBitWidth(), 1); |
4081 | APInt powten(significand.getBitWidth(), 10); |
4082 | while (true) { |
4083 | if (tensRemovable & 1) |
4084 | divisor *= powten; |
4085 | tensRemovable >>= 1; |
4086 | if (!tensRemovable) break; |
4087 | powten *= powten; |
4088 | } |
4089 | |
4090 | significand = significand.udiv(RHS: divisor); |
4091 | |
4092 | // Truncate the significand down to its active bit count. |
4093 | significand = significand.trunc(width: significand.getActiveBits()); |
4094 | } |
4095 | |
4096 | |
4097 | void AdjustToPrecision(SmallVectorImpl<char> &buffer, |
4098 | int &exp, unsigned FormatPrecision) { |
4099 | unsigned N = buffer.size(); |
4100 | if (N <= FormatPrecision) return; |
4101 | |
4102 | // The most significant figures are the last ones in the buffer. |
4103 | unsigned FirstSignificant = N - FormatPrecision; |
4104 | |
4105 | // Round. |
4106 | // FIXME: this probably shouldn't use 'round half up'. |
4107 | |
4108 | // Rounding down is just a truncation, except we also want to drop |
4109 | // trailing zeros from the new result. |
4110 | if (buffer[FirstSignificant - 1] < '5') { |
4111 | while (FirstSignificant < N && buffer[FirstSignificant] == '0') |
4112 | FirstSignificant++; |
4113 | |
4114 | exp += FirstSignificant; |
4115 | buffer.erase(CS: &buffer[0], CE: &buffer[FirstSignificant]); |
4116 | return; |
4117 | } |
4118 | |
4119 | // Rounding up requires a decimal add-with-carry. If we continue |
4120 | // the carry, the newly-introduced zeros will just be truncated. |
4121 | for (unsigned I = FirstSignificant; I != N; ++I) { |
4122 | if (buffer[I] == '9') { |
4123 | FirstSignificant++; |
4124 | } else { |
4125 | buffer[I]++; |
4126 | break; |
4127 | } |
4128 | } |
4129 | |
4130 | // If we carried through, we have exactly one digit of precision. |
4131 | if (FirstSignificant == N) { |
4132 | exp += FirstSignificant; |
4133 | buffer.clear(); |
4134 | buffer.push_back(Elt: '1'); |
4135 | return; |
4136 | } |
4137 | |
4138 | exp += FirstSignificant; |
4139 | buffer.erase(CS: &buffer[0], CE: &buffer[FirstSignificant]); |
4140 | } |
4141 | |
4142 | void toStringImpl(SmallVectorImpl<char> &Str, const bool isNeg, int exp, |
4143 | APInt significand, unsigned FormatPrecision, |
4144 | unsigned FormatMaxPadding, bool TruncateZero) { |
4145 | const int semanticsPrecision = significand.getBitWidth(); |
4146 | |
4147 | if (isNeg) |
4148 | Str.push_back(Elt: '-'); |
4149 | |
4150 | // Set FormatPrecision if zero. We want to do this before we |
4151 | // truncate trailing zeros, as those are part of the precision. |
4152 | if (!FormatPrecision) { |
4153 | // We use enough digits so the number can be round-tripped back to an |
4154 | // APFloat. The formula comes from "How to Print Floating-Point Numbers |
4155 | // Accurately" by Steele and White. |
4156 | // FIXME: Using a formula based purely on the precision is conservative; |
4157 | // we can print fewer digits depending on the actual value being printed. |
4158 | |
4159 | // FormatPrecision = 2 + floor(significandBits / lg_2(10)) |
4160 | FormatPrecision = 2 + semanticsPrecision * 59 / 196; |
4161 | } |
4162 | |
4163 | // Ignore trailing binary zeros. |
4164 | int trailingZeros = significand.countr_zero(); |
4165 | exp += trailingZeros; |
4166 | significand.lshrInPlace(ShiftAmt: trailingZeros); |
4167 | |
4168 | // Change the exponent from 2^e to 10^e. |
4169 | if (exp == 0) { |
4170 | // Nothing to do. |
4171 | } else if (exp > 0) { |
4172 | // Just shift left. |
4173 | significand = significand.zext(width: semanticsPrecision + exp); |
4174 | significand <<= exp; |
4175 | exp = 0; |
4176 | } else { /* exp < 0 */ |
4177 | int texp = -exp; |
4178 | |
4179 | // We transform this using the identity: |
4180 | // (N)(2^-e) == (N)(5^e)(10^-e) |
4181 | // This means we have to multiply N (the significand) by 5^e. |
4182 | // To avoid overflow, we have to operate on numbers large |
4183 | // enough to store N * 5^e: |
4184 | // log2(N * 5^e) == log2(N) + e * log2(5) |
4185 | // <= semantics->precision + e * 137 / 59 |
4186 | // (log_2(5) ~ 2.321928 < 2.322034 ~ 137/59) |
4187 | |
4188 | unsigned precision = semanticsPrecision + (137 * texp + 136) / 59; |
4189 | |
4190 | // Multiply significand by 5^e. |
4191 | // N * 5^0101 == N * 5^(1*1) * 5^(0*2) * 5^(1*4) * 5^(0*8) |
4192 | significand = significand.zext(width: precision); |
4193 | APInt five_to_the_i(precision, 5); |
4194 | while (true) { |
4195 | if (texp & 1) |
4196 | significand *= five_to_the_i; |
4197 | |
4198 | texp >>= 1; |
4199 | if (!texp) |
4200 | break; |
4201 | five_to_the_i *= five_to_the_i; |
4202 | } |
4203 | } |
4204 | |
4205 | AdjustToPrecision(significand, exp, FormatPrecision); |
4206 | |
4207 | SmallVector<char, 256> buffer; |
4208 | |
4209 | // Fill the buffer. |
4210 | unsigned precision = significand.getBitWidth(); |
4211 | if (precision < 4) { |
4212 | // We need enough precision to store the value 10. |
4213 | precision = 4; |
4214 | significand = significand.zext(width: precision); |
4215 | } |
4216 | APInt ten(precision, 10); |
4217 | APInt digit(precision, 0); |
4218 | |
4219 | bool inTrail = true; |
4220 | while (significand != 0) { |
4221 | // digit <- significand % 10 |
4222 | // significand <- significand / 10 |
4223 | APInt::udivrem(LHS: significand, RHS: ten, Quotient&: significand, Remainder&: digit); |
4224 | |
4225 | unsigned d = digit.getZExtValue(); |
4226 | |
4227 | // Drop trailing zeros. |
4228 | if (inTrail && !d) |
4229 | exp++; |
4230 | else { |
4231 | buffer.push_back(Elt: (char) ('0' + d)); |
4232 | inTrail = false; |
4233 | } |
4234 | } |
4235 | |
4236 | assert(!buffer.empty() && "no characters in buffer!" ); |
4237 | |
4238 | // Drop down to FormatPrecision. |
4239 | // TODO: don't do more precise calculations above than are required. |
4240 | AdjustToPrecision(buffer, exp, FormatPrecision); |
4241 | |
4242 | unsigned NDigits = buffer.size(); |
4243 | |
4244 | // Check whether we should use scientific notation. |
4245 | bool FormatScientific; |
4246 | if (!FormatMaxPadding) |
4247 | FormatScientific = true; |
4248 | else { |
4249 | if (exp >= 0) { |
4250 | // 765e3 --> 765000 |
4251 | // ^^^ |
4252 | // But we shouldn't make the number look more precise than it is. |
4253 | FormatScientific = ((unsigned) exp > FormatMaxPadding || |
4254 | NDigits + (unsigned) exp > FormatPrecision); |
4255 | } else { |
4256 | // Power of the most significant digit. |
