[clang][bytecode][NFC] Only get expr when checking for UB (#125397)
[llvm-project.git] / llvm / lib / Support / APFloat.cpp
blobb0d92ae37fe8f608c88f23bacd5524a198dd885b
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.
12 //===----------------------------------------------------------------------===//
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>
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)
40 using namespace llvm;
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))
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!");
54 namespace llvm {
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,
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,
72 // This behavior is present in Float6E3M2FN, Float6E2M3FN, and
73 // Float4E2M1FN types, which do not support Inf or NaN values.
74 FiniteOnly,
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,
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,
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,
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;
108 /* The smallest E such that 2^E is a normalized number; this
109 matches the definition of IEEE 754. */
110 APFloatBase::ExponentType minExponent;
112 /* Number of bits in the significand. This includes the integer
113 bit. */
114 unsigned int precision;
116 /* Number of bits actually used in the semantics. */
117 unsigned int sizeInBits;
119 fltNonfiniteBehavior nonFiniteBehavior = fltNonfiniteBehavior::IEEE754;
121 fltNanEncoding nanEncoding = fltNanEncoding::IEEE;
123 /* Whether this semantics has an encoding for Zero */
124 bool hasZero = true;
126 /* Whether this semantics can represent signed values */
127 bool hasSignedRepr = true;
130 static constexpr fltSemantics semIEEEhalf = {15, -14, 11, 16};
131 static constexpr fltSemantics semBFloat = {127, -126, 8, 16};
132 static constexpr fltSemantics semIEEEsingle = {127, -126, 24, 32};
133 static constexpr fltSemantics semIEEEdouble = {1023, -1022, 53, 64};
134 static constexpr fltSemantics semIEEEquad = {16383, -16382, 113, 128};
135 static constexpr fltSemantics semFloat8E5M2 = {15, -14, 3, 8};
136 static constexpr fltSemantics semFloat8E5M2FNUZ = {
137 15, -15, 3, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::NegativeZero};
138 static constexpr fltSemantics semFloat8E4M3 = {7, -6, 4, 8};
139 static constexpr fltSemantics semFloat8E4M3FN = {
140 8, -6, 4, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::AllOnes};
141 static constexpr fltSemantics semFloat8E4M3FNUZ = {
142 7, -7, 4, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::NegativeZero};
143 static constexpr fltSemantics semFloat8E4M3B11FNUZ = {
144 4, -10, 4, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::NegativeZero};
145 static constexpr fltSemantics semFloat8E3M4 = {3, -2, 5, 8};
146 static constexpr fltSemantics semFloatTF32 = {127, -126, 11, 19};
147 static constexpr fltSemantics semFloat8E8M0FNU = {
148 127, -127, 1, 8, fltNonfiniteBehavior::NanOnly, fltNanEncoding::AllOnes,
149 false, false};
151 static constexpr fltSemantics semFloat6E3M2FN = {
152 4, -2, 3, 6, fltNonfiniteBehavior::FiniteOnly};
153 static constexpr fltSemantics semFloat6E2M3FN = {
154 2, 0, 4, 6, fltNonfiniteBehavior::FiniteOnly};
155 static constexpr fltSemantics semFloat4E2M1FN = {
156 2, 0, 2, 4, fltNonfiniteBehavior::FiniteOnly};
157 static constexpr fltSemantics semX87DoubleExtended = {16383, -16382, 64, 80};
158 static constexpr fltSemantics semBogus = {0, 0, 0, 0};
159 static constexpr fltSemantics semPPCDoubleDouble = {-1, 0, 0, 128};
160 static constexpr fltSemantics semPPCDoubleDoubleLegacy = {1023, -1022 + 53,
161 53 + 53, 128};
163 const llvm::fltSemantics &APFloatBase::EnumToSemantics(Semantics S) {
164 switch (S) {
165 case S_IEEEhalf:
166 return IEEEhalf();
167 case S_BFloat:
168 return BFloat();
169 case S_IEEEsingle:
170 return IEEEsingle();
171 case S_IEEEdouble:
172 return IEEEdouble();
173 case S_IEEEquad:
174 return IEEEquad();
175 case S_PPCDoubleDouble:
176 return PPCDoubleDouble();
177 case S_PPCDoubleDoubleLegacy:
178 return PPCDoubleDoubleLegacy();
179 case S_Float8E5M2:
180 return Float8E5M2();
181 case S_Float8E5M2FNUZ:
182 return Float8E5M2FNUZ();
183 case S_Float8E4M3:
184 return Float8E4M3();
185 case S_Float8E4M3FN:
186 return Float8E4M3FN();
187 case S_Float8E4M3FNUZ:
188 return Float8E4M3FNUZ();
189 case S_Float8E4M3B11FNUZ:
190 return Float8E4M3B11FNUZ();
191 case S_Float8E3M4:
192 return Float8E3M4();
193 case S_FloatTF32:
194 return FloatTF32();
195 case S_Float8E8M0FNU:
196 return Float8E8M0FNU();
197 case S_Float6E3M2FN:
198 return Float6E3M2FN();
199 case S_Float6E2M3FN:
200 return Float6E2M3FN();
201 case S_Float4E2M1FN:
202 return Float4E2M1FN();
203 case S_x87DoubleExtended:
204 return x87DoubleExtended();
206 llvm_unreachable("Unrecognised floating semantics");
209 APFloatBase::Semantics
210 APFloatBase::SemanticsToEnum(const llvm::fltSemantics &Sem) {
211 if (&Sem == &llvm::APFloat::IEEEhalf())
212 return S_IEEEhalf;
213 else if (&Sem == &llvm::APFloat::BFloat())
214 return S_BFloat;
215 else if (&Sem == &llvm::APFloat::IEEEsingle())
216 return S_IEEEsingle;
217 else if (&Sem == &llvm::APFloat::IEEEdouble())
218 return S_IEEEdouble;
219 else if (&Sem == &llvm::APFloat::IEEEquad())
220 return S_IEEEquad;
221 else if (&Sem == &llvm::APFloat::PPCDoubleDouble())
222 return S_PPCDoubleDouble;
223 else if (&Sem == &llvm::APFloat::PPCDoubleDoubleLegacy())
224 return S_PPCDoubleDoubleLegacy;
225 else if (&Sem == &llvm::APFloat::Float8E5M2())
226 return S_Float8E5M2;
227 else if (&Sem == &llvm::APFloat::Float8E5M2FNUZ())
228 return S_Float8E5M2FNUZ;
229 else if (&Sem == &llvm::APFloat::Float8E4M3())
230 return S_Float8E4M3;
231 else if (&Sem == &llvm::APFloat::Float8E4M3FN())
232 return S_Float8E4M3FN;
233 else if (&Sem == &llvm::APFloat::Float8E4M3FNUZ())
234 return S_Float8E4M3FNUZ;
235 else if (&Sem == &llvm::APFloat::Float8E4M3B11FNUZ())
236 return S_Float8E4M3B11FNUZ;
237 else if (&Sem == &llvm::APFloat::Float8E3M4())
238 return S_Float8E3M4;
239 else if (&Sem == &llvm::APFloat::FloatTF32())
240 return S_FloatTF32;
241 else if (&Sem == &llvm::APFloat::Float8E8M0FNU())
242 return S_Float8E8M0FNU;
243 else if (&Sem == &llvm::APFloat::Float6E3M2FN())
244 return S_Float6E3M2FN;
245 else if (&Sem == &llvm::APFloat::Float6E2M3FN())
246 return S_Float6E2M3FN;
247 else if (&Sem == &llvm::APFloat::Float4E2M1FN())
248 return S_Float4E2M1FN;
249 else if (&Sem == &llvm::APFloat::x87DoubleExtended())
250 return S_x87DoubleExtended;
251 else
252 llvm_unreachable("Unknown floating semantics");
255 const fltSemantics &APFloatBase::IEEEhalf() { return semIEEEhalf; }
256 const fltSemantics &APFloatBase::BFloat() { return semBFloat; }
257 const fltSemantics &APFloatBase::IEEEsingle() { return semIEEEsingle; }
258 const fltSemantics &APFloatBase::IEEEdouble() { return semIEEEdouble; }
259 const fltSemantics &APFloatBase::IEEEquad() { return semIEEEquad; }
260 const fltSemantics &APFloatBase::PPCDoubleDouble() {
261 return semPPCDoubleDouble;
263 const fltSemantics &APFloatBase::PPCDoubleDoubleLegacy() {
264 return semPPCDoubleDoubleLegacy;
266 const fltSemantics &APFloatBase::Float8E5M2() { return semFloat8E5M2; }
267 const fltSemantics &APFloatBase::Float8E5M2FNUZ() { return semFloat8E5M2FNUZ; }
268 const fltSemantics &APFloatBase::Float8E4M3() { return semFloat8E4M3; }
269 const fltSemantics &APFloatBase::Float8E4M3FN() { return semFloat8E4M3FN; }
270 const fltSemantics &APFloatBase::Float8E4M3FNUZ() { return semFloat8E4M3FNUZ; }
271 const fltSemantics &APFloatBase::Float8E4M3B11FNUZ() {
272 return semFloat8E4M3B11FNUZ;
274 const fltSemantics &APFloatBase::Float8E3M4() { return semFloat8E3M4; }
275 const fltSemantics &APFloatBase::FloatTF32() { return semFloatTF32; }
276 const fltSemantics &APFloatBase::Float8E8M0FNU() { return semFloat8E8M0FNU; }
277 const fltSemantics &APFloatBase::Float6E3M2FN() { return semFloat6E3M2FN; }
278 const fltSemantics &APFloatBase::Float6E2M3FN() { return semFloat6E2M3FN; }
279 const fltSemantics &APFloatBase::Float4E2M1FN() { return semFloat4E2M1FN; }
280 const fltSemantics &APFloatBase::x87DoubleExtended() {
281 return semX87DoubleExtended;
283 const fltSemantics &APFloatBase::Bogus() { return semBogus; }
285 bool APFloatBase::isRepresentableBy(const fltSemantics &A,
286 const fltSemantics &B) {
287 return A.maxExponent <= B.maxExponent && A.minExponent >= B.minExponent &&
288 A.precision <= B.precision;
291 constexpr RoundingMode APFloatBase::rmNearestTiesToEven;
292 constexpr RoundingMode APFloatBase::rmTowardPositive;
293 constexpr RoundingMode APFloatBase::rmTowardNegative;
294 constexpr RoundingMode APFloatBase::rmTowardZero;
295 constexpr RoundingMode APFloatBase::rmNearestTiesToAway;
297 /* A tight upper bound on number of parts required to hold the value
298 pow(5, power) is
300 power * 815 / (351 * integerPartWidth) + 1
302 However, whilst the result may require only this many parts,
303 because we are multiplying two values to get it, the
304 multiplication may require an extra part with the excess part
305 being zero (consider the trivial case of 1 * 1, tcFullMultiply
306 requires two parts to hold the single-part result). So we add an
307 extra one to guarantee enough space whilst multiplying. */
308 const unsigned int maxExponent = 16383;
309 const unsigned int maxPrecision = 113;
310 const unsigned int maxPowerOfFiveExponent = maxExponent + maxPrecision - 1;
311 const unsigned int maxPowerOfFiveParts =
313 ((maxPowerOfFiveExponent * 815) / (351 * APFloatBase::integerPartWidth));
315 unsigned int APFloatBase::semanticsPrecision(const fltSemantics &semantics) {
316 return semantics.precision;
318 APFloatBase::ExponentType
319 APFloatBase::semanticsMaxExponent(const fltSemantics &semantics) {
320 return semantics.maxExponent;
322 APFloatBase::ExponentType
323 APFloatBase::semanticsMinExponent(const fltSemantics &semantics) {
324 return semantics.minExponent;
326 unsigned int APFloatBase::semanticsSizeInBits(const fltSemantics &semantics) {
327 return semantics.sizeInBits;
329 unsigned int APFloatBase::semanticsIntSizeInBits(const fltSemantics &semantics,
330 bool isSigned) {
331 // The max FP value is pow(2, MaxExponent) * (1 + MaxFraction), so we need
332 // at least one more bit than the MaxExponent to hold the max FP value.
333 unsigned int MinBitWidth = semanticsMaxExponent(semantics) + 1;
334 // Extra sign bit needed.
335 if (isSigned)
336 ++MinBitWidth;
337 return MinBitWidth;
340 bool APFloatBase::semanticsHasZero(const fltSemantics &semantics) {
341 return semantics.hasZero;
344 bool APFloatBase::semanticsHasSignedRepr(const fltSemantics &semantics) {
345 return semantics.hasSignedRepr;
348 bool APFloatBase::semanticsHasInf(const fltSemantics &semantics) {
349 return semantics.nonFiniteBehavior == fltNonfiniteBehavior::IEEE754;
352 bool APFloatBase::semanticsHasNaN(const fltSemantics &semantics) {
353 return semantics.nonFiniteBehavior != fltNonfiniteBehavior::FiniteOnly;
356 bool APFloatBase::isRepresentableAsNormalIn(const fltSemantics &Src,
357 const fltSemantics &Dst) {
358 // Exponent range must be larger.
359 if (Src.maxExponent >= Dst.maxExponent || Src.minExponent <= Dst.minExponent)
360 return false;
362 // If the mantissa is long enough, the result value could still be denormal
363 // with a larger exponent range.
365 // FIXME: This condition is probably not accurate but also shouldn't be a
366 // practical concern with existing types.
367 return Dst.precision >= Src.precision;
370 unsigned APFloatBase::getSizeInBits(const fltSemantics &Sem) {
371 return Sem.sizeInBits;
374 static constexpr APFloatBase::ExponentType
375 exponentZero(const fltSemantics &semantics) {
376 return semantics.minExponent - 1;
379 static constexpr APFloatBase::ExponentType
380 exponentInf(const fltSemantics &semantics) {
381 return semantics.maxExponent + 1;
384 static constexpr APFloatBase::ExponentType
385 exponentNaN(const fltSemantics &semantics) {
386 if (semantics.nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
387 if (semantics.nanEncoding == fltNanEncoding::NegativeZero)
388 return exponentZero(semantics);
389 if (semantics.hasSignedRepr)
390 return semantics.maxExponent;
392 return semantics.maxExponent + 1;
395 /* A bunch of private, handy routines. */
397 static inline Error createError(const Twine &Err) {
398 return make_error<StringError>(Err, inconvertibleErrorCode());
401 static constexpr inline unsigned int partCountForBits(unsigned int bits) {
402 return std::max(1u, (bits + APFloatBase::integerPartWidth - 1) /
403 APFloatBase::integerPartWidth);
406 /* Returns 0U-9U. Return values >= 10U are not digits. */
407 static inline unsigned int
408 decDigitValue(unsigned int c)
410 return c - '0';
413 /* Return the value of a decimal exponent of the form
414 [+-]ddddddd.