4257 | int MSD = exp + (int) (NDigits - 1); |
4258 | if (MSD >= 0) { |
4259 | // 765e-2 == 7.65 |
4260 | FormatScientific = false; |
4261 | } else { |
4262 | // 765e-5 == 0.00765 |
4263 | // ^ ^^ |
4264 | FormatScientific = ((unsigned) -MSD) > FormatMaxPadding; |
4265 | } |
4266 | } |
4267 | } |
4268 | |
4269 | // Scientific formatting is pretty straightforward. |
4270 | if (FormatScientific) { |
4271 | exp += (NDigits - 1); |
4272 | |
4273 | Str.push_back(Elt: buffer[NDigits-1]); |
4274 | Str.push_back(Elt: '.'); |
4275 | if (NDigits == 1 && TruncateZero) |
4276 | Str.push_back(Elt: '0'); |
4277 | else |
4278 | for (unsigned I = 1; I != NDigits; ++I) |
4279 | Str.push_back(Elt: buffer[NDigits-1-I]); |
4280 | // Fill with zeros up to FormatPrecision. |
4281 | if (!TruncateZero && FormatPrecision > NDigits - 1) |
4282 | Str.append(NumInputs: FormatPrecision - NDigits + 1, Elt: '0'); |
4283 | // For !TruncateZero we use lower 'e'. |
4284 | Str.push_back(Elt: TruncateZero ? 'E' : 'e'); |
4285 | |
4286 | Str.push_back(Elt: exp >= 0 ? '+' : '-'); |
4287 | if (exp < 0) |
4288 | exp = -exp; |
4289 | SmallVector<char, 6> expbuf; |
4290 | do { |
4291 | expbuf.push_back(Elt: (char) ('0' + (exp % 10))); |
4292 | exp /= 10; |
4293 | } while (exp); |
4294 | // Exponent always at least two digits if we do not truncate zeros. |
4295 | if (!TruncateZero && expbuf.size() < 2) |
4296 | expbuf.push_back(Elt: '0'); |
4297 | for (unsigned I = 0, E = expbuf.size(); I != E; ++I) |
4298 | Str.push_back(Elt: expbuf[E-1-I]); |
4299 | return; |
4300 | } |
4301 | |
4302 | // Non-scientific, positive exponents. |
4303 | if (exp >= 0) { |
4304 | for (unsigned I = 0; I != NDigits; ++I) |
4305 | Str.push_back(Elt: buffer[NDigits-1-I]); |
4306 | for (unsigned I = 0; I != (unsigned) exp; ++I) |
4307 | Str.push_back(Elt: '0'); |
4308 | return; |
4309 | } |
4310 | |
4311 | // Non-scientific, negative exponents. |
4312 | |
4313 | // The number of digits to the left of the decimal point. |
4314 | int NWholeDigits = exp + (int) NDigits; |
4315 | |
4316 | unsigned I = 0; |
4317 | if (NWholeDigits > 0) { |
4318 | for (; I != (unsigned) NWholeDigits; ++I) |
4319 | Str.push_back(Elt: buffer[NDigits-I-1]); |
4320 | Str.push_back(Elt: '.'); |
4321 | } else { |
4322 | unsigned NZeros = 1 + (unsigned) -NWholeDigits; |
4323 | |
4324 | Str.push_back(Elt: '0'); |
4325 | Str.push_back(Elt: '.'); |
4326 | for (unsigned Z = 1; Z != NZeros; ++Z) |
4327 | Str.push_back(Elt: '0'); |
4328 | } |
4329 | |
4330 | for (; I != NDigits; ++I) |
4331 | Str.push_back(Elt: buffer[NDigits-I-1]); |
4332 | |
4333 | } |
4334 | } // namespace |
4335 | |
4336 | void IEEEFloat::toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision, |
4337 | unsigned FormatMaxPadding, bool TruncateZero) const { |
4338 | switch (category) { |
4339 | case fcInfinity: |
4340 | if (isNegative()) |
4341 | return append(Buffer&: Str, Str: "-Inf" ); |
4342 | else |
4343 | return append(Buffer&: Str, Str: "+Inf" ); |
4344 | |
4345 | case fcNaN: return append(Buffer&: Str, Str: "NaN" ); |
4346 | |
4347 | case fcZero: |
4348 | if (isNegative()) |
4349 | Str.push_back(Elt: '-'); |
4350 | |
4351 | if (!FormatMaxPadding) { |
4352 | if (TruncateZero) |
4353 | append(Buffer&: Str, Str: "0.0E+0" ); |
4354 | else { |
4355 | append(Buffer&: Str, Str: "0.0" ); |
4356 | if (FormatPrecision > 1) |
4357 | Str.append(NumInputs: FormatPrecision - 1, Elt: '0'); |
4358 | append(Buffer&: Str, Str: "e+00" ); |
4359 | } |
4360 | } else |
4361 | Str.push_back(Elt: '0'); |
4362 | return; |
4363 | |
4364 | case fcNormal: |
4365 | break; |
4366 | } |
4367 | |
4368 | // Decompose the number into an APInt and an exponent. |
4369 | int exp = exponent - ((int) semantics->precision - 1); |
4370 | APInt significand( |
4371 | semantics->precision, |
4372 | ArrayRef(significandParts(), partCountForBits(bits: semantics->precision))); |
4373 | |
4374 | toStringImpl(Str, isNeg: isNegative(), exp, significand, FormatPrecision, |
4375 | FormatMaxPadding, TruncateZero); |
4376 | |
4377 | } |
4378 | |
4379 | bool IEEEFloat::getExactInverse(APFloat *inv) const { |
4380 | // Special floats and denormals have no exact inverse. |
4381 | if (!isFiniteNonZero()) |
4382 | return false; |
4383 | |
4384 | // Check that the number is a power of two by making sure that only the |
4385 | // integer bit is set in the significand. |
4386 | if (significandLSB() != semantics->precision - 1) |
4387 | return false; |
4388 | |
4389 | // Get the inverse. |
4390 | IEEEFloat reciprocal(*semantics, 1ULL); |
4391 | if (reciprocal.divide(rhs: *this, rounding_mode: rmNearestTiesToEven) != opOK) |
4392 | return false; |
4393 | |
4394 | // Avoid multiplication with a denormal, it is not safe on all platforms and |
4395 | // may be slower than a normal division. |
4396 | if (reciprocal.isDenormal()) |
4397 | return false; |
4398 | |
4399 | assert(reciprocal.isFiniteNonZero() && |
4400 | reciprocal.significandLSB() == reciprocal.semantics->precision - 1); |
4401 | |
4402 | if (inv) |
4403 | *inv = APFloat(reciprocal, *semantics); |
4404 | |
4405 | return true; |
4406 | } |
4407 | |
4408 | int IEEEFloat::getExactLog2Abs() const { |
4409 | if (!isFinite() || isZero()) |
4410 | return INT_MIN; |
4411 | |
4412 | const integerPart *Parts = significandParts(); |
4413 | const int PartCount = partCountForBits(bits: semantics->precision); |
4414 | |
4415 | int PopCount = 0; |
4416 | for (int i = 0; i < PartCount; ++i) { |
4417 | PopCount += llvm::popcount(Value: Parts[i]); |
4418 | if (PopCount > 1) |
4419 | return INT_MIN; |
4420 | } |
4421 | |
4422 | if (exponent != semantics->minExponent) |
4423 | return exponent; |
4424 | |
4425 | int CountrParts = 0; |
4426 | for (int i = 0; i < PartCount; |
4427 | ++i, CountrParts += APInt::APINT_BITS_PER_WORD) { |
4428 | if (Parts[i] != 0) { |
4429 | return exponent - semantics->precision + CountrParts + |
4430 | llvm::countr_zero(Val: Parts[i]) + 1; |
4431 | } |
4432 | } |
4433 | |
4434 | llvm_unreachable("didn't find the set bit" ); |
4435 | } |
4436 | |
4437 | bool IEEEFloat::isSignaling() const { |
4438 | if (!isNaN()) |
4439 | return false; |
4440 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly || |
4441 | semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
4442 | return false; |
4443 | |
4444 | // IEEE-754R 2008 6.2.1: A signaling NaN bit string should be encoded with the |
4445 | // first bit of the trailing significand being 0. |
4446 | return !APInt::tcExtractBit(significandParts(), bit: semantics->precision - 2); |
4447 | } |
4448 | |
4449 | /// IEEE-754R 2008 5.3.1: nextUp/nextDown. |
4450 | /// |
4451 | /// *NOTE* since nextDown(x) = -nextUp(-x), we only implement nextUp with |
4452 | /// appropriate sign switching before/after the computation. |
4453 | IEEEFloat::opStatus IEEEFloat::next(bool nextDown) { |
4454 | // If we are performing nextDown, swap sign so we have -x. |
4455 | if (nextDown) |
4456 | changeSign(); |
4457 | |
4458 | // Compute nextUp(x) |
4459 | opStatus result = opOK; |
4460 | |
4461 | // Handle each float category separately. |
4462 | switch (category) { |
4463 | case fcInfinity: |
4464 | // nextUp(+inf) = +inf |
4465 | if (!isNegative()) |
4466 | break; |
4467 | // nextUp(-inf) = -getLargest() |
4468 | makeLargest(Negative: true); |
4469 | break; |
4470 | case fcNaN: |
4471 | // IEEE-754R 2008 6.2 Par 2: nextUp(sNaN) = qNaN. Set Invalid flag. |
4472 | // IEEE-754R 2008 6.2: nextUp(qNaN) = qNaN. Must be identity so we do not |
4473 | // change the payload. |
4474 | if (isSignaling()) { |
4475 | result = opInvalidOp; |
4476 | // For consistency, propagate the sign of the sNaN to the qNaN. |
4477 | makeNaN(SNaN: false, Negative: isNegative(), fill: nullptr); |
4478 | } |
4479 | break; |
4480 | case fcZero: |
4481 | // nextUp(pm 0) = +getSmallest() |
4482 | makeSmallest(Negative: false); |
4483 | break; |
4484 | case fcNormal: |
4485 | // nextUp(-getSmallest()) = -0 |
4486 | if (isSmallest() && isNegative()) { |
4487 | APInt::tcSet(significandParts(), 0, partCount()); |
4488 | category = fcZero; |
4489 | exponent = 0; |
4490 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) |
4491 | sign = false; |
4492 | break; |
4493 | } |
4494 | |
4495 | if (isLargest() && !isNegative()) { |
4496 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
4497 | // nextUp(getLargest()) == NAN |
4498 | makeNaN(); |
4499 | break; |
4500 | } else if (semantics->nonFiniteBehavior == |
4501 | fltNonfiniteBehavior::FiniteOnly) { |
4502 | // nextUp(getLargest()) == getLargest() |
4503 | break; |
4504 | } else { |
4505 | // nextUp(getLargest()) == INFINITY |
4506 | APInt::tcSet(significandParts(), 0, partCount()); |
4507 | category = fcInfinity; |
4508 | exponent = semantics->maxExponent + 1; |
4509 | break; |
4510 | } |
4511 | } |
4512 | |
4513 | // nextUp(normal) == normal + inc. |
4514 | if (isNegative()) { |
4515 | // If we are negative, we need to decrement the significand. |
4516 | |
4517 | // We only cross a binade boundary that requires adjusting the exponent |
4518 | // if: |
4519 | // 1. exponent != semantics->minExponent. This implies we are not in the |
4520 | // smallest binade or are dealing with denormals. |
4521 | // 2. Our significand excluding the integral bit is all zeros. |
4522 | bool WillCrossBinadeBoundary = |
4523 | exponent != semantics->minExponent && isSignificandAllZeros(); |
4524 | |
4525 | // Decrement the significand. |
4526 | // |
4527 | // We always do this since: |
4528 | // 1. If we are dealing with a non-binade decrement, by definition we |
4529 | // just decrement the significand. |
4530 | // 2. If we are dealing with a normal -> normal binade decrement, since |
4531 | // we have an explicit integral bit the fact that all bits but the |
4532 | // integral bit are zero implies that subtracting one will yield a |
4533 | // significand with 0 integral bit and 1 in all other spots. Thus we |
4534 | // must just adjust the exponent and set the integral bit to 1. |
4535 | // 3. If we are dealing with a normal -> denormal binade decrement, |
4536 | // since we set the integral bit to 0 when we represent denormals, we |
4537 | // just decrement the significand. |
4538 | integerPart *Parts = significandParts(); |
4539 | APInt::tcDecrement(dst: Parts, parts: partCount()); |
4540 | |
4541 | if (WillCrossBinadeBoundary) { |
4542 | // Our result is a normal number. Do the following: |
4543 | // 1. Set the integral bit to 1. |
4544 | // 2. Decrement the exponent. |
4545 | APInt::tcSetBit(Parts, bit: semantics->precision - 1); |
4546 | exponent--; |
4547 | } |
4548 | } else { |
4549 | // If we are positive, we need to increment the significand. |
4550 | |
4551 | // We only cross a binade boundary that requires adjusting the exponent if |
4552 | // the input is not a denormal and all of said input's significand bits |
4553 | // are set. If all of said conditions are true: clear the significand, set |
4554 | // the integral bit to 1, and increment the exponent. If we have a |
4555 | // denormal always increment since moving denormals and the numbers in the |
4556 | // smallest normal binade have the same exponent in our representation. |
4557 | bool WillCrossBinadeBoundary = !isDenormal() && isSignificandAllOnes(); |
4558 | |
4559 | if (WillCrossBinadeBoundary) { |
4560 | integerPart *Parts = significandParts(); |
4561 | APInt::tcSet(Parts, 0, partCount()); |
4562 | APInt::tcSetBit(Parts, bit: semantics->precision - 1); |
4563 | assert(exponent != semantics->maxExponent && |
4564 | "We can not increment an exponent beyond the maxExponent allowed" |
4565 | " by the given floating point semantics." ); |
4566 | exponent++; |
4567 | } else { |
4568 | incrementSignificand(); |
4569 | } |
4570 | } |
4571 | break; |
4572 | } |
4573 | |
4574 | // If we are performing nextDown, swap sign so we have -nextUp(-x) |
4575 | if (nextDown) |
4576 | changeSign(); |
4577 | |
4578 | return result; |
4579 | } |
4580 | |
4581 | APFloatBase::ExponentType IEEEFloat::exponentNaN() const { |
4582 | return ::exponentNaN(semantics: *semantics); |
4583 | } |
4584 | |
4585 | APFloatBase::ExponentType IEEEFloat::exponentInf() const { |
4586 | return ::exponentInf(semantics: *semantics); |
4587 | } |
4588 | |
4589 | APFloatBase::ExponentType IEEEFloat::exponentZero() const { |
4590 | return ::exponentZero(semantics: *semantics); |
4591 | } |
4592 | |
4593 | void IEEEFloat::makeInf(bool Negative) { |
4594 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
4595 | llvm_unreachable("This floating point format does not support Inf" ); |
4596 | |
4597 | if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) { |
4598 | // There is no Inf, so make NaN instead. |
4599 | makeNaN(SNaN: false, Negative); |
4600 | return; |
4601 | } |
4602 | category = fcInfinity; |