416 If the exponent overflows, returns a large exponent with the
417 appropriate sign. */
418 static Expected<int> readExponent(StringRef::iterator begin,
419 StringRef::iterator end) {
420 bool isNegative;
421 unsigned int absExponent;
422 const unsigned int overlargeExponent = 24000; /* FIXME. */
423 StringRef::iterator p = begin;
425 // Treat no exponent as 0 to match binutils
426 if (p == end || ((*p == '-' || *p == '+') && (p + 1) == end)) {
427 return 0;
430 isNegative = (*p == '-');
431 if (*p == '-' || *p == '+') {
432 p++;
433 if (p == end)
434 return createError("Exponent has no digits");
437 absExponent = decDigitValue(*p++);
438 if (absExponent >= 10U)
439 return createError("Invalid character in exponent");
441 for (; p != end; ++p) {
442 unsigned int value;
444 value = decDigitValue(*p);
445 if (value >= 10U)
446 return createError("Invalid character in exponent");
448 absExponent = absExponent * 10U + value;
449 if (absExponent >= overlargeExponent) {
450 absExponent = overlargeExponent;
451 break;
455 if (isNegative)
456 return -(int) absExponent;
457 else
458 return (int) absExponent;
461 /* This is ugly and needs cleaning up, but I don't immediately see
462 how whilst remaining safe. */
463 static Expected<int> totalExponent(StringRef::iterator p,
464 StringRef::iterator end,
465 int exponentAdjustment) {
466 int unsignedExponent;
467 bool negative, overflow;
468 int exponent = 0;
470 if (p == end)
471 return createError("Exponent has no digits");
473 negative = *p == '-';
474 if (*p == '-' || *p == '+') {
475 p++;
476 if (p == end)
477 return createError("Exponent has no digits");
480 unsignedExponent = 0;
481 overflow = false;
482 for (; p != end; ++p) {
483 unsigned int value;
485 value = decDigitValue(*p);
486 if (value >= 10U)
487 return createError("Invalid character in exponent");
489 unsignedExponent = unsignedExponent * 10 + value;
490 if (unsignedExponent > 32767) {
491 overflow = true;
492 break;
496 if (exponentAdjustment > 32767 || exponentAdjustment < -32768)
497 overflow = true;
499 if (!overflow) {
500 exponent = unsignedExponent;
501 if (negative)
502 exponent = -exponent;
503 exponent += exponentAdjustment;
504 if (exponent > 32767 || exponent < -32768)
505 overflow = true;
508 if (overflow)
509 exponent = negative ? -32768: 32767;
511 return exponent;
514 static Expected<StringRef::iterator>
515 skipLeadingZeroesAndAnyDot(StringRef::iterator begin, StringRef::iterator end,
516 StringRef::iterator *dot) {
517 StringRef::iterator p = begin;
518 *dot = end;
519 while (p != end && *p == '0')
520 p++;
522 if (p != end && *p == '.') {
523 *dot = p++;
525 if (end - begin == 1)
526 return createError("Significand has no digits");
528 while (p != end && *p == '0')
529 p++;
532 return p;
535 /* Given a normal decimal floating point number of the form
537 dddd.dddd[eE][+-]ddd
539 where the decimal point and exponent are optional, fill out the
540 structure D. Exponent is appropriate if the significand is
541 treated as an integer, and normalizedExponent if the significand
542 is taken to have the decimal point after a single leading
543 non-zero digit.
545 If the value is zero, V->firstSigDigit points to a non-digit, and
546 the return exponent is zero.
548 struct decimalInfo {
549 const char *firstSigDigit;
550 const char *lastSigDigit;
551 int exponent;
552 int normalizedExponent;
555 static Error interpretDecimal(StringRef::iterator begin,
556 StringRef::iterator end, decimalInfo *D) {
557 StringRef::iterator dot = end;
559 auto PtrOrErr = skipLeadingZeroesAndAnyDot(begin, end, &dot);
560 if (!PtrOrErr)
561 return PtrOrErr.takeError();
562 StringRef::iterator p = *PtrOrErr;
564 D->firstSigDigit = p;
565 D->exponent = 0;
566 D->normalizedExponent = 0;
568 for (; p != end; ++p) {
569 if (*p == '.') {
570 if (dot != end)
571 return createError("String contains multiple dots");
572 dot = p++;
573 if (p == end)
574 break;
576 if (decDigitValue(*p) >= 10U)
577 break;
580 if (p != end) {
581 if (*p != 'e' && *p != 'E')
582 return createError("Invalid character in significand");
583 if (p == begin)
584 return createError("Significand has no digits");
585 if (dot != end && p - begin == 1)
586 return createError("Significand has no digits");
588 /* p points to the first non-digit in the string */
589 auto ExpOrErr = readExponent(p + 1, end);
590 if (!ExpOrErr)
591 return ExpOrErr.takeError();
592 D->exponent = *ExpOrErr;
594 /* Implied decimal point? */
595 if (dot == end)
596 dot = p;
599 /* If number is all zeroes accept any exponent. */
600 if (p != D->firstSigDigit) {
601 /* Drop insignificant trailing zeroes. */
602 if (p != begin) {
605 p--;
606 while (p != begin && *p == '0');
607 while (p != begin && *p == '.');
610 /* Adjust the exponents for any decimal point. */
611 D->exponent += static_cast<APFloat::ExponentType>((dot - p) - (dot > p));
612 D->normalizedExponent = (D->exponent +
613 static_cast<APFloat::ExponentType>((p - D->firstSigDigit)
614 - (dot > D->firstSigDigit && dot < p)));
617 D->lastSigDigit = p;
618 return Error::success();
621 /* Return the trailing fraction of a hexadecimal number.
622 DIGITVALUE is the first hex digit of the fraction, P points to
623 the next digit. */
624 static Expected<lostFraction>
625 trailingHexadecimalFraction(StringRef::iterator p, StringRef::iterator end,
626 unsigned int digitValue) {
627 unsigned int hexDigit;
629 /* If the first trailing digit isn't 0 or 8 we can work out the
630 fraction immediately. */
631 if (digitValue > 8)
632 return lfMoreThanHalf;
633 else if (digitValue < 8 && digitValue > 0)
634 return lfLessThanHalf;
636 // Otherwise we need to find the first non-zero digit.
637 while (p != end && (*p == '0' || *p == '.'))
638 p++;
640 if (p == end)
641 return createError("Invalid trailing hexadecimal fraction!");
643 hexDigit = hexDigitValue(*p);
645 /* If we ran off the end it is exactly zero or one-half, otherwise
646 a little more. */
647 if (hexDigit == UINT_MAX)
648 return digitValue == 0 ? lfExactlyZero: lfExactlyHalf;
649 else
650 return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf;
653 /* Return the fraction lost were a bignum truncated losing the least
654 significant BITS bits. */
655 static lostFraction
656 lostFractionThroughTruncation(const APFloatBase::integerPart *parts,
657 unsigned int partCount,
658 unsigned int bits)
660 unsigned int lsb;
662 lsb = APInt::tcLSB(parts, partCount);
664 /* Note this is guaranteed true if bits == 0, or LSB == UINT_MAX. */
665 if (bits <= lsb)
666 return lfExactlyZero;
667 if (bits == lsb + 1)
668 return lfExactlyHalf;
669 if (bits <= partCount * APFloatBase::integerPartWidth &&
670 APInt::tcExtractBit(parts, bits - 1))
671 return lfMoreThanHalf;
673 return lfLessThanHalf;
676 /* Shift DST right BITS bits noting lost fraction. */
677 static lostFraction
678 shiftRight(APFloatBase::integerPart *dst, unsigned int parts, unsigned int bits)
680 lostFraction lost_fraction;
682 lost_fraction = lostFractionThroughTruncation(dst, parts, bits);
684 APInt::tcShiftRight(dst, parts, bits);
686 return lost_fraction;
689 /* Combine the effect of two lost fractions. */
690 static lostFraction
691 combineLostFractions(lostFraction moreSignificant,
692 lostFraction lessSignificant)
694 if (lessSignificant != lfExactlyZero) {
695 if (moreSignificant == lfExactlyZero)
696 moreSignificant = lfLessThanHalf;
697 else if (moreSignificant == lfExactlyHalf)
698 moreSignificant = lfMoreThanHalf;
701 return moreSignificant;
704 /* The error from the true value, in half-ulps, on multiplying two
705 floating point numbers, which differ from the value they
706 approximate by at most HUE1 and HUE2 half-ulps, is strictly less
707 than the returned value.
709 See "How to Read Floating Point Numbers Accurately" by William D
710 Clinger. */
711 static unsigned int
712 HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2)
714 assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8));
716 if (HUerr1 + HUerr2 == 0)
717 return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */
718 else
719 return inexactMultiply + 2 * (HUerr1 + HUerr2);
722 /* The number of ulps from the boundary (zero, or half if ISNEAREST)
723 when the least significant BITS are truncated. BITS cannot be
724 zero. */
725 static APFloatBase::integerPart
726 ulpsFromBoundary(const APFloatBase::integerPart *parts, unsigned int bits,
727 bool isNearest) {
728 unsigned int count, partBits;
729 APFloatBase::integerPart part, boundary;
731 assert(bits != 0);
733 bits--;
734 count = bits / APFloatBase::integerPartWidth;
735 partBits = bits % APFloatBase::integerPartWidth + 1;
737 part = parts[count] & (~(APFloatBase::integerPart) 0 >> (APFloatBase::integerPartWidth - partBits));
739 if (isNearest)
740 boundary = (APFloatBase::integerPart) 1 << (partBits - 1);
741 else
742 boundary = 0;
744 if (count == 0) {
745 if (part - boundary <= boundary - part)
746 return part - boundary;
747 else
748 return boundary - part;
751 if (part == boundary) {
752 while (--count)
753 if (parts[count])
754 return ~(APFloatBase::integerPart) 0; /* A lot. */
756 return parts[0];
757 } else if (part == boundary - 1) {
758 while (--count)
759 if (~parts[count])
760 return ~(APFloatBase::integerPart) 0; /* A lot. */
762 return -parts[0];
765 return ~(APFloatBase::integerPart) 0; /* A lot. */
768 /* Place pow(5, power) in DST, and return the number of parts used.
769 DST must be at least one part larger than size of the answer. */
770 static unsigned int
771 powerOf5(APFloatBase::integerPart *dst, unsigned int power) {
772 static const APFloatBase::integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125, 15625, 78125 };
773 APFloatBase::integerPart pow5s[maxPowerOfFiveParts * 2 + 5];
774 pow5s[0] = 78125 * 5;
776 unsigned int partsCount = 1;
777 APFloatBase::integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5;
778 unsigned int result;
779 assert(power <= maxExponent);
781 p1 = dst;
782 p2 = scratch;
784 *p1 = firstEightPowers[power & 7];
785 power >>= 3;
787 result = 1;
788 pow5 = pow5s;
790 for (unsigned int n = 0; power; power >>= 1, n++) {
791 /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */
792 if (n != 0) {
793 APInt::tcFullMultiply(pow5, pow5 - partsCount, pow5 - partsCount,
794 partsCount, partsCount);
795 partsCount *= 2;
796 if (pow5[partsCount - 1] == 0)
797 partsCount--;
800 if (power & 1) {
801 APFloatBase::integerPart *tmp;
803 APInt::tcFullMultiply(p2, p1, pow5, result, partsCount);
804 result += partsCount;
805 if (p2[result - 1] == 0)
806 result--;
808 /* Now result is in p1 with partsCount parts and p2 is scratch
809 space. */
810 tmp = p1;
811 p1 = p2;
812 p2 = tmp;
815 pow5 += partsCount;
818 if (p1 != dst)
819 APInt::tcAssign(dst, p1, result);
821 return result;
824 /* Zero at the end to avoid modular arithmetic when adding one; used
825 when rounding up during hexadecimal output. */
826 static const char hexDigitsLower[] = "0123456789abcdef0";
827 static const char hexDigitsUpper[] = "0123456789ABCDEF0";
828 static const char infinityL[] = "infinity";
829 static const char infinityU[] = "INFINITY";
830 static const char NaNL[] = "nan";
831 static const char NaNU[] = "NAN";
833 /* Write out an integerPart in hexadecimal, starting with the most
834 significant nibble. Write out exactly COUNT hexdigits, return
835 COUNT. */
836 static unsigned int
837 partAsHex (char *dst, APFloatBase::integerPart part, unsigned int count,
838 const char *hexDigitChars)
840 unsigned int result = count;
842 assert(count != 0 && count <= APFloatBase::integerPartWidth / 4);
844 part >>= (APFloatBase::integerPartWidth - 4 * count);
845 while (count--) {
846 dst[count] = hexDigitChars[part & 0xf];
847 part >>= 4;
850 return result;
853 /* Write out an unsigned decimal integer. */
854 static char *
855 writeUnsignedDecimal (char *dst, unsigned int n)
857 char buff[40], *p;
859 p = buff;
861 *p++ = '0' + n % 10;
862 while (n /= 10);
865 *dst++ = *--p;
866 while (p != buff);
868 return dst;
871 /* Write out a signed decimal integer. */
872 static char *
873 writeSignedDecimal (char *dst, int value)
875 if (value < 0) {
876 *dst++ = '-';
877 dst = writeUnsignedDecimal(dst, -(unsigned) value);
878 } else
879 dst = writeUnsignedDecimal(dst, value);
881 return dst;
884 namespace detail {
885 /* Constructors. */
886 void IEEEFloat::initialize(const fltSemantics *ourSemantics) {
887 unsigned int count;
889 semantics = ourSemantics;
890 count = partCount();
891 if (count > 1)
892 significand.parts = new integerPart[count];
895 void IEEEFloat::freeSignificand() {
896 if (needsCleanup())
897 delete [] significand.parts;
900 void IEEEFloat::assign(const IEEEFloat &rhs) {
901 assert(semantics == rhs.semantics);
903 sign = rhs.sign;
904 category = rhs.category;
905 exponent = rhs.exponent;
906 if (isFiniteNonZero() || category == fcNaN)
907 copySignificand(rhs);
910 void IEEEFloat::copySignificand(const IEEEFloat &rhs) {
911 assert(isFiniteNonZero() || category == fcNaN);
912 assert(rhs.partCount() >= partCount());
914 APInt::tcAssign(significandParts(), rhs.significandParts(),
915 partCount());
918 /* Make this number a NaN, with an arbitrary but deterministic value
919 for the significand. If double or longer, this is a signalling NaN,
920 which may not be ideal. If float, this is QNaN(0). */
921 void IEEEFloat::makeNaN(bool SNaN, bool Negative, const APInt *fill) {
922 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)
923 llvm_unreachable("This floating point format does not support NaN");
925 if (Negative && !semantics->hasSignedRepr)
926 llvm_unreachable(
927 "This floating point format does not support signed values");
929 category = fcNaN;
930 sign = Negative;
931 exponent = exponentNaN();
933 integerPart *significand = significandParts();
934 unsigned numParts = partCount();
936 APInt fill_storage;
937 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
938 // Finite-only types do not distinguish signalling and quiet NaN, so
939 // make them all signalling.
940 SNaN = false;
941 if (semantics->nanEncoding == fltNanEncoding::NegativeZero) {
942 sign = true;
943 fill_storage = APInt::getZero(semantics->precision - 1);
944 } else {
945 fill_storage = APInt::getAllOnes(semantics->precision - 1);
947 fill = &fill_storage;
950 // Set the significand bits to the fill.
951 if (!fill || fill->getNumWords() < numParts)
952 APInt::tcSet(significand, 0, numParts);
953 if (fill) {
954 APInt::tcAssign(significand, fill->getRawData(),
955 std::min(fill->getNumWords(), numParts));
957 // Zero out the excess bits of the significand.
958 unsigned bitsToPreserve = semantics->precision - 1;
959 unsigned part = bitsToPreserve / 64;
960 bitsToPreserve %= 64;
961 significand[part] &= ((1ULL << bitsToPreserve) - 1);
962 for (part++; part != numParts; ++part)
963 significand[part] = 0;
966 unsigned QNaNBit =
967 (semantics->precision >= 2) ? (semantics->precision - 2) : 0;
969 if (SNaN) {
970 // We always have to clear the QNaN bit to make it an SNaN.
971 APInt::tcClearBit(significand, QNaNBit);
973 // If there are no bits set in the payload, we have to set
974 // *something* to make it a NaN instead of an infinity;
975 // conventionally, this is the next bit down from the QNaN bit.
976 if (APInt::tcIsZero(significand, numParts))
977 APInt::tcSetBit(significand, QNaNBit - 1);
978 } else if (semantics->nanEncoding == fltNanEncoding::NegativeZero) {
979 // The only NaN is a quiet NaN, and it has no bits sets in the significand.
980 // Do nothing.
981 } else {
982 // We always have to set the QNaN bit to make it a QNaN.
983 APInt::tcSetBit(significand, QNaNBit);
986 // For x87 extended precision, we want to make a NaN, not a
987 // pseudo-NaN. Maybe we should expose the ability to make
988 // pseudo-NaNs?
989 if (semantics == &semX87DoubleExtended)
990 APInt::tcSetBit(significand, QNaNBit + 1);
993 IEEEFloat &IEEEFloat::operator=(const IEEEFloat &rhs) {
994 if (this != &rhs) {
995 if (semantics != rhs.semantics) {
996 freeSignificand();
997 initialize(rhs.semantics);
999 assign(rhs);
1002 return *this;
1005 IEEEFloat &IEEEFloat::operator=(IEEEFloat &&rhs) {
1006 freeSignificand();
1008 semantics = rhs.semantics;
1009 significand = rhs.significand;
1010 exponent = rhs.exponent;
1011 category = rhs.category;
1012 sign = rhs.sign;
1014 rhs.semantics = &semBogus;
1015 return *this;
1018 bool IEEEFloat::isDenormal() const {
1019 return isFiniteNonZero() && (exponent == semantics->minExponent) &&
1020 (APInt::tcExtractBit(significandParts(),
1021 semantics->precision - 1) == 0);
1024 bool IEEEFloat::isSmallest() const {
1025 // The smallest number by magnitude in our format will be the smallest
1026 // denormal, i.e. the floating point number with exponent being minimum
1027 // exponent and significand bitwise equal to 1 (i.e. with MSB equal to 0).
1028 return isFiniteNonZero() && exponent == semantics->minExponent &&
1029 significandMSB() == 0;
1032 bool IEEEFloat::isSmallestNormalized() const {
1033 return getCategory() == fcNormal && exponent == semantics->minExponent &&
1034 isSignificandAllZerosExceptMSB();
1037 unsigned int IEEEFloat::getNumHighBits() const {
1038 const unsigned int PartCount = partCountForBits(semantics->precision);
1039 const unsigned int Bits = PartCount * integerPartWidth;
1041 // Compute how many bits are used in the final word.
1042 // When precision is just 1, it represents the 'Pth'
1043 // Precision bit and not the actual significand bit.
1044 const unsigned int NumHighBits = (semantics->precision > 1)
1045 ? (Bits - semantics->precision + 1)
1046 : (Bits - semantics->precision);
1047 return NumHighBits;
1050 bool IEEEFloat::isSignificandAllOnes() const {
1051 // Test if the significand excluding the integral bit is all ones. This allows
1052 // us to test for binade boundaries.
1053 const integerPart *Parts = significandParts();
1054 const unsigned PartCount = partCountForBits(semantics->precision);
1055 for (unsigned i = 0; i < PartCount - 1; i++)
1056 if (~Parts[i])
1057 return false;
1059 // Set the unused high bits to all ones when we compare.
1060 const unsigned NumHighBits = getNumHighBits();
1061 assert(NumHighBits <= integerPartWidth && NumHighBits > 0 &&
1062 "Can not have more high bits to fill than integerPartWidth");
1063 const integerPart HighBitFill =
1064 ~integerPart(0) << (integerPartWidth - NumHighBits);
1065 if ((semantics->precision <= 1) || (~(Parts[PartCount - 1] | HighBitFill)))
1066 return false;
1068 return true;
1071 bool IEEEFloat::isSignificandAllOnesExceptLSB() const {
1072 // Test if the significand excluding the integral bit is all ones except for
1073 // the least significant bit.
1074 const integerPart *Parts = significandParts();
1076 if (Parts[0] & 1)
1077 return false;
1079 const unsigned PartCount = partCountForBits(semantics->precision);
1080 for (unsigned i = 0; i < PartCount - 1; i++) {
1081 if (~Parts[i] & ~unsigned{!i})
1082 return false;
1085 // Set the unused high bits to all ones when we compare.
1086 const unsigned NumHighBits = getNumHighBits();
1087 assert(NumHighBits <= integerPartWidth && NumHighBits > 0 &&
1088 "Can not have more high bits to fill than integerPartWidth");
1089 const integerPart HighBitFill = ~integerPart(0)
1090 << (integerPartWidth - NumHighBits);
1091 if (~(Parts[PartCount - 1] | HighBitFill | 0x1))
1092 return false;
1094 return true;
1097 bool IEEEFloat::isSignificandAllZeros() const {
1098 // Test if the significand excluding the integral bit is all zeros. This
1099 // allows us to test for binade boundaries.
1100 const integerPart *Parts = significandParts();
1101 const unsigned PartCount = partCountForBits(semantics->precision);
1103 for (unsigned i = 0; i < PartCount - 1; i++)
1104 if (Parts[i])
1105 return false;
1107 // Compute how many bits are used in the final word.
1108 const unsigned NumHighBits = getNumHighBits();
1109 assert(NumHighBits < integerPartWidth && "Can not have more high bits to "
1110 "clear than integerPartWidth");
1111 const integerPart HighBitMask = ~integerPart(0) >> NumHighBits;
1113 if ((semantics->precision > 1) && (Parts[PartCount - 1] & HighBitMask))
1114 return false;
1116 return true;
1119 bool IEEEFloat::isSignificandAllZerosExceptMSB() const {
1120 const integerPart *Parts = significandParts();
1121 const unsigned PartCount = partCountForBits(semantics->precision);
1123 for (unsigned i = 0; i < PartCount - 1; i++) {
1124 if (Parts[i])
1125 return false;
1128 const unsigned NumHighBits = getNumHighBits();
1129 const integerPart MSBMask = integerPart(1)
1130 << (integerPartWidth - NumHighBits);
1131 return ((semantics->precision <= 1) || (Parts[PartCount - 1] == MSBMask));
1134 bool IEEEFloat::isLargest() const {
1135 bool IsMaxExp = isFiniteNonZero() && exponent == semantics->maxExponent;
1136 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly &&
1137 semantics->nanEncoding == fltNanEncoding::AllOnes) {
1138 // The largest number by magnitude in our format will be the floating point
1139 // number with maximum exponent and with significand that is all ones except
1140 // the LSB.
1141 return (IsMaxExp && APFloat::hasSignificand(*semantics))
1142 ? isSignificandAllOnesExceptLSB()
1143 : IsMaxExp;
1144 } else {
1145 // The largest number by magnitude in our format will be the floating point
1146 // number with maximum exponent and with significand that is all ones.
1147 return IsMaxExp && isSignificandAllOnes();
1151 bool IEEEFloat::isInteger() const {
1152 // This could be made more efficient; I'm going for obviously correct.
1153 if (!isFinite()) return false;
1154 IEEEFloat truncated = *this;
1155 truncated.roundToIntegral(rmTowardZero);
1156 return compare(truncated) == cmpEqual;
1159 bool IEEEFloat::bitwiseIsEqual(const IEEEFloat &rhs) const {
1160 if (this == &rhs)
1161 return true;
1162 if (semantics != rhs.semantics ||
1163 category != rhs.category ||
1164 sign != rhs.sign)
1165 return false;
1166 if (category==fcZero || category==fcInfinity)
1167 return true;
1169 if (isFiniteNonZero() && exponent != rhs.exponent)
1170 return false;
1172 return std::equal(significandParts(), significandParts() + partCount(),
1173 rhs.significandParts());
1176 IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics, integerPart value) {
1177 initialize(&ourSemantics);
1178 sign = 0;
1179 category = fcNormal;
1180 zeroSignificand();
1181 exponent = ourSemantics.precision - 1;
1182 significandParts()[0] = value;
1183 normalize(rmNearestTiesToEven, lfExactlyZero);
1186 IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics) {
1187 initialize(&ourSemantics);
1188 // The Float8E8MOFNU format does not have a representation
1189 // for zero. So, use the closest representation instead.
1190 // Moreover, the all-zero encoding represents a valid
1191 // normal value (which is the smallestNormalized here).
1192 // Hence, we call makeSmallestNormalized (where category is
1193 // 'fcNormal') instead of makeZero (where category is 'fcZero').
1194 ourSemantics.hasZero ? makeZero(false) : makeSmallestNormalized(false);
1197 // Delegate to the previous constructor, because later copy constructor may
1198 // actually inspects category, which can't be garbage.
1199 IEEEFloat::IEEEFloat(const fltSemantics &ourSemantics, uninitializedTag tag)
1200 : IEEEFloat(ourSemantics) {}
1202 IEEEFloat::IEEEFloat(const IEEEFloat &rhs) {
1203 initialize(rhs.semantics);
1204 assign(rhs);
1207 IEEEFloat::IEEEFloat(IEEEFloat &&rhs) : semantics(&semBogus) {
1208 *this = std::move(rhs);
1211 IEEEFloat::~IEEEFloat() { freeSignificand(); }
1213 unsigned int IEEEFloat::partCount() const {
1214 return partCountForBits(semantics->precision + 1);
1217 const APFloat::integerPart *IEEEFloat::significandParts() const {
1218 return const_cast<IEEEFloat *>(this)->significandParts();
1221 APFloat::integerPart *IEEEFloat::significandParts() {
1222 if (partCount() > 1)
1223 return significand.parts;
1224 else
1225 return &significand.part;
1228 void IEEEFloat::zeroSignificand() {
1229 APInt::tcSet(significandParts(), 0, partCount());
1232 /* Increment an fcNormal floating point number's significand. */
1233 void IEEEFloat::incrementSignificand() {
1234 integerPart carry;
1236 carry = APInt::tcIncrement(significandParts(), partCount());
1238 /* Our callers should never cause us to overflow. */
1239 assert(carry == 0);
1240 (void)carry;
1243 /* Add the significand of the RHS. Returns the carry flag. */
1244 APFloat::integerPart IEEEFloat::addSignificand(const IEEEFloat &rhs) {
1245 integerPart *parts;
1247 parts = significandParts();
1249 assert(semantics == rhs.semantics);
1250 assert(exponent == rhs.exponent);
1252 return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount());
1255 /* Subtract the significand of the RHS with a borrow flag. Returns
1256 the borrow flag. */
1257 APFloat::integerPart IEEEFloat::subtractSignificand(const IEEEFloat &rhs,
1258 integerPart borrow) {
1259 integerPart *parts;
1261 parts = significandParts();
1263 assert(semantics == rhs.semantics);
1264 assert(exponent == rhs.exponent);
1266 return APInt::tcSubtract(parts, rhs.significandParts(), borrow,
1267 partCount());
1270 /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it
1271 on to the full-precision result of the multiplication. Returns the
1272 lost fraction. */
1273 lostFraction IEEEFloat::multiplySignificand(const IEEEFloat &rhs,
1274 IEEEFloat addend,
1275 bool ignoreAddend) {
1276 unsigned int omsb; // One, not zero, based MSB.
1277 unsigned int partsCount, newPartsCount, precision;
1278 integerPart *lhsSignificand;
1279 integerPart scratch[4];
1280 integerPart *fullSignificand;
1281 lostFraction lost_fraction;
1282 bool ignored;
1284 assert(semantics == rhs.semantics);
1286 precision = semantics->precision;
1288 // Allocate space for twice as many bits as the original significand, plus one
1289 // extra bit for the addition to overflow into.
1290 newPartsCount = partCountForBits(precision * 2 + 1);
1292 if (newPartsCount > 4)
1293 fullSignificand = new integerPart[newPartsCount];
1294 else
1295 fullSignificand = scratch;
1297 lhsSignificand = significandParts();
1298 partsCount = partCount();
1300 APInt::tcFullMultiply(fullSignificand, lhsSignificand,
1301 rhs.significandParts(), partsCount, partsCount);
1303 lost_fraction = lfExactlyZero;
1304 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
1305 exponent += rhs.exponent;
1307 // Assume the operands involved in the multiplication are single-precision
1308 // FP, and the two multiplicants are:
1309 // *this = a23 . a22 ... a0 * 2^e1
1310 // rhs = b23 . b22 ... b0 * 2^e2
1311 // the result of multiplication is:
1312 // *this = c48 c47 c46 . c45 ... c0 * 2^(e1+e2)
1313 // Note that there are three significant bits at the left-hand side of the
1314 // radix point: two for the multiplication, and an overflow bit for the
1315 // addition (that will always be zero at this point). Move the radix point
1316 // toward left by two bits, and adjust exponent accordingly.
1317 exponent += 2;
1319 if (!ignoreAddend && addend.isNonZero()) {
1320 // The intermediate result of the multiplication has "2 * precision"
1321 // signicant bit; adjust the addend to be consistent with mul result.
1323 Significand savedSignificand = significand;
1324 const fltSemantics *savedSemantics = semantics;
1325 fltSemantics extendedSemantics;
1326 opStatus status;
1327 unsigned int extendedPrecision;
1329 // Normalize our MSB to one below the top bit to allow for overflow.
1330 extendedPrecision = 2 * precision + 1;
1331 if (omsb != extendedPrecision - 1) {
1332 assert(extendedPrecision > omsb);
1333 APInt::tcShiftLeft(fullSignificand, newPartsCount,
1334 (extendedPrecision - 1) - omsb);
1335 exponent -= (extendedPrecision - 1) - omsb;
1338 /* Create new semantics. */
1339 extendedSemantics = *semantics;
1340 extendedSemantics.precision = extendedPrecision;
1342 if (newPartsCount == 1)
1343 significand.part = fullSignificand[0];
1344 else
1345 significand.parts = fullSignificand;
1346 semantics = &extendedSemantics;
1348 // Make a copy so we can convert it to the extended semantics.
1349 // Note that we cannot convert the addend directly, as the extendedSemantics
1350 // is a local variable (which we take a reference to).
1351 IEEEFloat extendedAddend(addend);
1352 status = extendedAddend.convert(extendedSemantics, APFloat::rmTowardZero,
1353 &ignored);
1354 assert(status == APFloat::opOK);
1355 (void)status;
1357 // Shift the significand of the addend right by one bit. This guarantees
1358 // that the high bit of the significand is zero (same as fullSignificand),
1359 // so the addition will overflow (if it does overflow at all) into the top bit.
1360 lost_fraction = extendedAddend.shiftSignificandRight(1);
1361 assert(lost_fraction == lfExactlyZero &&
1362 "Lost precision while shifting addend for fused-multiply-add.");
1364 lost_fraction = addOrSubtractSignificand(extendedAddend, false);
1366 /* Restore our state. */
1367 if (newPartsCount == 1)
1368 fullSignificand[0] = significand.part;
1369 significand = savedSignificand;
1370 semantics = savedSemantics;
1372 omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1;
1375 // Convert the result having "2 * precision" significant-bits back to the one
1376 // having "precision" significant-bits. First, move the radix point from
1377 // poision "2*precision - 1" to "precision - 1". The exponent need to be
1378 // adjusted by "2*precision - 1" - "precision - 1" = "precision".