4603 | sign = Negative; |
4604 | exponent = exponentInf(); |
4605 | APInt::tcSet(significandParts(), 0, partCount()); |
4606 | } |
4607 | |
4608 | void IEEEFloat::makeZero(bool Negative) { |
4609 | category = fcZero; |
4610 | sign = Negative; |
4611 | if (semantics->nanEncoding == fltNanEncoding::NegativeZero) { |
4612 | // Merge negative zero to positive because 0b10000...000 is used for NaN |
4613 | sign = false; |
4614 | } |
4615 | exponent = exponentZero(); |
4616 | APInt::tcSet(significandParts(), 0, partCount()); |
4617 | } |
4618 | |
4619 | void IEEEFloat::makeQuiet() { |
4620 | assert(isNaN()); |
4621 | if (semantics->nonFiniteBehavior != fltNonfiniteBehavior::NanOnly) |
4622 | APInt::tcSetBit(significandParts(), bit: semantics->precision - 2); |
4623 | } |
4624 | |
4625 | int ilogb(const IEEEFloat &Arg) { |
4626 | if (Arg.isNaN()) |
4627 | return IEEEFloat::IEK_NaN; |
4628 | if (Arg.isZero()) |
4629 | return IEEEFloat::IEK_Zero; |
4630 | if (Arg.isInfinity()) |
4631 | return IEEEFloat::IEK_Inf; |
4632 | if (!Arg.isDenormal()) |
4633 | return Arg.exponent; |
4634 | |
4635 | IEEEFloat Normalized(Arg); |
4636 | int SignificandBits = Arg.getSemantics().precision - 1; |
4637 | |
4638 | Normalized.exponent += SignificandBits; |
4639 | Normalized.normalize(rounding_mode: IEEEFloat::rmNearestTiesToEven, lost_fraction: lfExactlyZero); |
4640 | return Normalized.exponent - SignificandBits; |
4641 | } |
4642 | |
4643 | IEEEFloat scalbn(IEEEFloat X, int Exp, IEEEFloat::roundingMode RoundingMode) { |
4644 | auto MaxExp = X.getSemantics().maxExponent; |
4645 | auto MinExp = X.getSemantics().minExponent; |
4646 | |
4647 | // If Exp is wildly out-of-scale, simply adding it to X.exponent will |
4648 | // overflow; clamp it to a safe range before adding, but ensure that the range |
4649 | // is large enough that the clamp does not change the result. The range we |
4650 | // need to support is the difference between the largest possible exponent and |
4651 | // the normalized exponent of half the smallest denormal. |
4652 | |
4653 | int SignificandBits = X.getSemantics().precision - 1; |
4654 | int MaxIncrement = MaxExp - (MinExp - SignificandBits) + 1; |
4655 | |
4656 | // Clamp to one past the range ends to let normalize handle overlflow. |
4657 | X.exponent += std::clamp(val: Exp, lo: -MaxIncrement - 1, hi: MaxIncrement); |
4658 | X.normalize(rounding_mode: RoundingMode, lost_fraction: lfExactlyZero); |
4659 | if (X.isNaN()) |
4660 | X.makeQuiet(); |
4661 | return X; |
4662 | } |
4663 | |
4664 | IEEEFloat frexp(const IEEEFloat &Val, int &Exp, IEEEFloat::roundingMode RM) { |
4665 | Exp = ilogb(Arg: Val); |
4666 | |
4667 | // Quiet signalling nans. |
4668 | if (Exp == IEEEFloat::IEK_NaN) { |
4669 | IEEEFloat Quiet(Val); |
4670 | Quiet.makeQuiet(); |
4671 | return Quiet; |
4672 | } |
4673 | |
4674 | if (Exp == IEEEFloat::IEK_Inf) |
4675 | return Val; |
4676 | |
4677 | // 1 is added because frexp is defined to return a normalized fraction in |
4678 | // +/-[0.5, 1.0), rather than the usual +/-[1.0, 2.0). |
4679 | Exp = Exp == IEEEFloat::IEK_Zero ? 0 : Exp + 1; |
4680 | return scalbn(X: Val, Exp: -Exp, RoundingMode: RM); |
4681 | } |
4682 | |
4683 | DoubleAPFloat::DoubleAPFloat(const fltSemantics &S) |
4684 | : Semantics(&S), |
4685 | Floats(new APFloat[2]{APFloat(semIEEEdouble), APFloat(semIEEEdouble)}) { |
4686 | assert(Semantics == &semPPCDoubleDouble); |
4687 | } |
4688 | |
4689 | DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, uninitializedTag) |
4690 | : Semantics(&S), |
4691 | Floats(new APFloat[2]{APFloat(semIEEEdouble, uninitialized), |
4692 | APFloat(semIEEEdouble, uninitialized)}) { |
4693 | assert(Semantics == &semPPCDoubleDouble); |
4694 | } |
4695 | |
4696 | DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, integerPart I) |
4697 | : Semantics(&S), Floats(new APFloat[2]{APFloat(semIEEEdouble, I), |
4698 | APFloat(semIEEEdouble)}) { |
4699 | assert(Semantics == &semPPCDoubleDouble); |
4700 | } |
4701 | |
4702 | DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, const APInt &I) |
4703 | : Semantics(&S), |
4704 | Floats(new APFloat[2]{ |
4705 | APFloat(semIEEEdouble, APInt(64, I.getRawData()[0])), |
4706 | APFloat(semIEEEdouble, APInt(64, I.getRawData()[1]))}) { |
4707 | assert(Semantics == &semPPCDoubleDouble); |
4708 | } |
4709 | |
4710 | DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, APFloat &&First, |
4711 | APFloat &&Second) |
4712 | : Semantics(&S), |
4713 | Floats(new APFloat[2]{std::move(First), std::move(Second)}) { |
4714 | assert(Semantics == &semPPCDoubleDouble); |
4715 | assert(&Floats[0].getSemantics() == &semIEEEdouble); |
4716 | assert(&Floats[1].getSemantics() == &semIEEEdouble); |
4717 | } |
4718 | |
4719 | DoubleAPFloat::DoubleAPFloat(const DoubleAPFloat &RHS) |
4720 | : Semantics(RHS.Semantics), |
4721 | Floats(RHS.Floats ? new APFloat[2]{APFloat(RHS.Floats[0]), |
4722 | APFloat(RHS.Floats[1])} |
4723 | : nullptr) { |
4724 | assert(Semantics == &semPPCDoubleDouble); |
4725 | } |
4726 | |
4727 | DoubleAPFloat::DoubleAPFloat(DoubleAPFloat &&RHS) |
4728 | : Semantics(RHS.Semantics), Floats(std::move(RHS.Floats)) { |
4729 | RHS.Semantics = &semBogus; |
4730 | assert(Semantics == &semPPCDoubleDouble); |
4731 | } |
4732 | |
4733 | DoubleAPFloat &DoubleAPFloat::operator=(const DoubleAPFloat &RHS) { |
4734 | if (Semantics == RHS.Semantics && RHS.Floats) { |
4735 | Floats[0] = RHS.Floats[0]; |
4736 | Floats[1] = RHS.Floats[1]; |
4737 | } else if (this != &RHS) { |
4738 | this->~DoubleAPFloat(); |
4739 | new (this) DoubleAPFloat(RHS); |
4740 | } |
4741 | return *this; |
4742 | } |
4743 | |
4744 | // Implement addition, subtraction, multiplication and division based on: |
4745 | // "Software for Doubled-Precision Floating-Point Computations", |
4746 | // by Seppo Linnainmaa, ACM TOMS vol 7 no 3, September 1981, pages 272-283. |
4747 | APFloat::opStatus DoubleAPFloat::addImpl(const APFloat &a, const APFloat &aa, |
4748 | const APFloat &c, const APFloat &cc, |
4749 | roundingMode RM) { |
4750 | int Status = opOK; |
4751 | APFloat z = a; |
4752 | Status |= z.add(RHS: c, RM); |
4753 | if (!z.isFinite()) { |
4754 | if (!z.isInfinity()) { |
4755 | Floats[0] = std::move(z); |
4756 | Floats[1].makeZero(/* Neg = */ false); |
4757 | return (opStatus)Status; |
4758 | } |
4759 | Status = opOK; |
4760 | auto AComparedToC = a.compareAbsoluteValue(RHS: c); |
4761 | z = cc; |
4762 | Status |= z.add(RHS: aa, RM); |
4763 | if (AComparedToC == APFloat::cmpGreaterThan) { |
4764 | // z = cc + aa + c + a; |
4765 | Status |= z.add(RHS: c, RM); |
4766 | Status |= z.add(RHS: a, RM); |
4767 | } else { |
4768 | // z = cc + aa + a + c; |
4769 | Status |= z.add(RHS: a, RM); |
4770 | Status |= z.add(RHS: c, RM); |
4771 | } |
4772 | if (!z.isFinite()) { |
4773 | Floats[0] = std::move(z); |
4774 | Floats[1].makeZero(/* Neg = */ false); |
4775 | return (opStatus)Status; |
4776 | } |
4777 | Floats[0] = z; |
4778 | APFloat zz = aa; |
4779 | Status |= zz.add(RHS: cc, RM); |
4780 | if (AComparedToC == APFloat::cmpGreaterThan) { |
4781 | // Floats[1] = a - z + c + zz; |
4782 | Floats[1] = a; |
4783 | Status |= Floats[1].subtract(RHS: z, RM); |
4784 | Status |= Floats[1].add(RHS: c, RM); |
4785 | Status |= Floats[1].add(RHS: zz, RM); |
4786 | } else { |
4787 | // Floats[1] = c - z + a + zz; |
4788 | Floats[1] = c; |
4789 | Status |= Floats[1].subtract(RHS: z, RM); |
4790 | Status |= Floats[1].add(RHS: a, RM); |
4791 | Status |= Floats[1].add(RHS: zz, RM); |
4792 | } |
4793 | } else { |
4794 | // q = a - z; |
4795 | APFloat q = a; |
4796 | Status |= q.subtract(RHS: z, RM); |
4797 | |
4798 | // zz = q + c + (a - (q + z)) + aa + cc; |
4799 | // Compute a - (q + z) as -((q + z) - a) to avoid temporary copies. |
4800 | auto zz = q; |
4801 | Status |= zz.add(RHS: c, RM); |
4802 | Status |= q.add(RHS: z, RM); |
4803 | Status |= q.subtract(RHS: a, RM); |
4804 | q.changeSign(); |
4805 | Status |= zz.add(RHS: q, RM); |
4806 | Status |= zz.add(RHS: aa, RM); |
4807 | Status |= zz.add(RHS: cc, RM); |
4808 | if (zz.isZero() && !zz.isNegative()) { |
4809 | Floats[0] = std::move(z); |
4810 | Floats[1].makeZero(/* Neg = */ false); |
4811 | return opOK; |
4812 | } |
4813 | Floats[0] = z; |
4814 | Status |= Floats[0].add(RHS: zz, RM); |
4815 | if (!Floats[0].isFinite()) { |
4816 | Floats[1].makeZero(/* Neg = */ false); |
4817 | return (opStatus)Status; |
4818 | } |
4819 | Floats[1] = std::move(z); |
4820 | Status |= Floats[1].subtract(RHS: Floats[0], RM); |
4821 | Status |= Floats[1].add(RHS: zz, RM); |
4822 | } |
4823 | return (opStatus)Status; |
4824 | } |
4825 | |
4826 | APFloat::opStatus DoubleAPFloat::addWithSpecial(const DoubleAPFloat &LHS, |
4827 | const DoubleAPFloat &RHS, |
4828 | DoubleAPFloat &Out, |
4829 | roundingMode RM) { |
4830 | if (LHS.getCategory() == fcNaN) { |
4831 | Out = LHS; |
4832 | return opOK; |
4833 | } |
4834 | if (RHS.getCategory() == fcNaN) { |
4835 | Out = RHS; |
4836 | return opOK; |
4837 | } |
4838 | if (LHS.getCategory() == fcZero) { |
4839 | Out = RHS; |
4840 | return opOK; |
4841 | } |
4842 | if (RHS.getCategory() == fcZero) { |
4843 | Out = LHS; |
4844 | return opOK; |
4845 | } |
4846 | if (LHS.getCategory() == fcInfinity && RHS.getCategory() == fcInfinity && |
4847 | LHS.isNegative() != RHS.isNegative()) { |
4848 | Out.makeNaN(SNaN: false, Neg: Out.isNegative(), fill: nullptr); |
4849 | return opInvalidOp; |
4850 | } |
4851 | if (LHS.getCategory() == fcInfinity) { |
4852 | Out = LHS; |
4853 | return opOK; |
4854 | } |
4855 | if (RHS.getCategory() == fcInfinity) { |
4856 | Out = RHS; |
4857 | return opOK; |
4858 | } |
4859 | assert(LHS.getCategory() == fcNormal && RHS.getCategory() == fcNormal); |
4860 | |
4861 | APFloat A(LHS.Floats[0]), AA(LHS.Floats[1]), C(RHS.Floats[0]), |
4862 | CC(RHS.Floats[1]); |
4863 | assert(&A.getSemantics() == &semIEEEdouble); |
4864 | assert(&AA.getSemantics() == &semIEEEdouble); |
4865 | assert(&C.getSemantics() == &semIEEEdouble); |
4866 | assert(&CC.getSemantics() == &semIEEEdouble); |
4867 | assert(&Out.Floats[0].getSemantics() == &semIEEEdouble); |
4868 | assert(&Out.Floats[1].getSemantics() == &semIEEEdouble); |
4869 | return Out.addImpl(a: A, aa: AA, c: C, cc: CC, RM); |
4870 | } |
4871 | |
4872 | APFloat::opStatus DoubleAPFloat::add(const DoubleAPFloat &RHS, |
4873 | roundingMode RM) { |
4874 | return addWithSpecial(LHS: *this, RHS, Out&: *this, RM); |
4875 | } |
4876 | |
4877 | APFloat::opStatus DoubleAPFloat::subtract(const DoubleAPFloat &RHS, |
4878 | roundingMode RM) { |
4879 | changeSign(); |
4880 | auto Ret = add(RHS, RM); |
4881 | changeSign(); |
4882 | return Ret; |
4883 | } |
4884 | |
4885 | APFloat::opStatus DoubleAPFloat::multiply(const DoubleAPFloat &RHS, |
4886 | APFloat::roundingMode RM) { |
4887 | const auto &LHS = *this; |
4888 | auto &Out = *this; |
4889 | /* Interesting observation: For special categories, finding the lowest |
4890 | common ancestor of the following layered graph gives the correct |
4891 | return category: |
4892 | |
4893 | NaN |
4894 | / \ |
4895 | Zero Inf |
4896 | \ / |
4897 | Normal |
4898 | |
4899 | e.g. NaN * NaN = NaN |
4900 | Zero * Inf = NaN |
4901 | Normal * Zero = Zero |
4902 | Normal * Inf = Inf |
4903 | */ |
4904 | if (LHS.getCategory() == fcNaN) { |
4905 | Out = LHS; |
4906 | return opOK; |
4907 | } |
4908 | if (RHS.getCategory() == fcNaN) { |
4909 | Out = RHS; |
4910 | return opOK; |
4911 | } |
4912 | if ((LHS.getCategory() == fcZero && RHS.getCategory() == fcInfinity) || |
4913 | (LHS.getCategory() == fcInfinity && RHS.getCategory() == fcZero)) { |
4914 | Out.makeNaN(SNaN: false, Neg: false, fill: nullptr); |
4915 | return opOK; |
4916 | } |
4917 | if (LHS.getCategory() == fcZero || LHS.getCategory() == fcInfinity) { |
4918 | Out = LHS; |
4919 | return opOK; |
4920 | } |
4921 | if (RHS.getCategory() == fcZero || RHS.getCategory() == fcInfinity) { |
4922 | Out = RHS; |
4923 | return opOK; |
4924 | } |
4925 | assert(LHS.getCategory() == fcNormal && RHS.getCategory() == fcNormal && |
4926 | "Special cases not handled exhaustively" ); |
4927 | |
4928 | int Status = opOK; |
4929 | APFloat A = Floats[0], B = Floats[1], C = RHS.Floats[0], D = RHS.Floats[1]; |
4930 | // t = a * c |
4931 | APFloat T = A; |
4932 | Status |= T.multiply(RHS: C, RM); |
4933 | if (!T.isFiniteNonZero()) { |
4934 | Floats[0] = T; |
4935 | Floats[1].makeZero(/* Neg = */ false); |
4936 | return (opStatus)Status; |
4937 | } |
4938 | |
4939 | // tau = fmsub(a, c, t), that is -fmadd(-a, c, t). |
4940 | APFloat Tau = A; |
4941 | T.changeSign(); |
4942 | Status |= Tau.fusedMultiplyAdd(Multiplicand: C, Addend: T, RM); |
4943 | T.changeSign(); |
4944 | { |
4945 | // v = a * d |
4946 | APFloat V = A; |
4947 | Status |= V.multiply(RHS: D, RM); |
4948 | // w = b * c |
4949 | APFloat W = B; |