1379 exponent -= precision + 1;
1381 // In case MSB resides at the left-hand side of radix point, shift the
1382 // mantissa right by some amount to make sure the MSB reside right before
1383 // the radix point (i.e. "MSB . rest-significant-bits").
1385 // Note that the result is not normalized when "omsb < precision". So, the
1386 // caller needs to call IEEEFloat::normalize() if normalized value is
1387 // expected.
1388 if (omsb > precision) {
1389 unsigned int bits, significantParts;
1390 lostFraction lf;
1392 bits = omsb - precision;
1393 significantParts = partCountForBits(omsb);
1394 lf = shiftRight(fullSignificand, significantParts, bits);
1395 lost_fraction = combineLostFractions(lf, lost_fraction);
1396 exponent += bits;
1399 APInt::tcAssign(lhsSignificand, fullSignificand, partsCount);
1401 if (newPartsCount > 4)
1402 delete [] fullSignificand;
1404 return lost_fraction;
1407 lostFraction IEEEFloat::multiplySignificand(const IEEEFloat &rhs) {
1408 // When the given semantics has zero, the addend here is a zero.
1409 // i.e . it belongs to the 'fcZero' category.
1410 // But when the semantics does not support zero, we need to
1411 // explicitly convey that this addend should be ignored
1412 // for multiplication.
1413 return multiplySignificand(rhs, IEEEFloat(*semantics), !semantics->hasZero);
1416 /* Multiply the significands of LHS and RHS to DST. */
1417 lostFraction IEEEFloat::divideSignificand(const IEEEFloat &rhs) {
1418 unsigned int bit, i, partsCount;
1419 const integerPart *rhsSignificand;
1420 integerPart *lhsSignificand, *dividend, *divisor;
1421 integerPart scratch[4];
1422 lostFraction lost_fraction;
1424 assert(semantics == rhs.semantics);
1426 lhsSignificand = significandParts();
1427 rhsSignificand = rhs.significandParts();
1428 partsCount = partCount();
1430 if (partsCount > 2)
1431 dividend = new integerPart[partsCount * 2];
1432 else
1433 dividend = scratch;
1435 divisor = dividend + partsCount;
1437 /* Copy the dividend and divisor as they will be modified in-place. */
1438 for (i = 0; i < partsCount; i++) {
1439 dividend[i] = lhsSignificand[i];
1440 divisor[i] = rhsSignificand[i];
1441 lhsSignificand[i] = 0;
1444 exponent -= rhs.exponent;
1446 unsigned int precision = semantics->precision;
1448 /* Normalize the divisor. */
1449 bit = precision - APInt::tcMSB(divisor, partsCount) - 1;
1450 if (bit) {
1451 exponent += bit;
1452 APInt::tcShiftLeft(divisor, partsCount, bit);
1455 /* Normalize the dividend. */
1456 bit = precision - APInt::tcMSB(dividend, partsCount) - 1;
1457 if (bit) {
1458 exponent -= bit;
1459 APInt::tcShiftLeft(dividend, partsCount, bit);
1462 /* Ensure the dividend >= divisor initially for the loop below.
1463 Incidentally, this means that the division loop below is
1464 guaranteed to set the integer bit to one. */
1465 if (APInt::tcCompare(dividend, divisor, partsCount) < 0) {
1466 exponent--;
1467 APInt::tcShiftLeft(dividend, partsCount, 1);
1468 assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0);
1471 /* Long division. */
1472 for (bit = precision; bit; bit -= 1) {
1473 if (APInt::tcCompare(dividend, divisor, partsCount) >= 0) {
1474 APInt::tcSubtract(dividend, divisor, 0, partsCount);
1475 APInt::tcSetBit(lhsSignificand, bit - 1);
1478 APInt::tcShiftLeft(dividend, partsCount, 1);
1481 /* Figure out the lost fraction. */
1482 int cmp = APInt::tcCompare(dividend, divisor, partsCount);
1484 if (cmp > 0)
1485 lost_fraction = lfMoreThanHalf;
1486 else if (cmp == 0)
1487 lost_fraction = lfExactlyHalf;
1488 else if (APInt::tcIsZero(dividend, partsCount))
1489 lost_fraction = lfExactlyZero;
1490 else
1491 lost_fraction = lfLessThanHalf;
1493 if (partsCount > 2)
1494 delete [] dividend;
1496 return lost_fraction;
1499 unsigned int IEEEFloat::significandMSB() const {
1500 return APInt::tcMSB(significandParts(), partCount());
1503 unsigned int IEEEFloat::significandLSB() const {
1504 return APInt::tcLSB(significandParts(), partCount());
1507 /* Note that a zero result is NOT normalized to fcZero. */
1508 lostFraction IEEEFloat::shiftSignificandRight(unsigned int bits) {
1509 /* Our exponent should not overflow. */
1510 assert((ExponentType) (exponent + bits) >= exponent);
1512 exponent += bits;
1514 return shiftRight(significandParts(), partCount(), bits);
1517 /* Shift the significand left BITS bits, subtract BITS from its exponent. */
1518 void IEEEFloat::shiftSignificandLeft(unsigned int bits) {
1519 assert(bits < semantics->precision ||
1520 (semantics->precision == 1 && bits <= 1));
1522 if (bits) {
1523 unsigned int partsCount = partCount();
1525 APInt::tcShiftLeft(significandParts(), partsCount, bits);
1526 exponent -= bits;
1528 assert(!APInt::tcIsZero(significandParts(), partsCount));
1532 APFloat::cmpResult IEEEFloat::compareAbsoluteValue(const IEEEFloat &rhs) const {
1533 int compare;
1535 assert(semantics == rhs.semantics);
1536 assert(isFiniteNonZero());
1537 assert(rhs.isFiniteNonZero());
1539 compare = exponent - rhs.exponent;
1541 /* If exponents are equal, do an unsigned bignum comparison of the
1542 significands. */
1543 if (compare == 0)
1544 compare = APInt::tcCompare(significandParts(), rhs.significandParts(),
1545 partCount());
1547 if (compare > 0)
1548 return cmpGreaterThan;
1549 else if (compare < 0)
1550 return cmpLessThan;
1551 else
1552 return cmpEqual;
1555 /* Set the least significant BITS bits of a bignum, clear the
1556 rest. */
1557 static void tcSetLeastSignificantBits(APInt::WordType *dst, unsigned parts,
1558 unsigned bits) {
1559 unsigned i = 0;
1560 while (bits > APInt::APINT_BITS_PER_WORD) {
1561 dst[i++] = ~(APInt::WordType)0;
1562 bits -= APInt::APINT_BITS_PER_WORD;
1565 if (bits)
1566 dst[i++] = ~(APInt::WordType)0 >> (APInt::APINT_BITS_PER_WORD - bits);
1568 while (i < parts)
1569 dst[i++] = 0;
1572 /* Handle overflow. Sign is preserved. We either become infinity or
1573 the largest finite number. */
1574 APFloat::opStatus IEEEFloat::handleOverflow(roundingMode rounding_mode) {
1575 if (semantics->nonFiniteBehavior != fltNonfiniteBehavior::FiniteOnly) {
1576 /* Infinity? */
1577 if (rounding_mode == rmNearestTiesToEven ||
1578 rounding_mode == rmNearestTiesToAway ||
1579 (rounding_mode == rmTowardPositive && !sign) ||
1580 (rounding_mode == rmTowardNegative && sign)) {
1581 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly)
1582 makeNaN(false, sign);
1583 else
1584 category = fcInfinity;
1585 return static_cast<opStatus>(opOverflow | opInexact);
1589 /* Otherwise we become the largest finite number. */
1590 category = fcNormal;
1591 exponent = semantics->maxExponent;
1592 tcSetLeastSignificantBits(significandParts(), partCount(),
1593 semantics->precision);
1594 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly &&
1595 semantics->nanEncoding == fltNanEncoding::AllOnes)
1596 APInt::tcClearBit(significandParts(), 0);
1598 return opInexact;
1601 /* Returns TRUE if, when truncating the current number, with BIT the
1602 new LSB, with the given lost fraction and rounding mode, the result
1603 would need to be rounded away from zero (i.e., by increasing the
1604 signficand). This routine must work for fcZero of both signs, and
1605 fcNormal numbers. */
1606 bool IEEEFloat::roundAwayFromZero(roundingMode rounding_mode,
1607 lostFraction lost_fraction,
1608 unsigned int bit) const {
1609 /* NaNs and infinities should not have lost fractions. */
1610 assert(isFiniteNonZero() || category == fcZero);
1612 /* Current callers never pass this so we don't handle it. */
1613 assert(lost_fraction != lfExactlyZero);
1615 switch (rounding_mode) {
1616 case rmNearestTiesToAway:
1617 return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf;
1619 case rmNearestTiesToEven:
1620 if (lost_fraction == lfMoreThanHalf)
1621 return true;
1623 /* Our zeroes don't have a significand to test. */
1624 if (lost_fraction == lfExactlyHalf && category != fcZero)
1625 return APInt::tcExtractBit(significandParts(), bit);
1627 return false;
1629 case rmTowardZero:
1630 return false;
1632 case rmTowardPositive:
1633 return !sign;
1635 case rmTowardNegative:
1636 return sign;
1638 default:
1639 break;
1641 llvm_unreachable("Invalid rounding mode found");
1644 APFloat::opStatus IEEEFloat::normalize(roundingMode rounding_mode,
1645 lostFraction lost_fraction) {
1646 unsigned int omsb; /* One, not zero, based MSB. */
1647 int exponentChange;
1649 if (!isFiniteNonZero())
1650 return opOK;
1652 /* Before rounding normalize the exponent of fcNormal numbers. */
1653 omsb = significandMSB() + 1;
1655 if (omsb) {
1656 /* OMSB is numbered from 1. We want to place it in the integer
1657 bit numbered PRECISION if possible, with a compensating change in
1658 the exponent. */
1659 exponentChange = omsb - semantics->precision;
1661 /* If the resulting exponent is too high, overflow according to
1662 the rounding mode. */
1663 if (exponent + exponentChange > semantics->maxExponent)
1664 return handleOverflow(rounding_mode);
1666 /* Subnormal numbers have exponent minExponent, and their MSB
1667 is forced based on that. */
1668 if (exponent + exponentChange < semantics->minExponent)
1669 exponentChange = semantics->minExponent - exponent;
1671 /* Shifting left is easy as we don't lose precision. */
1672 if (exponentChange < 0) {
1673 assert(lost_fraction == lfExactlyZero);
1675 shiftSignificandLeft(-exponentChange);
1677 return opOK;
1680 if (exponentChange > 0) {
1681 lostFraction lf;
1683 /* Shift right and capture any new lost fraction. */
1684 lf = shiftSignificandRight(exponentChange);
1686 lost_fraction = combineLostFractions(lf, lost_fraction);
1688 /* Keep OMSB up-to-date. */
1689 if (omsb > (unsigned) exponentChange)
1690 omsb -= exponentChange;
1691 else
1692 omsb = 0;
1696 // The all-ones values is an overflow if NaN is all ones. If NaN is
1697 // represented by negative zero, then it is a valid finite value.
1698 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly &&
1699 semantics->nanEncoding == fltNanEncoding::AllOnes &&
1700 exponent == semantics->maxExponent && isSignificandAllOnes())
1701 return handleOverflow(rounding_mode);
1703 /* Now round the number according to rounding_mode given the lost
1704 fraction. */
1706 /* As specified in IEEE 754, since we do not trap we do not report
1707 underflow for exact results. */
1708 if (lost_fraction == lfExactlyZero) {
1709 /* Canonicalize zeroes. */
1710 if (omsb == 0) {
1711 category = fcZero;
1712 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
1713 sign = false;
1714 if (!semantics->hasZero)
1715 makeSmallestNormalized(false);
1718 return opOK;
1721 /* Increment the significand if we're rounding away from zero. */
1722 if (roundAwayFromZero(rounding_mode, lost_fraction, 0)) {
1723 if (omsb == 0)
1724 exponent = semantics->minExponent;
1726 incrementSignificand();
1727 omsb = significandMSB() + 1;
1729 /* Did the significand increment overflow? */
1730 if (omsb == (unsigned) semantics->precision + 1) {
1731 /* Renormalize by incrementing the exponent and shifting our
1732 significand right one. However if we already have the
1733 maximum exponent we overflow to infinity. */
1734 if (exponent == semantics->maxExponent)
1735 // Invoke overflow handling with a rounding mode that will guarantee
1736 // that the result gets turned into the correct infinity representation.
1737 // This is needed instead of just setting the category to infinity to
1738 // account for 8-bit floating point types that have no inf, only NaN.
1739 return handleOverflow(sign ? rmTowardNegative : rmTowardPositive);
1741 shiftSignificandRight(1);
1743 return opInexact;
1746 // The all-ones values is an overflow if NaN is all ones. If NaN is
1747 // represented by negative zero, then it is a valid finite value.
1748 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly &&
1749 semantics->nanEncoding == fltNanEncoding::AllOnes &&
1750 exponent == semantics->maxExponent && isSignificandAllOnes())
1751 return handleOverflow(rounding_mode);
1754 /* The normal case - we were and are not denormal, and any
1755 significand increment above didn't overflow. */
1756 if (omsb == semantics->precision)
1757 return opInexact;
1759 /* We have a non-zero denormal. */
1760 assert(omsb < semantics->precision);
1762 /* Canonicalize zeroes. */
1763 if (omsb == 0) {
1764 category = fcZero;
1765 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
1766 sign = false;
1767 // This condition handles the case where the semantics
1768 // does not have zero but uses the all-zero encoding
1769 // to represent the smallest normal value.
1770 if (!semantics->hasZero)
1771 makeSmallestNormalized(false);
1774 /* The fcZero case is a denormal that underflowed to zero. */
1775 return (opStatus) (opUnderflow | opInexact);
1778 APFloat::opStatus IEEEFloat::addOrSubtractSpecials(const IEEEFloat &rhs,
1779 bool subtract) {
1780 switch (PackCategoriesIntoKey(category, rhs.category)) {
1781 default:
1782 llvm_unreachable(nullptr);
1784 case PackCategoriesIntoKey(fcZero, fcNaN):
1785 case PackCategoriesIntoKey(fcNormal, fcNaN):
1786 case PackCategoriesIntoKey(fcInfinity, fcNaN):
1787 assign(rhs);
1788 [[fallthrough]];
1789 case PackCategoriesIntoKey(fcNaN, fcZero):
1790 case PackCategoriesIntoKey(fcNaN, fcNormal):
1791 case PackCategoriesIntoKey(fcNaN, fcInfinity):
1792 case PackCategoriesIntoKey(fcNaN, fcNaN):
1793 if (isSignaling()) {
1794 makeQuiet();
1795 return opInvalidOp;
1797 return rhs.isSignaling() ? opInvalidOp : opOK;
1799 case PackCategoriesIntoKey(fcNormal, fcZero):
1800 case PackCategoriesIntoKey(fcInfinity, fcNormal):
1801 case PackCategoriesIntoKey(fcInfinity, fcZero):
1802 return opOK;
1804 case PackCategoriesIntoKey(fcNormal, fcInfinity):
1805 case PackCategoriesIntoKey(fcZero, fcInfinity):
1806 category = fcInfinity;
1807 sign = rhs.sign ^ subtract;
1808 return opOK;
1810 case PackCategoriesIntoKey(fcZero, fcNormal):
1811 assign(rhs);
1812 sign = rhs.sign ^ subtract;
1813 return opOK;
1815 case PackCategoriesIntoKey(fcZero, fcZero):
1816 /* Sign depends on rounding mode; handled by caller. */
1817 return opOK;
1819 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
1820 /* Differently signed infinities can only be validly
1821 subtracted. */
1822 if (((sign ^ rhs.sign)!=0) != subtract) {
1823 makeNaN();
1824 return opInvalidOp;
1827 return opOK;
1829 case PackCategoriesIntoKey(fcNormal, fcNormal):
1830 return opDivByZero;
1834 /* Add or subtract two normal numbers. */
1835 lostFraction IEEEFloat::addOrSubtractSignificand(const IEEEFloat &rhs,
1836 bool subtract) {
1837 integerPart carry;
1838 lostFraction lost_fraction;
1839 int bits;
1841 /* Determine if the operation on the absolute values is effectively
1842 an addition or subtraction. */
1843 subtract ^= static_cast<bool>(sign ^ rhs.sign);
1845 /* Are we bigger exponent-wise than the RHS? */
1846 bits = exponent - rhs.exponent;
1848 /* Subtraction is more subtle than one might naively expect. */
1849 if (subtract) {
1850 if ((bits < 0) && !semantics->hasSignedRepr)
1851 llvm_unreachable(
1852 "This floating point format does not support signed values");
1854 IEEEFloat temp_rhs(rhs);
1856 if (bits == 0)
1857 lost_fraction = lfExactlyZero;
1858 else if (bits > 0) {
1859 lost_fraction = temp_rhs.shiftSignificandRight(bits - 1);
1860 shiftSignificandLeft(1);
1861 } else {
1862 lost_fraction = shiftSignificandRight(-bits - 1);
1863 temp_rhs.shiftSignificandLeft(1);
1866 // Should we reverse the subtraction.
1867 if (compareAbsoluteValue(temp_rhs) == cmpLessThan) {
1868 carry = temp_rhs.subtractSignificand
1869 (*this, lost_fraction != lfExactlyZero);
1870 copySignificand(temp_rhs);
1871 sign = !sign;
1872 } else {
1873 carry = subtractSignificand
1874 (temp_rhs, lost_fraction != lfExactlyZero);
1877 /* Invert the lost fraction - it was on the RHS and
1878 subtracted. */
1879 if (lost_fraction == lfLessThanHalf)
1880 lost_fraction = lfMoreThanHalf;
1881 else if (lost_fraction == lfMoreThanHalf)
1882 lost_fraction = lfLessThanHalf;
1884 /* The code above is intended to ensure that no borrow is
1885 necessary. */
1886 assert(!carry);
1887 (void)carry;
1888 } else {
1889 if (bits > 0) {
1890 IEEEFloat temp_rhs(rhs);
1892 lost_fraction = temp_rhs.shiftSignificandRight(bits);
1893 carry = addSignificand(temp_rhs);
1894 } else {
1895 lost_fraction = shiftSignificandRight(-bits);
1896 carry = addSignificand(rhs);
1899 /* We have a guard bit; generating a carry cannot happen. */
1900 assert(!carry);
1901 (void)carry;
1904 return lost_fraction;
1907 APFloat::opStatus IEEEFloat::multiplySpecials(const IEEEFloat &rhs) {
1908 switch (PackCategoriesIntoKey(category, rhs.category)) {
1909 default:
1910 llvm_unreachable(nullptr);
1912 case PackCategoriesIntoKey(fcZero, fcNaN):
1913 case PackCategoriesIntoKey(fcNormal, fcNaN):
1914 case PackCategoriesIntoKey(fcInfinity, fcNaN):
1915 assign(rhs);
1916 sign = false;
1917 [[fallthrough]];
1918 case PackCategoriesIntoKey(fcNaN, fcZero):
1919 case PackCategoriesIntoKey(fcNaN, fcNormal):
1920 case PackCategoriesIntoKey(fcNaN, fcInfinity):
1921 case PackCategoriesIntoKey(fcNaN, fcNaN):
1922 sign ^= rhs.sign; // restore the original sign
1923 if (isSignaling()) {
1924 makeQuiet();
1925 return opInvalidOp;
1927 return rhs.isSignaling() ? opInvalidOp : opOK;
1929 case PackCategoriesIntoKey(fcNormal, fcInfinity):
1930 case PackCategoriesIntoKey(fcInfinity, fcNormal):
1931 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
1932 category = fcInfinity;
1933 return opOK;
1935 case PackCategoriesIntoKey(fcZero, fcNormal):
1936 case PackCategoriesIntoKey(fcNormal, fcZero):
1937 case PackCategoriesIntoKey(fcZero, fcZero):
1938 category = fcZero;
1939 return opOK;
1941 case PackCategoriesIntoKey(fcZero, fcInfinity):
1942 case PackCategoriesIntoKey(fcInfinity, fcZero):
1943 makeNaN();
1944 return opInvalidOp;
1946 case PackCategoriesIntoKey(fcNormal, fcNormal):
1947 return opOK;
1951 APFloat::opStatus IEEEFloat::divideSpecials(const IEEEFloat &rhs) {
1952 switch (PackCategoriesIntoKey(category, rhs.category)) {
1953 default:
1954 llvm_unreachable(nullptr);
1956 case PackCategoriesIntoKey(fcZero, fcNaN):
1957 case PackCategoriesIntoKey(fcNormal, fcNaN):
1958 case PackCategoriesIntoKey(fcInfinity, fcNaN):
1959 assign(rhs);
1960 sign = false;
1961 [[fallthrough]];
1962 case PackCategoriesIntoKey(fcNaN, fcZero):
1963 case PackCategoriesIntoKey(fcNaN, fcNormal):
1964 case PackCategoriesIntoKey(fcNaN, fcInfinity):
1965 case PackCategoriesIntoKey(fcNaN, fcNaN):
1966 sign ^= rhs.sign; // restore the original sign
1967 if (isSignaling()) {
1968 makeQuiet();
1969 return opInvalidOp;
1971 return rhs.isSignaling() ? opInvalidOp : opOK;
1973 case PackCategoriesIntoKey(fcInfinity, fcZero):
1974 case PackCategoriesIntoKey(fcInfinity, fcNormal):
1975 case PackCategoriesIntoKey(fcZero, fcInfinity):
1976 case PackCategoriesIntoKey(fcZero, fcNormal):
1977 return opOK;
1979 case PackCategoriesIntoKey(fcNormal, fcInfinity):
1980 category = fcZero;
1981 return opOK;
1983 case PackCategoriesIntoKey(fcNormal, fcZero):
1984 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly)
1985 makeNaN(false, sign);
1986 else
1987 category = fcInfinity;
1988 return opDivByZero;
1990 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
1991 case PackCategoriesIntoKey(fcZero, fcZero):
1992 makeNaN();
1993 return opInvalidOp;
1995 case PackCategoriesIntoKey(fcNormal, fcNormal):
1996 return opOK;
2000 APFloat::opStatus IEEEFloat::modSpecials(const IEEEFloat &rhs) {
2001 switch (PackCategoriesIntoKey(category, rhs.category)) {
2002 default:
2003 llvm_unreachable(nullptr);
2005 case PackCategoriesIntoKey(fcZero, fcNaN):
2006 case PackCategoriesIntoKey(fcNormal, fcNaN):
2007 case PackCategoriesIntoKey(fcInfinity, fcNaN):
2008 assign(rhs);
2009 [[fallthrough]];
2010 case PackCategoriesIntoKey(fcNaN, fcZero):
2011 case PackCategoriesIntoKey(fcNaN, fcNormal):
2012 case PackCategoriesIntoKey(fcNaN, fcInfinity):
2013 case PackCategoriesIntoKey(fcNaN, fcNaN):
2014 if (isSignaling()) {
2015 makeQuiet();
2016 return opInvalidOp;
2018 return rhs.isSignaling() ? opInvalidOp : opOK;
2020 case PackCategoriesIntoKey(fcZero, fcInfinity):
2021 case PackCategoriesIntoKey(fcZero, fcNormal):
2022 case PackCategoriesIntoKey(fcNormal, fcInfinity):
2023 return opOK;
2025 case PackCategoriesIntoKey(fcNormal, fcZero):
2026 case PackCategoriesIntoKey(fcInfinity, fcZero):
2027 case PackCategoriesIntoKey(fcInfinity, fcNormal):
2028 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
2029 case PackCategoriesIntoKey(fcZero, fcZero):
2030 makeNaN();
2031 return opInvalidOp;
2033 case PackCategoriesIntoKey(fcNormal, fcNormal):
2034 return opOK;
2038 APFloat::opStatus IEEEFloat::remainderSpecials(const IEEEFloat &rhs) {
2039 switch (PackCategoriesIntoKey(category, rhs.category)) {
2040 default:
2041 llvm_unreachable(nullptr);
2043 case PackCategoriesIntoKey(fcZero, fcNaN):
2044 case PackCategoriesIntoKey(fcNormal, fcNaN):
2045 case PackCategoriesIntoKey(fcInfinity, fcNaN):
2046 assign(rhs);
2047 [[fallthrough]];
2048 case PackCategoriesIntoKey(fcNaN, fcZero):
2049 case PackCategoriesIntoKey(fcNaN, fcNormal):
2050 case PackCategoriesIntoKey(fcNaN, fcInfinity):
2051 case PackCategoriesIntoKey(fcNaN, fcNaN):
2052 if (isSignaling()) {
2053 makeQuiet();
2054 return opInvalidOp;
2056 return rhs.isSignaling() ? opInvalidOp : opOK;
2058 case PackCategoriesIntoKey(fcZero, fcInfinity):
2059 case PackCategoriesIntoKey(fcZero, fcNormal):
2060 case PackCategoriesIntoKey(fcNormal, fcInfinity):
2061 return opOK;
2063 case PackCategoriesIntoKey(fcNormal, fcZero):
2064 case PackCategoriesIntoKey(fcInfinity, fcZero):
2065 case PackCategoriesIntoKey(fcInfinity, fcNormal):
2066 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
2067 case PackCategoriesIntoKey(fcZero, fcZero):
2068 makeNaN();
2069 return opInvalidOp;
2071 case PackCategoriesIntoKey(fcNormal, fcNormal):
2072 return opDivByZero; // fake status, indicating this is not a special case
2076 /* Change sign. */
2077 void IEEEFloat::changeSign() {
2078 // With NaN-as-negative-zero, neither NaN or negative zero can change
2079 // their signs.
2080 if (semantics->nanEncoding == fltNanEncoding::NegativeZero &&
2081 (isZero() || isNaN()))
2082 return;
2083 /* Look mummy, this one's easy. */
2084 sign = !sign;
2087 /* Normalized addition or subtraction. */
2088 APFloat::opStatus IEEEFloat::addOrSubtract(const IEEEFloat &rhs,
2089 roundingMode rounding_mode,
2090 bool subtract) {
2091 opStatus fs;
2093 fs = addOrSubtractSpecials(rhs, subtract);
2095 /* This return code means it was not a simple case. */
2096 if (fs == opDivByZero) {
2097 lostFraction lost_fraction;
2099 lost_fraction = addOrSubtractSignificand(rhs, subtract);
2100 fs = normalize(rounding_mode, lost_fraction);
2102 /* Can only be zero if we lost no fraction. */
2103 assert(category != fcZero || lost_fraction == lfExactlyZero);
2106 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
2107 positive zero unless rounding to minus infinity, except that
2108 adding two like-signed zeroes gives that zero. */
2109 if (category == fcZero) {
2110 if (rhs.category != fcZero || (sign == rhs.sign) == subtract)
2111 sign = (rounding_mode == rmTowardNegative);
2112 // NaN-in-negative-zero means zeros need to be normalized to +0.
2113 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
2114 sign = false;
2117 return fs;
2120 /* Normalized addition. */
2121 APFloat::opStatus IEEEFloat::add(const IEEEFloat &rhs,
2122 roundingMode rounding_mode) {
2123 return addOrSubtract(rhs, rounding_mode, false);
2126 /* Normalized subtraction. */
2127 APFloat::opStatus IEEEFloat::subtract(const IEEEFloat &rhs,
2128 roundingMode rounding_mode) {
2129 return addOrSubtract(rhs, rounding_mode, true);
2132 /* Normalized multiply. */
2133 APFloat::opStatus IEEEFloat::multiply(const IEEEFloat &rhs,
2134 roundingMode rounding_mode) {
2135 opStatus fs;
2137 sign ^= rhs.sign;
2138 fs = multiplySpecials(rhs);
2140 if (isZero() && semantics->nanEncoding == fltNanEncoding::NegativeZero)
2141 sign = false;
2142 if (isFiniteNonZero()) {
2143 lostFraction lost_fraction = multiplySignificand(rhs);
2144 fs = normalize(rounding_mode, lost_fraction);
2145 if (lost_fraction != lfExactlyZero)
2146 fs = (opStatus) (fs | opInexact);
2149 return fs;
2152 /* Normalized divide. */
2153 APFloat::opStatus IEEEFloat::divide(const IEEEFloat &rhs,
2154 roundingMode rounding_mode) {
2155 opStatus fs;
2157 sign ^= rhs.sign;
2158 fs = divideSpecials(rhs);
2160 if (isZero() && semantics->nanEncoding == fltNanEncoding::NegativeZero)
2161 sign = false;
2162 if (isFiniteNonZero()) {
2163 lostFraction lost_fraction = divideSignificand(rhs);
2164 fs = normalize(rounding_mode, lost_fraction);
2165 if (lost_fraction != lfExactlyZero)
2166 fs = (opStatus) (fs | opInexact);
2169 return fs;
2172 /* Normalized remainder. */
2173 APFloat::opStatus IEEEFloat::remainder(const IEEEFloat &rhs) {
2174 opStatus fs;
2175 unsigned int origSign = sign;
2177 // First handle the special cases.
2178 fs = remainderSpecials(rhs);
2179 if (fs != opDivByZero)
2180 return fs;
2182 fs = opOK;
2184 // Make sure the current value is less than twice the denom. If the addition
2185 // did not succeed (an overflow has happened), which means that the finite
2186 // value we currently posses must be less than twice the denom (as we are
2187 // using the same semantics).
2188 IEEEFloat P2 = rhs;
2189 if (P2.add(rhs, rmNearestTiesToEven) == opOK) {
2190 fs = mod(P2);
2191 assert(fs == opOK);
2194 // Lets work with absolute numbers.
2195 IEEEFloat P = rhs;
2196 P.sign = false;
2197 sign = false;
2200 // To calculate the remainder we use the following scheme.
2202 // The remainder is defained as follows:
2204 // remainder = numer - rquot * denom = x - r * p
2206 // Where r is the result of: x/p, rounded toward the nearest integral value
2207 // (with halfway cases rounded toward the even number).
2209 // Currently, (after x mod 2p):
2210 // r is the number of 2p's present inside x, which is inherently, an even
2211 // number of p's.
2213 // We may split the remaining calculation into 4 options:
2214 // - if x < 0.5p then we round to the nearest number with is 0, and are done.