4950 | Status |= W.multiply(RHS: C, RM); |
4951 | Status |= V.add(RHS: W, RM); |
4952 | // tau += v + w |
4953 | Status |= Tau.add(RHS: V, RM); |
4954 | } |
4955 | // u = t + tau |
4956 | APFloat U = T; |
4957 | Status |= U.add(RHS: Tau, RM); |
4958 | |
4959 | Floats[0] = U; |
4960 | if (!U.isFinite()) { |
4961 | Floats[1].makeZero(/* Neg = */ false); |
4962 | } else { |
4963 | // Floats[1] = (t - u) + tau |
4964 | Status |= T.subtract(RHS: U, RM); |
4965 | Status |= T.add(RHS: Tau, RM); |
4966 | Floats[1] = T; |
4967 | } |
4968 | return (opStatus)Status; |
4969 | } |
4970 | |
4971 | APFloat::opStatus DoubleAPFloat::divide(const DoubleAPFloat &RHS, |
4972 | APFloat::roundingMode RM) { |
4973 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
4974 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
4975 | auto Ret = |
4976 | Tmp.divide(RHS: APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt()), RM); |
4977 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
4978 | return Ret; |
4979 | } |
4980 | |
4981 | APFloat::opStatus DoubleAPFloat::remainder(const DoubleAPFloat &RHS) { |
4982 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
4983 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
4984 | auto Ret = |
4985 | Tmp.remainder(RHS: APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt())); |
4986 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
4987 | return Ret; |
4988 | } |
4989 | |
4990 | APFloat::opStatus DoubleAPFloat::mod(const DoubleAPFloat &RHS) { |
4991 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
4992 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
4993 | auto Ret = Tmp.mod(RHS: APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt())); |
4994 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
4995 | return Ret; |
4996 | } |
4997 | |
4998 | APFloat::opStatus |
4999 | DoubleAPFloat::fusedMultiplyAdd(const DoubleAPFloat &Multiplicand, |
5000 | const DoubleAPFloat &Addend, |
5001 | APFloat::roundingMode RM) { |
5002 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5003 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
5004 | auto Ret = Tmp.fusedMultiplyAdd( |
5005 | Multiplicand: APFloat(semPPCDoubleDoubleLegacy, Multiplicand.bitcastToAPInt()), |
5006 | Addend: APFloat(semPPCDoubleDoubleLegacy, Addend.bitcastToAPInt()), RM); |
5007 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5008 | return Ret; |
5009 | } |
5010 | |
5011 | APFloat::opStatus DoubleAPFloat::roundToIntegral(APFloat::roundingMode RM) { |
5012 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5013 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
5014 | auto Ret = Tmp.roundToIntegral(RM); |
5015 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5016 | return Ret; |
5017 | } |
5018 | |
5019 | void DoubleAPFloat::changeSign() { |
5020 | Floats[0].changeSign(); |
5021 | Floats[1].changeSign(); |
5022 | } |
5023 | |
5024 | APFloat::cmpResult |
5025 | DoubleAPFloat::compareAbsoluteValue(const DoubleAPFloat &RHS) const { |
5026 | auto Result = Floats[0].compareAbsoluteValue(RHS: RHS.Floats[0]); |
5027 | if (Result != cmpEqual) |
5028 | return Result; |
5029 | Result = Floats[1].compareAbsoluteValue(RHS: RHS.Floats[1]); |
5030 | if (Result == cmpLessThan || Result == cmpGreaterThan) { |
5031 | auto Against = Floats[0].isNegative() ^ Floats[1].isNegative(); |
5032 | auto RHSAgainst = RHS.Floats[0].isNegative() ^ RHS.Floats[1].isNegative(); |
5033 | if (Against && !RHSAgainst) |
5034 | return cmpLessThan; |
5035 | if (!Against && RHSAgainst) |
5036 | return cmpGreaterThan; |
5037 | if (!Against && !RHSAgainst) |
5038 | return Result; |
5039 | if (Against && RHSAgainst) |
5040 | return (cmpResult)(cmpLessThan + cmpGreaterThan - Result); |
5041 | } |
5042 | return Result; |
5043 | } |
5044 | |
5045 | APFloat::fltCategory DoubleAPFloat::getCategory() const { |
5046 | return Floats[0].getCategory(); |
5047 | } |
5048 | |
5049 | bool DoubleAPFloat::isNegative() const { return Floats[0].isNegative(); } |
5050 | |
5051 | void DoubleAPFloat::makeInf(bool Neg) { |
5052 | Floats[0].makeInf(Neg); |
5053 | Floats[1].makeZero(/* Neg = */ false); |
5054 | } |
5055 | |
5056 | void DoubleAPFloat::makeZero(bool Neg) { |
5057 | Floats[0].makeZero(Neg); |
5058 | Floats[1].makeZero(/* Neg = */ false); |
5059 | } |
5060 | |
5061 | void DoubleAPFloat::makeLargest(bool Neg) { |
5062 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5063 | Floats[0] = APFloat(semIEEEdouble, APInt(64, 0x7fefffffffffffffull)); |
5064 | Floats[1] = APFloat(semIEEEdouble, APInt(64, 0x7c8ffffffffffffeull)); |
5065 | if (Neg) |
5066 | changeSign(); |
5067 | } |
5068 | |
5069 | void DoubleAPFloat::makeSmallest(bool Neg) { |
5070 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5071 | Floats[0].makeSmallest(Neg); |
5072 | Floats[1].makeZero(/* Neg = */ false); |
5073 | } |
5074 | |
5075 | void DoubleAPFloat::makeSmallestNormalized(bool Neg) { |
5076 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5077 | Floats[0] = APFloat(semIEEEdouble, APInt(64, 0x0360000000000000ull)); |
5078 | if (Neg) |
5079 | Floats[0].changeSign(); |
5080 | Floats[1].makeZero(/* Neg = */ false); |
5081 | } |
5082 | |
5083 | void DoubleAPFloat::makeNaN(bool SNaN, bool Neg, const APInt *fill) { |
5084 | Floats[0].makeNaN(SNaN, Neg, fill); |
5085 | Floats[1].makeZero(/* Neg = */ false); |
5086 | } |
5087 | |
5088 | APFloat::cmpResult DoubleAPFloat::compare(const DoubleAPFloat &RHS) const { |
5089 | auto Result = Floats[0].compare(RHS: RHS.Floats[0]); |
5090 | // |Float[0]| > |Float[1]| |
5091 | if (Result == APFloat::cmpEqual) |
5092 | return Floats[1].compare(RHS: RHS.Floats[1]); |
5093 | return Result; |
5094 | } |
5095 | |
5096 | bool DoubleAPFloat::bitwiseIsEqual(const DoubleAPFloat &RHS) const { |
5097 | return Floats[0].bitwiseIsEqual(RHS: RHS.Floats[0]) && |
5098 | Floats[1].bitwiseIsEqual(RHS: RHS.Floats[1]); |
5099 | } |
5100 | |
5101 | hash_code hash_value(const DoubleAPFloat &Arg) { |
5102 | if (Arg.Floats) |
5103 | return hash_combine(args: hash_value(Arg: Arg.Floats[0]), args: hash_value(Arg: Arg.Floats[1])); |
5104 | return hash_combine(args: Arg.Semantics); |
5105 | } |
5106 | |
5107 | APInt DoubleAPFloat::bitcastToAPInt() const { |
5108 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5109 | uint64_t Data[] = { |
5110 | Floats[0].bitcastToAPInt().getRawData()[0], |
5111 | Floats[1].bitcastToAPInt().getRawData()[0], |