2215 // - if x == 0.5p then we round to the nearest even number which is 0, and we
2216 // are done as well.
2217 // - if 0.5p < x < p then we round to nearest number which is 1, and we have
2218 // to subtract 1p at least once.
2219 // - if x >= p then we must subtract p at least once, as x must be a
2220 // remainder.
2222 // By now, we were done, or we added 1 to r, which in turn, now an odd number.
2224 // We can now split the remaining calculation to the following 3 options:
2225 // - if x < 0.5p then we round to the nearest number with is 0, and are done.
2226 // - if x == 0.5p then we round to the nearest even number. As r is odd, we
2227 // must round up to the next even number. so we must subtract p once more.
2228 // - if x > 0.5p (and inherently x < p) then we must round r up to the next
2229 // integral, and subtract p once more.
2232 // Extend the semantics to prevent an overflow/underflow or inexact result.
2233 bool losesInfo;
2234 fltSemantics extendedSemantics = *semantics;
2235 extendedSemantics.maxExponent++;
2236 extendedSemantics.minExponent--;
2237 extendedSemantics.precision += 2;
2239 IEEEFloat VEx = *this;
2240 fs = VEx.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo);
2241 assert(fs == opOK && !losesInfo);
2242 IEEEFloat PEx = P;
2243 fs = PEx.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo);
2244 assert(fs == opOK && !losesInfo);
2246 // It is simpler to work with 2x instead of 0.5p, and we do not need to lose
2247 // any fraction.
2248 fs = VEx.add(VEx, rmNearestTiesToEven);
2249 assert(fs == opOK);
2251 if (VEx.compare(PEx) == cmpGreaterThan) {
2252 fs = subtract(P, rmNearestTiesToEven);
2253 assert(fs == opOK);
2255 // Make VEx = this.add(this), but because we have different semantics, we do
2256 // not want to `convert` again, so we just subtract PEx twice (which equals
2257 // to the desired value).
2258 fs = VEx.subtract(PEx, rmNearestTiesToEven);
2259 assert(fs == opOK);
2260 fs = VEx.subtract(PEx, rmNearestTiesToEven);
2261 assert(fs == opOK);
2263 cmpResult result = VEx.compare(PEx);
2264 if (result == cmpGreaterThan || result == cmpEqual) {
2265 fs = subtract(P, rmNearestTiesToEven);
2266 assert(fs == opOK);
2270 if (isZero()) {
2271 sign = origSign; // IEEE754 requires this
2272 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
2273 // But some 8-bit floats only have positive 0.
2274 sign = false;
2277 else
2278 sign ^= origSign;
2279 return fs;
2282 /* Normalized llvm frem (C fmod). */
2283 APFloat::opStatus IEEEFloat::mod(const IEEEFloat &rhs) {
2284 opStatus fs;
2285 fs = modSpecials(rhs);
2286 unsigned int origSign = sign;
2288 while (isFiniteNonZero() && rhs.isFiniteNonZero() &&
2289 compareAbsoluteValue(rhs) != cmpLessThan) {
2290 int Exp = ilogb(*this) - ilogb(rhs);
2291 IEEEFloat V = scalbn(rhs, Exp, rmNearestTiesToEven);
2292 // V can overflow to NaN with fltNonfiniteBehavior::NanOnly, so explicitly
2293 // check for it.
2294 if (V.isNaN() || compareAbsoluteValue(V) == cmpLessThan)
2295 V = scalbn(rhs, Exp - 1, rmNearestTiesToEven);
2296 V.sign = sign;
2298 fs = subtract(V, rmNearestTiesToEven);
2300 // When the semantics supports zero, this loop's
2301 // exit-condition is handled by the 'isFiniteNonZero'
2302 // category check above. However, when the semantics
2303 // does not have 'fcZero' and we have reached the
2304 // minimum possible value, (and any further subtract
2305 // will underflow to the same value) explicitly
2306 // provide an exit-path here.
2307 if (!semantics->hasZero && this->isSmallest())
2308 break;
2310 assert(fs==opOK);
2312 if (isZero()) {
2313 sign = origSign; // fmod requires this
2314 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
2315 sign = false;
2317 return fs;
2320 /* Normalized fused-multiply-add. */
2321 APFloat::opStatus IEEEFloat::fusedMultiplyAdd(const IEEEFloat &multiplicand,
2322 const IEEEFloat &addend,
2323 roundingMode rounding_mode) {
2324 opStatus fs;
2326 /* Post-multiplication sign, before addition. */
2327 sign ^= multiplicand.sign;
2329 /* If and only if all arguments are normal do we need to do an
2330 extended-precision calculation. */
2331 if (isFiniteNonZero() &&
2332 multiplicand.isFiniteNonZero() &&
2333 addend.isFinite()) {
2334 lostFraction lost_fraction;
2336 lost_fraction = multiplySignificand(multiplicand, addend);
2337 fs = normalize(rounding_mode, lost_fraction);
2338 if (lost_fraction != lfExactlyZero)
2339 fs = (opStatus) (fs | opInexact);
2341 /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a
2342 positive zero unless rounding to minus infinity, except that
2343 adding two like-signed zeroes gives that zero. */
2344 if (category == fcZero && !(fs & opUnderflow) && sign != addend.sign) {
2345 sign = (rounding_mode == rmTowardNegative);
2346 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
2347 sign = false;
2349 } else {
2350 fs = multiplySpecials(multiplicand);
2352 /* FS can only be opOK or opInvalidOp. There is no more work
2353 to do in the latter case. The IEEE-754R standard says it is
2354 implementation-defined in this case whether, if ADDEND is a
2355 quiet NaN, we raise invalid op; this implementation does so.
2357 If we need to do the addition we can do so with normal
2358 precision. */
2359 if (fs == opOK)
2360 fs = addOrSubtract(addend, rounding_mode, false);
2363 return fs;
2366 /* Rounding-mode correct round to integral value. */
2367 APFloat::opStatus IEEEFloat::roundToIntegral(roundingMode rounding_mode) {
2368 opStatus fs;
2370 if (isInfinity())
2371 // [IEEE Std 754-2008 6.1]:
2372 // The behavior of infinity in floating-point arithmetic is derived from the
2373 // limiting cases of real arithmetic with operands of arbitrarily
2374 // large magnitude, when such a limit exists.
2375 // ...
2376 // Operations on infinite operands are usually exact and therefore signal no
2377 // exceptions ...
2378 return opOK;
2380 if (isNaN()) {
2381 if (isSignaling()) {
2382 // [IEEE Std 754-2008 6.2]:
2383 // Under default exception handling, any operation signaling an invalid
2384 // operation exception and for which a floating-point result is to be
2385 // delivered shall deliver a quiet NaN.
2386 makeQuiet();
2387 // [IEEE Std 754-2008 6.2]:
2388 // Signaling NaNs shall be reserved operands that, under default exception
2389 // handling, signal the invalid operation exception(see 7.2) for every
2390 // general-computational and signaling-computational operation except for
2391 // the conversions described in 5.12.
2392 return opInvalidOp;
2393 } else {
2394 // [IEEE Std 754-2008 6.2]:
2395 // For an operation with quiet NaN inputs, other than maximum and minimum
2396 // operations, if a floating-point result is to be delivered the result
2397 // shall be a quiet NaN which should be one of the input NaNs.
2398 // ...
2399 // Every general-computational and quiet-computational operation involving
2400 // one or more input NaNs, none of them signaling, shall signal no
2401 // exception, except fusedMultiplyAdd might signal the invalid operation
2402 // exception(see 7.2).
2403 return opOK;
2407 if (isZero()) {
2408 // [IEEE Std 754-2008 6.3]:
2409 // ... the sign of the result of conversions, the quantize operation, the
2410 // roundToIntegral operations, and the roundToIntegralExact(see 5.3.1) is
2411 // the sign of the first or only operand.
2412 return opOK;
2415 // If the exponent is large enough, we know that this value is already
2416 // integral, and the arithmetic below would potentially cause it to saturate
2417 // to +/-Inf. Bail out early instead.
2418 if (exponent + 1 >= (int)APFloat::semanticsPrecision(*semantics))
2419 return opOK;
2421 // The algorithm here is quite simple: we add 2^(p-1), where p is the
2422 // precision of our format, and then subtract it back off again. The choice
2423 // of rounding modes for the addition/subtraction determines the rounding mode
2424 // for our integral rounding as well.
2425 // NOTE: When the input value is negative, we do subtraction followed by
2426 // addition instead.
2427 APInt IntegerConstant(NextPowerOf2(APFloat::semanticsPrecision(*semantics)),
2429 IntegerConstant <<= APFloat::semanticsPrecision(*semantics) - 1;
2430 IEEEFloat MagicConstant(*semantics);
2431 fs = MagicConstant.convertFromAPInt(IntegerConstant, false,
2432 rmNearestTiesToEven);
2433 assert(fs == opOK);
2434 MagicConstant.sign = sign;
2436 // Preserve the input sign so that we can handle the case of zero result
2437 // correctly.
2438 bool inputSign = isNegative();
2440 fs = add(MagicConstant, rounding_mode);
2442 // Current value and 'MagicConstant' are both integers, so the result of the
2443 // subtraction is always exact according to Sterbenz' lemma.
2444 subtract(MagicConstant, rounding_mode);
2446 // Restore the input sign.
2447 if (inputSign != isNegative())
2448 changeSign();
2450 return fs;
2453 /* Comparison requires normalized numbers. */
2454 APFloat::cmpResult IEEEFloat::compare(const IEEEFloat &rhs) const {
2455 cmpResult result;
2457 assert(semantics == rhs.semantics);
2459 switch (PackCategoriesIntoKey(category, rhs.category)) {
2460 default:
2461 llvm_unreachable(nullptr);
2463 case PackCategoriesIntoKey(fcNaN, fcZero):
2464 case PackCategoriesIntoKey(fcNaN, fcNormal):
2465 case PackCategoriesIntoKey(fcNaN, fcInfinity):
2466 case PackCategoriesIntoKey(fcNaN, fcNaN):
2467 case PackCategoriesIntoKey(fcZero, fcNaN):
2468 case PackCategoriesIntoKey(fcNormal, fcNaN):
2469 case PackCategoriesIntoKey(fcInfinity, fcNaN):
2470 return cmpUnordered;
2472 case PackCategoriesIntoKey(fcInfinity, fcNormal):
2473 case PackCategoriesIntoKey(fcInfinity, fcZero):
2474 case PackCategoriesIntoKey(fcNormal, fcZero):
2475 if (sign)
2476 return cmpLessThan;
2477 else
2478 return cmpGreaterThan;
2480 case PackCategoriesIntoKey(fcNormal, fcInfinity):
2481 case PackCategoriesIntoKey(fcZero, fcInfinity):
2482 case PackCategoriesIntoKey(fcZero, fcNormal):
2483 if (rhs.sign)
2484 return cmpGreaterThan;
2485 else
2486 return cmpLessThan;
2488 case PackCategoriesIntoKey(fcInfinity, fcInfinity):
2489 if (sign == rhs.sign)
2490 return cmpEqual;
2491 else if (sign)
2492 return cmpLessThan;
2493 else
2494 return cmpGreaterThan;
2496 case PackCategoriesIntoKey(fcZero, fcZero):
2497 return cmpEqual;
2499 case PackCategoriesIntoKey(fcNormal, fcNormal):
2500 break;
2503 /* Two normal numbers. Do they have the same sign? */
2504 if (sign != rhs.sign) {
2505 if (sign)
2506 result = cmpLessThan;
2507 else
2508 result = cmpGreaterThan;
2509 } else {
2510 /* Compare absolute values; invert result if negative. */
2511 result = compareAbsoluteValue(rhs);
2513 if (sign) {
2514 if (result == cmpLessThan)
2515 result = cmpGreaterThan;
2516 else if (result == cmpGreaterThan)
2517 result = cmpLessThan;
2521 return result;
2524 /// IEEEFloat::convert - convert a value of one floating point type to another.
2525 /// The return value corresponds to the IEEE754 exceptions. *losesInfo
2526 /// records whether the transformation lost information, i.e. whether
2527 /// converting the result back to the original type will produce the
2528 /// original value (this is almost the same as return value==fsOK, but there
2529 /// are edge cases where this is not so).
2531 APFloat::opStatus IEEEFloat::convert(const fltSemantics &toSemantics,
2532 roundingMode rounding_mode,
2533 bool *losesInfo) {
2534 lostFraction lostFraction;
2535 unsigned int newPartCount, oldPartCount;
2536 opStatus fs;
2537 int shift;
2538 const fltSemantics &fromSemantics = *semantics;
2539 bool is_signaling = isSignaling();
2541 lostFraction = lfExactlyZero;
2542 newPartCount = partCountForBits(toSemantics.precision + 1);
2543 oldPartCount = partCount();
2544 shift = toSemantics.precision - fromSemantics.precision;
2546 bool X86SpecialNan = false;
2547 if (&fromSemantics == &semX87DoubleExtended &&
2548 &toSemantics != &semX87DoubleExtended && category == fcNaN &&
2549 (!(*significandParts() & 0x8000000000000000ULL) ||
2550 !(*significandParts() & 0x4000000000000000ULL))) {
2551 // x86 has some unusual NaNs which cannot be represented in any other
2552 // format; note them here.
2553 X86SpecialNan = true;
2556 // If this is a truncation of a denormal number, and the target semantics
2557 // has larger exponent range than the source semantics (this can happen
2558 // when truncating from PowerPC double-double to double format), the
2559 // right shift could lose result mantissa bits. Adjust exponent instead
2560 // of performing excessive shift.
2561 // Also do a similar trick in case shifting denormal would produce zero
2562 // significand as this case isn't handled correctly by normalize.
2563 if (shift < 0 && isFiniteNonZero()) {
2564 int omsb = significandMSB() + 1;
2565 int exponentChange = omsb - fromSemantics.precision;
2566 if (exponent + exponentChange < toSemantics.minExponent)
2567 exponentChange = toSemantics.minExponent - exponent;
2568 if (exponentChange < shift)
2569 exponentChange = shift;
2570 if (exponentChange < 0) {
2571 shift -= exponentChange;
2572 exponent += exponentChange;
2573 } else if (omsb <= -shift) {
2574 exponentChange = omsb + shift - 1; // leave at least one bit set
2575 shift -= exponentChange;
2576 exponent += exponentChange;
2580 // If this is a truncation, perform the shift before we narrow the storage.
2581 if (shift < 0 && (isFiniteNonZero() ||
2582 (category == fcNaN && semantics->nonFiniteBehavior !=
2583 fltNonfiniteBehavior::NanOnly)))
2584 lostFraction = shiftRight(significandParts(), oldPartCount, -shift);
2586 // Fix the storage so it can hold to new value.
2587 if (newPartCount > oldPartCount) {
2588 // The new type requires more storage; make it available.
2589 integerPart *newParts;
2590 newParts = new integerPart[newPartCount];
2591 APInt::tcSet(newParts, 0, newPartCount);
2592 if (isFiniteNonZero() || category==fcNaN)
2593 APInt::tcAssign(newParts, significandParts(), oldPartCount);
2594 freeSignificand();
2595 significand.parts = newParts;
2596 } else if (newPartCount == 1 && oldPartCount != 1) {
2597 // Switch to built-in storage for a single part.