5112 | }; |
5113 | return APInt(128, 2, Data); |
5114 | } |
5115 | |
5116 | Expected<APFloat::opStatus> DoubleAPFloat::convertFromString(StringRef S, |
5117 | roundingMode RM) { |
5118 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5119 | APFloat Tmp(semPPCDoubleDoubleLegacy); |
5120 | auto Ret = Tmp.convertFromString(S, RM); |
5121 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5122 | return Ret; |
5123 | } |
5124 | |
5125 | APFloat::opStatus DoubleAPFloat::next(bool nextDown) { |
5126 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5127 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
5128 | auto Ret = Tmp.next(nextDown); |
5129 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5130 | return Ret; |
5131 | } |
5132 | |
5133 | APFloat::opStatus |
5134 | DoubleAPFloat::convertToInteger(MutableArrayRef<integerPart> Input, |
5135 | unsigned int Width, bool IsSigned, |
5136 | roundingMode RM, bool *IsExact) const { |
5137 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5138 | return APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt()) |
5139 | .convertToInteger(Input, Width, IsSigned, RM, IsExact); |
5140 | } |
5141 | |
5142 | APFloat::opStatus DoubleAPFloat::convertFromAPInt(const APInt &Input, |
5143 | bool IsSigned, |
5144 | roundingMode RM) { |
5145 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5146 | APFloat Tmp(semPPCDoubleDoubleLegacy); |
5147 | auto Ret = Tmp.convertFromAPInt(Input, IsSigned, RM); |
5148 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5149 | return Ret; |
5150 | } |
5151 | |
5152 | APFloat::opStatus |
5153 | DoubleAPFloat::convertFromSignExtendedInteger(const integerPart *Input, |
5154 | unsigned int InputSize, |
5155 | bool IsSigned, roundingMode RM) { |
5156 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5157 | APFloat Tmp(semPPCDoubleDoubleLegacy); |
5158 | auto Ret = Tmp.convertFromSignExtendedInteger(Input, InputSize, IsSigned, RM); |
5159 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5160 | return Ret; |
5161 | } |
5162 | |
5163 | APFloat::opStatus |
5164 | DoubleAPFloat::convertFromZeroExtendedInteger(const integerPart *Input, |
5165 | unsigned int InputSize, |
5166 | bool IsSigned, roundingMode RM) { |
5167 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5168 | APFloat Tmp(semPPCDoubleDoubleLegacy); |
5169 | auto Ret = Tmp.convertFromZeroExtendedInteger(Input, InputSize, IsSigned, RM); |
5170 | *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt()); |
5171 | return Ret; |
5172 | } |
5173 | |
5174 | unsigned int DoubleAPFloat::convertToHexString(char *DST, |
5175 | unsigned int HexDigits, |
5176 | bool UpperCase, |
5177 | roundingMode RM) const { |
5178 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5179 | return APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt()) |
5180 | .convertToHexString(DST, HexDigits, UpperCase, RM); |
5181 | } |
5182 | |
5183 | bool DoubleAPFloat::isDenormal() const { |
5184 | return getCategory() == fcNormal && |
5185 | (Floats[0].isDenormal() || Floats[1].isDenormal() || |
5186 | // (double)(Hi + Lo) == Hi defines a normal number. |
5187 | Floats[0] != Floats[0] + Floats[1]); |
5188 | } |
5189 | |
5190 | bool DoubleAPFloat::isSmallest() const { |
5191 | if (getCategory() != fcNormal) |
5192 | return false; |
5193 | DoubleAPFloat Tmp(*this); |
5194 | Tmp.makeSmallest(Neg: this->isNegative()); |
5195 | return Tmp.compare(RHS: *this) == cmpEqual; |
5196 | } |
5197 | |
5198 | bool DoubleAPFloat::isSmallestNormalized() const { |
5199 | if (getCategory() != fcNormal) |
5200 | return false; |
5201 | |
5202 | DoubleAPFloat Tmp(*this); |
5203 | Tmp.makeSmallestNormalized(Neg: this->isNegative()); |
5204 | return Tmp.compare(RHS: *this) == cmpEqual; |
5205 | } |
5206 | |
5207 | bool DoubleAPFloat::isLargest() const { |
5208 | if (getCategory() != fcNormal) |
5209 | return false; |
5210 | DoubleAPFloat Tmp(*this); |
5211 | Tmp.makeLargest(Neg: this->isNegative()); |
5212 | return Tmp.compare(RHS: *this) == cmpEqual; |
5213 | } |
5214 | |
5215 | bool DoubleAPFloat::isInteger() const { |
5216 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5217 | return Floats[0].isInteger() && Floats[1].isInteger(); |
5218 | } |
5219 | |
5220 | void DoubleAPFloat::toString(SmallVectorImpl<char> &Str, |
5221 | unsigned FormatPrecision, |
5222 | unsigned FormatMaxPadding, |
5223 | bool TruncateZero) const { |
5224 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5225 | APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt()) |
5226 | .toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero); |
5227 | } |
5228 | |
5229 | bool DoubleAPFloat::getExactInverse(APFloat *inv) const { |
5230 | assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5231 | APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt()); |
5232 | if (!inv) |
5233 | return Tmp.getExactInverse(inv: nullptr); |
5234 | APFloat Inv(semPPCDoubleDoubleLegacy); |
5235 | auto Ret = Tmp.getExactInverse(inv: &Inv); |
5236 | *inv = APFloat(semPPCDoubleDouble, Inv.bitcastToAPInt()); |
5237 | return Ret; |
5238 | } |
5239 | |
5240 | int DoubleAPFloat::getExactLog2() const { |
5241 | // TODO: Implement me |
5242 | return INT_MIN; |
5243 | } |
5244 | |
5245 | int DoubleAPFloat::getExactLog2Abs() const { |
5246 | // TODO: Implement me |
5247 | return INT_MIN; |
5248 | } |
5249 | |
5250 | DoubleAPFloat scalbn(const DoubleAPFloat &Arg, int Exp, |
5251 | APFloat::roundingMode RM) { |
5252 | assert(Arg.Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5253 | return DoubleAPFloat(semPPCDoubleDouble, scalbn(X: Arg.Floats[0], Exp, RM), |
5254 | scalbn(X: Arg.Floats[1], Exp, RM)); |
5255 | } |
5256 | |
5257 | DoubleAPFloat frexp(const DoubleAPFloat &Arg, int &Exp, |
5258 | APFloat::roundingMode RM) { |
5259 | assert(Arg.Semantics == &semPPCDoubleDouble && "Unexpected Semantics" ); |
5260 | APFloat First = frexp(X: Arg.Floats[0], Exp, RM); |
5261 | APFloat Second = Arg.Floats[1]; |
5262 | if (Arg.getCategory() == APFloat::fcNormal) |
5263 | Second = scalbn(X: Second, Exp: -Exp, RM); |
5264 | return DoubleAPFloat(semPPCDoubleDouble, std::move(First), std::move(Second)); |
5265 | } |
5266 | |
5267 | } // namespace detail |
5268 | |
5269 | APFloat::Storage::Storage(IEEEFloat F, const fltSemantics &Semantics) { |
5270 | if (usesLayout<IEEEFloat>(Semantics)) { |
5271 | new (&IEEE) IEEEFloat(std::move(F)); |