2598 integerPart newPart = 0;
2599 if (isFiniteNonZero() || category==fcNaN)
2600 newPart = significandParts()[0];
2601 freeSignificand();
2602 significand.part = newPart;
2605 // Now that we have the right storage, switch the semantics.
2606 semantics = &toSemantics;
2608 // If this is an extension, perform the shift now that the storage is
2609 // available.
2610 if (shift > 0 && (isFiniteNonZero() || category==fcNaN))
2611 APInt::tcShiftLeft(significandParts(), newPartCount, shift);
2613 if (isFiniteNonZero()) {
2614 fs = normalize(rounding_mode, lostFraction);
2615 *losesInfo = (fs != opOK);
2616 } else if (category == fcNaN) {
2617 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
2618 *losesInfo =
2619 fromSemantics.nonFiniteBehavior != fltNonfiniteBehavior::NanOnly;
2620 makeNaN(false, sign);
2621 return is_signaling ? opInvalidOp : opOK;
2624 // If NaN is negative zero, we need to create a new NaN to avoid converting
2625 // NaN to -Inf.
2626 if (fromSemantics.nanEncoding == fltNanEncoding::NegativeZero &&
2627 semantics->nanEncoding != fltNanEncoding::NegativeZero)
2628 makeNaN(false, false);
2630 *losesInfo = lostFraction != lfExactlyZero || X86SpecialNan;
2632 // For x87 extended precision, we want to make a NaN, not a special NaN if
2633 // the input wasn't special either.
2634 if (!X86SpecialNan && semantics == &semX87DoubleExtended)
2635 APInt::tcSetBit(significandParts(), semantics->precision - 1);
2637 // Convert of sNaN creates qNaN and raises an exception (invalid op).
2638 // This also guarantees that a sNaN does not become Inf on a truncation
2639 // that loses all payload bits.
2640 if (is_signaling) {
2641 makeQuiet();
2642 fs = opInvalidOp;
2643 } else {
2644 fs = opOK;
2646 } else if (category == fcInfinity &&
2647 semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
2648 makeNaN(false, sign);
2649 *losesInfo = true;
2650 fs = opInexact;
2651 } else if (category == fcZero &&
2652 semantics->nanEncoding == fltNanEncoding::NegativeZero) {
2653 // Negative zero loses info, but positive zero doesn't.
2654 *losesInfo =
2655 fromSemantics.nanEncoding != fltNanEncoding::NegativeZero && sign;
2656 fs = *losesInfo ? opInexact : opOK;
2657 // NaN is negative zero means -0 -> +0, which can lose information
2658 sign = false;
2659 } else {
2660 *losesInfo = false;
2661 fs = opOK;
2664 if (category == fcZero && !semantics->hasZero)
2665 makeSmallestNormalized(false);
2666 return fs;
2669 /* Convert a floating point number to an integer according to the
2670 rounding mode. If the rounded integer value is out of range this
2671 returns an invalid operation exception and the contents of the
2672 destination parts are unspecified. If the rounded value is in
2673 range but the floating point number is not the exact integer, the C
2674 standard doesn't require an inexact exception to be raised. IEEE
2675 854 does require it so we do that.
2677 Note that for conversions to integer type the C standard requires
2678 round-to-zero to always be used. */
2679 APFloat::opStatus IEEEFloat::convertToSignExtendedInteger(
2680 MutableArrayRef<integerPart> parts, unsigned int width, bool isSigned,
2681 roundingMode rounding_mode, bool *isExact) const {
2682 lostFraction lost_fraction;
2683 const integerPart *src;
2684 unsigned int dstPartsCount, truncatedBits;
2686 *isExact = false;
2688 /* Handle the three special cases first. */
2689 if (category == fcInfinity || category == fcNaN)
2690 return opInvalidOp;
2692 dstPartsCount = partCountForBits(width);
2693 assert(dstPartsCount <= parts.size() && "Integer too big");
2695 if (category == fcZero) {
2696 APInt::tcSet(parts.data(), 0, dstPartsCount);
2697 // Negative zero can't be represented as an int.
2698 *isExact = !sign;
2699 return opOK;
2702 src = significandParts();
2704 /* Step 1: place our absolute value, with any fraction truncated, in
2705 the destination. */
2706 if (exponent < 0) {
2707 /* Our absolute value is less than one; truncate everything. */
2708 APInt::tcSet(parts.data(), 0, dstPartsCount);
2709 /* For exponent -1 the integer bit represents .5, look at that.
2710 For smaller exponents leftmost truncated bit is 0. */
2711 truncatedBits = semantics->precision -1U - exponent;
2712 } else {
2713 /* We want the most significant (exponent + 1) bits; the rest are
2714 truncated. */
2715 unsigned int bits = exponent + 1U;
2717 /* Hopelessly large in magnitude? */
2718 if (bits > width)
2719 return opInvalidOp;
2721 if (bits < semantics->precision) {
2722 /* We truncate (semantics->precision - bits) bits. */
2723 truncatedBits = semantics->precision - bits;
2724 APInt::tcExtract(parts.data(), dstPartsCount, src, bits, truncatedBits);
2725 } else {
2726 /* We want at least as many bits as are available. */
2727 APInt::tcExtract(parts.data(), dstPartsCount, src, semantics->precision,
2729 APInt::tcShiftLeft(parts.data(), dstPartsCount,
2730 bits - semantics->precision);
2731 truncatedBits = 0;
2735 /* Step 2: work out any lost fraction, and increment the absolute
2736 value if we would round away from zero. */
2737 if (truncatedBits) {
2738 lost_fraction = lostFractionThroughTruncation(src, partCount(),
2739 truncatedBits);
2740 if (lost_fraction != lfExactlyZero &&
2741 roundAwayFromZero(rounding_mode, lost_fraction, truncatedBits)) {
2742 if (APInt::tcIncrement(parts.data(), dstPartsCount))
2743 return opInvalidOp; /* Overflow. */
2745 } else {
2746 lost_fraction = lfExactlyZero;
2749 /* Step 3: check if we fit in the destination. */
2750 unsigned int omsb = APInt::tcMSB(parts.data(), dstPartsCount) + 1;
2752 if (sign) {
2753 if (!isSigned) {
2754 /* Negative numbers cannot be represented as unsigned. */
2755 if (omsb != 0)
2756 return opInvalidOp;
2757 } else {
2758 /* It takes omsb bits to represent the unsigned integer value.
2759 We lose a bit for the sign, but care is needed as the
2760 maximally negative integer is a special case. */
2761 if (omsb == width &&
2762 APInt::tcLSB(parts.data(), dstPartsCount) + 1 != omsb)
2763 return opInvalidOp;
2765 /* This case can happen because of rounding. */
2766 if (omsb > width)
2767 return opInvalidOp;
2770 APInt::tcNegate (parts.data(), dstPartsCount);
2771 } else {
2772 if (omsb >= width + !isSigned)
2773 return opInvalidOp;
2776 if (lost_fraction == lfExactlyZero) {
2777 *isExact = true;
2778 return opOK;
2779 } else
2780 return opInexact;
2783 /* Same as convertToSignExtendedInteger, except we provide
2784 deterministic values in case of an invalid operation exception,
2785 namely zero for NaNs and the minimal or maximal value respectively
2786 for underflow or overflow.
2787 The *isExact output tells whether the result is exact, in the sense
2788 that converting it back to the original floating point type produces
2789 the original value. This is almost equivalent to result==opOK,
2790 except for negative zeroes.
2792 APFloat::opStatus
2793 IEEEFloat::convertToInteger(MutableArrayRef<integerPart> parts,
2794 unsigned int width, bool isSigned,
2795 roundingMode rounding_mode, bool *isExact) const {
2796 opStatus fs;
2798 fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode,
2799 isExact);
2801 if (fs == opInvalidOp) {
2802 unsigned int bits, dstPartsCount;
2804 dstPartsCount = partCountForBits(width);
2805 assert(dstPartsCount <= parts.size() && "Integer too big");
2807 if (category == fcNaN)
2808 bits = 0;
2809 else if (sign)
2810 bits = isSigned;
2811 else
2812 bits = width - isSigned;
2814 tcSetLeastSignificantBits(parts.data(), dstPartsCount, bits);
2815 if (sign && isSigned)
2816 APInt::tcShiftLeft(parts.data(), dstPartsCount, width - 1);
2819 return fs;
2822 /* Convert an unsigned integer SRC to a floating point number,
2823 rounding according to ROUNDING_MODE. The sign of the floating
2824 point number is not modified. */
2825 APFloat::opStatus IEEEFloat::convertFromUnsignedParts(
2826 const integerPart *src, unsigned int srcCount, roundingMode rounding_mode) {
2827 unsigned int omsb, precision, dstCount;
2828 integerPart *dst;
2829 lostFraction lost_fraction;
2831 category = fcNormal;
2832 omsb = APInt::tcMSB(src, srcCount) + 1;
2833 dst = significandParts();
2834 dstCount = partCount();
2835 precision = semantics->precision;
2837 /* We want the most significant PRECISION bits of SRC. There may not
2838 be that many; extract what we can. */
2839 if (precision <= omsb) {
2840 exponent = omsb - 1;
2841 lost_fraction = lostFractionThroughTruncation(src, srcCount,
2842 omsb - precision);
2843 APInt::tcExtract(dst, dstCount, src, precision, omsb - precision);
2844 } else {
2845 exponent = precision - 1;
2846 lost_fraction = lfExactlyZero;
2847 APInt::tcExtract(dst, dstCount, src, omsb, 0);
2850 return normalize(rounding_mode, lost_fraction);
2853 APFloat::opStatus IEEEFloat::convertFromAPInt(const APInt &Val, bool isSigned,
2854 roundingMode rounding_mode) {
2855 unsigned int partCount = Val.getNumWords();
2856 APInt api = Val;
2858 sign = false;
2859 if (isSigned && api.isNegative()) {
2860 sign = true;
2861 api = -api;
2864 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
2867 /* Convert a two's complement integer SRC to a floating point number,
2868 rounding according to ROUNDING_MODE. ISSIGNED is true if the
2869 integer is signed, in which case it must be sign-extended. */
2870 APFloat::opStatus
2871 IEEEFloat::convertFromSignExtendedInteger(const integerPart *src,
2872 unsigned int srcCount, bool isSigned,
2873 roundingMode rounding_mode) {
2874 opStatus status;
2876 if (isSigned &&
2877 APInt::tcExtractBit(src, srcCount * integerPartWidth - 1)) {
2878 integerPart *copy;
2880 /* If we're signed and negative negate a copy. */
2881 sign = true;
2882 copy = new integerPart[srcCount];
2883 APInt::tcAssign(copy, src, srcCount);
2884 APInt::tcNegate(copy, srcCount);
2885 status = convertFromUnsignedParts(copy, srcCount, rounding_mode);
2886 delete [] copy;
2887 } else {
2888 sign = false;
2889 status = convertFromUnsignedParts(src, srcCount, rounding_mode);
2892 return status;
2895 /* FIXME: should this just take a const APInt reference? */
2896 APFloat::opStatus
2897 IEEEFloat::convertFromZeroExtendedInteger(const integerPart *parts,
2898 unsigned int width, bool isSigned,
2899 roundingMode rounding_mode) {
2900 unsigned int partCount = partCountForBits(width);
2901 APInt api = APInt(width, ArrayRef(parts, partCount));
2903 sign = false;
2904 if (isSigned && APInt::tcExtractBit(parts, width - 1)) {
2905 sign = true;
2906 api = -api;
2909 return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode);
2912 Expected<APFloat::opStatus>
2913 IEEEFloat::convertFromHexadecimalString(StringRef s,
2914 roundingMode rounding_mode) {
2915 lostFraction lost_fraction = lfExactlyZero;
2917 category = fcNormal;
2918 zeroSignificand();
2919 exponent = 0;
2921 integerPart *significand = significandParts();
2922 unsigned partsCount = partCount();
2923 unsigned bitPos = partsCount * integerPartWidth;
2924 bool computedTrailingFraction = false;
2926 // Skip leading zeroes and any (hexa)decimal point.
2927 StringRef::iterator begin = s.begin();
2928 StringRef::iterator end = s.end();
2929 StringRef::iterator dot;
2930 auto PtrOrErr = skipLeadingZeroesAndAnyDot(begin, end, &dot);
2931 if (!PtrOrErr)
2932 return PtrOrErr.takeError();
2933 StringRef::iterator p = *PtrOrErr;
2934 StringRef::iterator firstSignificantDigit = p;
2936 while (p != end) {
2937 integerPart hex_value;
2939 if (*p == '.') {
2940 if (dot != end)
2941 return createError("String contains multiple dots");
2942 dot = p++;
2943 continue;
2946 hex_value = hexDigitValue(*p);
2947 if (hex_value == UINT_MAX)
2948 break;
2950 p++;
2952 // Store the number while we have space.