5272 | return; |
5273 | } |
5274 | if (usesLayout<DoubleAPFloat>(Semantics)) { |
5275 | const fltSemantics& S = F.getSemantics(); |
5276 | new (&Double) |
5277 | DoubleAPFloat(Semantics, APFloat(std::move(F), S), |
5278 | APFloat(semIEEEdouble)); |
5279 | return; |
5280 | } |
5281 | llvm_unreachable("Unexpected semantics" ); |
5282 | } |
5283 | |
5284 | Expected<APFloat::opStatus> APFloat::convertFromString(StringRef Str, |
5285 | roundingMode RM) { |
5286 | APFLOAT_DISPATCH_ON_SEMANTICS(convertFromString(Str, RM)); |
5287 | } |
5288 | |
5289 | hash_code hash_value(const APFloat &Arg) { |
5290 | if (APFloat::usesLayout<detail::IEEEFloat>(Semantics: Arg.getSemantics())) |
5291 | return hash_value(Arg: Arg.U.IEEE); |
5292 | if (APFloat::usesLayout<detail::DoubleAPFloat>(Semantics: Arg.getSemantics())) |
5293 | return hash_value(Arg: Arg.U.Double); |
5294 | llvm_unreachable("Unexpected semantics" ); |
5295 | } |
5296 | |
5297 | APFloat::APFloat(const fltSemantics &Semantics, StringRef S) |
5298 | : APFloat(Semantics) { |
5299 | auto StatusOrErr = convertFromString(Str: S, RM: rmNearestTiesToEven); |
5300 | assert(StatusOrErr && "Invalid floating point representation" ); |
5301 | consumeError(Err: StatusOrErr.takeError()); |
5302 | } |
5303 | |
5304 | FPClassTest APFloat::classify() const { |
5305 | if (isZero()) |
5306 | return isNegative() ? fcNegZero : fcPosZero; |
5307 | if (isNormal()) |
5308 | return isNegative() ? fcNegNormal : fcPosNormal; |
5309 | if (isDenormal()) |
5310 | return isNegative() ? fcNegSubnormal : fcPosSubnormal; |
5311 | if (isInfinity()) |
5312 | return isNegative() ? fcNegInf : fcPosInf; |
5313 | assert(isNaN() && "Other class of FP constant" ); |
5314 | return isSignaling() ? fcSNan : fcQNan; |
5315 | } |
5316 | |
5317 | APFloat::opStatus APFloat::convert(const fltSemantics &ToSemantics, |
5318 | roundingMode RM, bool *losesInfo) { |
5319 | if (&getSemantics() == &ToSemantics) { |
5320 | *losesInfo = false; |
5321 | return opOK; |
5322 | } |
5323 | if (usesLayout<IEEEFloat>(Semantics: getSemantics()) && |
5324 | usesLayout<IEEEFloat>(Semantics: ToSemantics)) |
5325 | return U.IEEE.convert(toSemantics: ToSemantics, rounding_mode: RM, losesInfo); |
5326 | if (usesLayout<IEEEFloat>(Semantics: getSemantics()) && |
5327 | usesLayout<DoubleAPFloat>(Semantics: ToSemantics)) { |
5328 | assert(&ToSemantics == &semPPCDoubleDouble); |
5329 | auto Ret = U.IEEE.convert(toSemantics: semPPCDoubleDoubleLegacy, rounding_mode: RM, losesInfo); |
5330 | *this = APFloat(ToSemantics, U.IEEE.bitcastToAPInt()); |
5331 | return Ret; |
5332 | } |
5333 | if (usesLayout<DoubleAPFloat>(Semantics: getSemantics()) && |
5334 | usesLayout<IEEEFloat>(Semantics: ToSemantics)) { |
5335 | auto Ret = getIEEE().convert(toSemantics: ToSemantics, rounding_mode: RM, losesInfo); |
5336 | *this = APFloat(std::move(getIEEE()), ToSemantics); |
5337 | return Ret; |
5338 | } |
5339 | llvm_unreachable("Unexpected semantics" ); |
5340 | } |
5341 | |
5342 | APFloat APFloat::getAllOnesValue(const fltSemantics &Semantics) { |
5343 | return APFloat(Semantics, APInt::getAllOnes(numBits: Semantics.sizeInBits)); |
5344 | } |
5345 | |
5346 | void APFloat::print(raw_ostream &OS) const { |
5347 | SmallVector<char, 16> Buffer; |
5348 | toString(Str&: Buffer); |
5349 | OS << Buffer << "\n" ; |
5350 | } |
5351 | |
5352 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
5353 | LLVM_DUMP_METHOD void APFloat::dump() const { print(dbgs()); } |
5354 | #endif |
5355 | |
5356 | void APFloat::Profile(FoldingSetNodeID &NID) const { |
5357 | NID.Add(x: bitcastToAPInt()); |
5358 | } |
5359 | |
5360 | /* Same as convertToInteger(integerPart*, ...), except the result is returned in |
5361 | an APSInt, whose initial bit-width and signed-ness are used to determine the |
5362 | precision of the conversion. |
5363 | */ |
5364 | APFloat::opStatus APFloat::convertToInteger(APSInt &result, |
5365 | roundingMode rounding_mode, |
5366 | bool *isExact) const { |
5367 | unsigned bitWidth = result.getBitWidth(); |
5368 | SmallVector<uint64_t, 4> parts(result.getNumWords()); |
5369 | opStatus status = convertToInteger(Input: parts, Width: bitWidth, IsSigned: result.isSigned(), |
5370 | RM: rounding_mode, IsExact: isExact); |
5371 | // Keeps the original signed-ness. |
5372 | result = APInt(bitWidth, parts); |
5373 | return status; |
5374 | } |
5375 | |
5376 | double APFloat::convertToDouble() const { |
5377 | if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEdouble) |
5378 | return getIEEE().convertToDouble(); |
5379 | assert(getSemantics().isRepresentableBy(semIEEEdouble) && |
5380 | "Float semantics is not representable by IEEEdouble" ); |
5381 | APFloat Temp = *this; |
5382 | bool LosesInfo; |
5383 | opStatus St = Temp.convert(ToSemantics: semIEEEdouble, RM: rmNearestTiesToEven, losesInfo: &LosesInfo); |
5384 | assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision" ); |
5385 | (void)St; |
5386 | return Temp.getIEEE().convertToDouble(); |
5387 | } |
5388 | |
5389 | #ifdef HAS_IEE754_FLOAT128 |
5390 | float128 APFloat::convertToQuad() const { |
5391 | if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEquad) |
5392 | return getIEEE().convertToQuad(); |
5393 | assert(getSemantics().isRepresentableBy(semIEEEquad) && |
5394 | "Float semantics is not representable by IEEEquad" ); |
5395 | APFloat Temp = *this; |
5396 | bool LosesInfo; |
5397 | opStatus St = Temp.convert(ToSemantics: semIEEEquad, RM: rmNearestTiesToEven, losesInfo: &LosesInfo); |
5398 | assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision" ); |
5399 | (void)St; |
5400 | return Temp.getIEEE().convertToQuad(); |
5401 | } |
5402 | #endif |
5403 | |
5404 | float APFloat::convertToFloat() const { |
5405 | if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEsingle) |
5406 | return getIEEE().convertToFloat(); |
5407 | assert(getSemantics().isRepresentableBy(semIEEEsingle) && |
5408 | "Float semantics is not representable by IEEEsingle" ); |
5409 | APFloat Temp = *this; |
5410 | bool LosesInfo; |
5411 | opStatus St = Temp.convert(ToSemantics: semIEEEsingle, RM: rmNearestTiesToEven, losesInfo: &LosesInfo); |
5412 | assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision" ); |
5413 | (void)St; |
5414 | return Temp.getIEEE().convertToFloat(); |
5415 | } |
5416 | |
5417 | } // namespace llvm |
5418 | |
5419 | #undef APFLOAT_DISPATCH_ON_SEMANTICS |
5420 | |