2953 if (bitPos) {
2954 bitPos -= 4;
2955 hex_value <<= bitPos % integerPartWidth;
2956 significand[bitPos / integerPartWidth] |= hex_value;
2957 } else if (!computedTrailingFraction) {
2958 auto FractOrErr = trailingHexadecimalFraction(p, end, hex_value);
2959 if (!FractOrErr)
2960 return FractOrErr.takeError();
2961 lost_fraction = *FractOrErr;
2962 computedTrailingFraction = true;
2966 /* Hex floats require an exponent but not a hexadecimal point. */
2967 if (p == end)
2968 return createError("Hex strings require an exponent");
2969 if (*p != 'p' && *p != 'P')
2970 return createError("Invalid character in significand");
2971 if (p == begin)
2972 return createError("Significand has no digits");
2973 if (dot != end && p - begin == 1)
2974 return createError("Significand has no digits");
2976 /* Ignore the exponent if we are zero. */
2977 if (p != firstSignificantDigit) {
2978 int expAdjustment;
2980 /* Implicit hexadecimal point? */
2981 if (dot == end)
2982 dot = p;
2984 /* Calculate the exponent adjustment implicit in the number of
2985 significant digits. */
2986 expAdjustment = static_cast<int>(dot - firstSignificantDigit);
2987 if (expAdjustment < 0)
2988 expAdjustment++;
2989 expAdjustment = expAdjustment * 4 - 1;
2991 /* Adjust for writing the significand starting at the most
2992 significant nibble. */
2993 expAdjustment += semantics->precision;
2994 expAdjustment -= partsCount * integerPartWidth;
2996 /* Adjust for the given exponent. */
2997 auto ExpOrErr = totalExponent(p + 1, end, expAdjustment);
2998 if (!ExpOrErr)
2999 return ExpOrErr.takeError();
3000 exponent = *ExpOrErr;
3003 return normalize(rounding_mode, lost_fraction);
3006 APFloat::opStatus
3007 IEEEFloat::roundSignificandWithExponent(const integerPart *decSigParts,
3008 unsigned sigPartCount, int exp,
3009 roundingMode rounding_mode) {
3010 unsigned int parts, pow5PartCount;
3011 fltSemantics calcSemantics = { 32767, -32767, 0, 0 };
3012 integerPart pow5Parts[maxPowerOfFiveParts];
3013 bool isNearest;
3015 isNearest = (rounding_mode == rmNearestTiesToEven ||
3016 rounding_mode == rmNearestTiesToAway);
3018 parts = partCountForBits(semantics->precision + 11);
3020 /* Calculate pow(5, abs(exp)). */
3021 pow5PartCount = powerOf5(pow5Parts, exp >= 0 ? exp: -exp);
3023 for (;; parts *= 2) {
3024 opStatus sigStatus, powStatus;
3025 unsigned int excessPrecision, truncatedBits;
3027 calcSemantics.precision = parts * integerPartWidth - 1;
3028 excessPrecision = calcSemantics.precision - semantics->precision;
3029 truncatedBits = excessPrecision;
3031 IEEEFloat decSig(calcSemantics, uninitialized);
3032 decSig.makeZero(sign);
3033 IEEEFloat pow5(calcSemantics);
3035 sigStatus = decSig.convertFromUnsignedParts(decSigParts, sigPartCount,
3036 rmNearestTiesToEven);
3037 powStatus = pow5.convertFromUnsignedParts(pow5Parts, pow5PartCount,
3038 rmNearestTiesToEven);
3039 /* Add exp, as 10^n = 5^n * 2^n. */
3040 decSig.exponent += exp;
3042 lostFraction calcLostFraction;
3043 integerPart HUerr, HUdistance;
3044 unsigned int powHUerr;
3046 if (exp >= 0) {
3047 /* multiplySignificand leaves the precision-th bit set to 1. */
3048 calcLostFraction = decSig.multiplySignificand(pow5);
3049 powHUerr = powStatus != opOK;
3050 } else {
3051 calcLostFraction = decSig.divideSignificand(pow5);
3052 /* Denormal numbers have less precision. */
3053 if (decSig.exponent < semantics->minExponent) {
3054 excessPrecision += (semantics->minExponent - decSig.exponent);
3055 truncatedBits = excessPrecision;
3056 if (excessPrecision > calcSemantics.precision)
3057 excessPrecision = calcSemantics.precision;
3059 /* Extra half-ulp lost in reciprocal of exponent. */
3060 powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2;
3063 /* Both multiplySignificand and divideSignificand return the
3064 result with the integer bit set. */
3065 assert(APInt::tcExtractBit
3066 (decSig.significandParts(), calcSemantics.precision - 1) == 1);
3068 HUerr = HUerrBound(calcLostFraction != lfExactlyZero, sigStatus != opOK,
3069 powHUerr);
3070 HUdistance = 2 * ulpsFromBoundary(decSig.significandParts(),
3071 excessPrecision, isNearest);
3073 /* Are we guaranteed to round correctly if we truncate? */
3074 if (HUdistance >= HUerr) {
3075 APInt::tcExtract(significandParts(), partCount(), decSig.significandParts(),
3076 calcSemantics.precision - excessPrecision,
3077 excessPrecision);
3078 /* Take the exponent of decSig. If we tcExtract-ed less bits
3079 above we must adjust our exponent to compensate for the
3080 implicit right shift. */
3081 exponent = (decSig.exponent + semantics->precision
3082 - (calcSemantics.precision - excessPrecision));
3083 calcLostFraction = lostFractionThroughTruncation(decSig.significandParts(),
3084 decSig.partCount(),
3085 truncatedBits);
3086 return normalize(rounding_mode, calcLostFraction);
3091 Expected<APFloat::opStatus>
3092 IEEEFloat::convertFromDecimalString(StringRef str, roundingMode rounding_mode) {
3093 decimalInfo D;
3094 opStatus fs;
3096 /* Scan the text. */
3097 StringRef::iterator p = str.begin();
3098 if (Error Err = interpretDecimal(p, str.end(), &D))
3099 return std::move(Err);
3101 /* Handle the quick cases. First the case of no significant digits,
3102 i.e. zero, and then exponents that are obviously too large or too
3103 small. Writing L for log 10 / log 2, a number d.ddddd*10^exp
3104 definitely overflows if
3106 (exp - 1) * L >= maxExponent
3108 and definitely underflows to zero where
3110 (exp + 1) * L <= minExponent - precision
3112 With integer arithmetic the tightest bounds for L are
3114 93/28 < L < 196/59 [ numerator <= 256 ]
3115 42039/12655 < L < 28738/8651 [ numerator <= 65536 ]
3118 // Test if we have a zero number allowing for strings with no null terminators
3119 // and zero decimals with non-zero exponents.
3121 // We computed firstSigDigit by ignoring all zeros and dots. Thus if
3122 // D->firstSigDigit equals str.end(), every digit must be a zero and there can
3123 // be at most one dot. On the other hand, if we have a zero with a non-zero
3124 // exponent, then we know that D.firstSigDigit will be non-numeric.
3125 if (D.firstSigDigit == str.end() || decDigitValue(*D.firstSigDigit) >= 10U) {
3126 category = fcZero;
3127 fs = opOK;
3128 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
3129 sign = false;
3130 if (!semantics->hasZero)
3131 makeSmallestNormalized(false);
3133 /* Check whether the normalized exponent is high enough to overflow
3134 max during the log-rebasing in the max-exponent check below. */
3135 } else if (D.normalizedExponent - 1 > INT_MAX / 42039) {
3136 fs = handleOverflow(rounding_mode);
3138 /* If it wasn't, then it also wasn't high enough to overflow max
3139 during the log-rebasing in the min-exponent check. Check that it
3140 won't overflow min in either check, then perform the min-exponent
3141 check. */
3142 } else if (D.normalizedExponent - 1 < INT_MIN / 42039 ||
3143 (D.normalizedExponent + 1) * 28738 <=
3144 8651 * (semantics->minExponent - (int) semantics->precision)) {
3145 /* Underflow to zero and round. */
3146 category = fcNormal;
3147 zeroSignificand();
3148 fs = normalize(rounding_mode, lfLessThanHalf);
3150 /* We can finally safely perform the max-exponent check. */
3151 } else if ((D.normalizedExponent - 1) * 42039
3152 >= 12655 * semantics->maxExponent) {
3153 /* Overflow and round. */
3154 fs = handleOverflow(rounding_mode);
3155 } else {
3156 integerPart *decSignificand;
3157 unsigned int partCount;
3159 /* A tight upper bound on number of bits required to hold an
3160 N-digit decimal integer is N * 196 / 59. Allocate enough space
3161 to hold the full significand, and an extra part required by
3162 tcMultiplyPart. */
3163 partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1;
3164 partCount = partCountForBits(1 + 196 * partCount / 59);
3165 decSignificand = new integerPart[partCount + 1];
3166 partCount = 0;
3168 /* Convert to binary efficiently - we do almost all multiplication
3169 in an integerPart. When this would overflow do we do a single
3170 bignum multiplication, and then revert again to multiplication
3171 in an integerPart. */
3172 do {
3173 integerPart decValue, val, multiplier;
3175 val = 0;
3176 multiplier = 1;
3178 do {
3179 if (*p == '.') {
3180 p++;
3181 if (p == str.end()) {
3182 break;
3185 decValue = decDigitValue(*p++);
3186 if (decValue >= 10U) {
3187 delete[] decSignificand;
3188 return createError("Invalid character in significand");
3190 multiplier *= 10;
3191 val = val * 10 + decValue;
3192 /* The maximum number that can be multiplied by ten with any
3193 digit added without overflowing an integerPart. */
3194 } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10);
3196 /* Multiply out the current part. */
3197 APInt::tcMultiplyPart(decSignificand, decSignificand, multiplier, val,
3198 partCount, partCount + 1, false);
3200 /* If we used another part (likely but not guaranteed), increase
3201 the count. */
3202 if (decSignificand[partCount])
3203 partCount++;
3204 } while (p <= D.lastSigDigit);
3206 category = fcNormal;
3207 fs = roundSignificandWithExponent(decSignificand, partCount,
3208 D.exponent, rounding_mode);
3210 delete [] decSignificand;
3213 return fs;
3216 bool IEEEFloat::convertFromStringSpecials(StringRef str) {
3217 const size_t MIN_NAME_SIZE = 3;
3219 if (str.size() < MIN_NAME_SIZE)
3220 return false;
3222 if (str == "inf" || str == "INFINITY" || str == "+Inf") {
3223 makeInf(false);
3224 return true;
3227 bool IsNegative = str.front() == '-';
3228 if (IsNegative) {
3229 str = str.drop_front();
3230 if (str.size() < MIN_NAME_SIZE)
3231 return false;
3233 if (str == "inf" || str == "INFINITY" || str == "Inf") {
3234 makeInf(true);
3235 return true;
3239 // If we have a 's' (or 'S') prefix, then this is a Signaling NaN.
3240 bool IsSignaling = str.front() == 's' || str.front() == 'S';
3241 if (IsSignaling) {
3242 str = str.drop_front();
3243 if (str.size() < MIN_NAME_SIZE)
3244 return false;
3247 if (str.starts_with("nan") || str.starts_with("NaN")) {
3248 str = str.drop_front(3);
3250 // A NaN without payload.
3251 if (str.empty()) {
3252 makeNaN(IsSignaling, IsNegative);
3253 return true;
3256 // Allow the payload to be inside parentheses.
3257 if (str.front() == '(') {
3258 // Parentheses should be balanced (and not empty).
3259 if (str.size() <= 2 || str.back() != ')')
3260 return false;
3262 str = str.slice(1, str.size() - 1);
3265 // Determine the payload number's radix.
3266 unsigned Radix = 10;
3267 if (str[0] == '0') {
3268 if (str.size() > 1 && tolower(str[1]) == 'x') {
3269 str = str.drop_front(2);
3270 Radix = 16;
3271 } else
3272 Radix = 8;
3275 // Parse the payload and make the NaN.
3276 APInt Payload;
3277 if (!str.getAsInteger(Radix, Payload)) {
3278 makeNaN(IsSignaling, IsNegative, &Payload);
3279 return true;
3283 return false;
3286 Expected<APFloat::opStatus>
3287 IEEEFloat::convertFromString(StringRef str, roundingMode rounding_mode) {
3288 if (str.empty())
3289 return createError("Invalid string length");
3291 // Handle special cases.
3292 if (convertFromStringSpecials(str))
3293 return opOK;
3295 /* Handle a leading minus sign. */
3296 StringRef::iterator p = str.begin();
3297 size_t slen = str.size();
3298 sign = *p == '-' ? 1 : 0;
3299 if (sign && !semantics->hasSignedRepr)
3300 llvm_unreachable(
3301 "This floating point format does not support signed values");
3303 if (*p == '-' || *p == '+') {
3304 p++;
3305 slen--;
3306 if (!slen)
3307 return createError("String has no digits");
3310 if (slen >= 2 && p[0] == '0' && (p[1] == 'x' || p[1] == 'X')) {
3311 if (slen == 2)
3312 return createError("Invalid string");
3313 return convertFromHexadecimalString(StringRef(p + 2, slen - 2),
3314 rounding_mode);
3317 return convertFromDecimalString(StringRef(p, slen), rounding_mode);
3320 /* Write out a hexadecimal representation of the floating point value
3321 to DST, which must be of sufficient size, in the C99 form
3322 [-]0xh.hhhhp[+-]d. Return the number of characters written,
3323 excluding the terminating NUL.
3325 If UPPERCASE, the output is in upper case, otherwise in lower case.
3327 HEXDIGITS digits appear altogether, rounding the value if
3328 necessary. If HEXDIGITS is 0, the minimal precision to display the
3329 number precisely is used instead. If nothing would appear after
3330 the decimal point it is suppressed.
3332 The decimal exponent is always printed and has at least one digit.
3333 Zero values display an exponent of zero. Infinities and NaNs
3334 appear as "infinity" or "nan" respectively.
3336 The above rules are as specified by C99. There is ambiguity about
3337 what the leading hexadecimal digit should be. This implementation
3338 uses whatever is necessary so that the exponent is displayed as
3339 stored. This implies the exponent will fall within the IEEE format
3340 range, and the leading hexadecimal digit will be 0 (for denormals),
3341 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with
3342 any other digits zero).
3344 unsigned int IEEEFloat::convertToHexString(char *dst, unsigned int hexDigits,
3345 bool upperCase,
3346 roundingMode rounding_mode) const {
3347 char *p;
3349 p = dst;
3350 if (sign)
3351 *dst++ = '-';
3353 switch (category) {
3354 case fcInfinity:
3355 memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1);
3356 dst += sizeof infinityL - 1;
3357 break;
3359 case fcNaN:
3360 memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1);
3361 dst += sizeof NaNU - 1;
3362 break;
3364 case fcZero:
3365 *dst++ = '0';
3366 *dst++ = upperCase ? 'X': 'x';
3367 *dst++ = '0';
3368 if (hexDigits > 1) {
3369 *dst++ = '.';
3370 memset (dst, '0', hexDigits - 1);
3371 dst += hexDigits - 1;
3373 *dst++ = upperCase ? 'P': 'p';
3374 *dst++ = '0';
3375 break;
3377 case fcNormal:
3378 dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode);
3379 break;
3382 *dst = 0;
3384 return static_cast<unsigned int>(dst - p);
3387 /* Does the hard work of outputting the correctly rounded hexadecimal
3388 form of a normal floating point number with the specified number of
3389 hexadecimal digits. If HEXDIGITS is zero the minimum number of
3390 digits necessary to print the value precisely is output. */
3391 char *IEEEFloat::convertNormalToHexString(char *dst, unsigned int hexDigits,
3392 bool upperCase,
3393 roundingMode rounding_mode) const {
3394 unsigned int count, valueBits, shift, partsCount, outputDigits;
3395 const char *hexDigitChars;
3396 const integerPart *significand;
3397 char *p;
3398 bool roundUp;
3400 *dst++ = '0';
3401 *dst++ = upperCase ? 'X': 'x';
3403 roundUp = false;
3404 hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower;
3406 significand = significandParts();
3407 partsCount = partCount();
3409 /* +3 because the first digit only uses the single integer bit, so
3410 we have 3 virtual zero most-significant-bits. */
3411 valueBits = semantics->precision + 3;
3412 shift = integerPartWidth - valueBits % integerPartWidth;
3414 /* The natural number of digits required ignoring trailing
3415 insignificant zeroes. */
3416 outputDigits = (valueBits - significandLSB () + 3) / 4;
3418 /* hexDigits of zero means use the required number for the
3419 precision. Otherwise, see if we are truncating. If we are,
3420 find out if we need to round away from zero. */
3421 if (hexDigits) {
3422 if (hexDigits < outputDigits) {
3423 /* We are dropping non-zero bits, so need to check how to round.
3424 "bits" is the number of dropped bits. */
3425 unsigned int bits;
3426 lostFraction fraction;
3428 bits = valueBits - hexDigits * 4;
3429 fraction = lostFractionThroughTruncation (significand, partsCount, bits);
3430 roundUp = roundAwayFromZero(rounding_mode, fraction, bits);
3432 outputDigits = hexDigits;
3435 /* Write the digits consecutively, and start writing in the location
3436 of the hexadecimal point. We move the most significant digit
3437 left and add the hexadecimal point later. */
3438 p = ++dst;
3440 count = (valueBits + integerPartWidth - 1) / integerPartWidth;
3442 while (outputDigits && count) {
3443 integerPart part;
3445 /* Put the most significant integerPartWidth bits in "part". */
3446 if (--count == partsCount)
3447 part = 0; /* An imaginary higher zero part. */
3448 else
3449 part = significand[count] << shift;
3451 if (count && shift)
3452 part |= significand[count - 1] >> (integerPartWidth - shift);
3454 /* Convert as much of "part" to hexdigits as we can. */
3455 unsigned int curDigits = integerPartWidth / 4;
3457 if (curDigits > outputDigits)
3458 curDigits = outputDigits;
3459 dst += partAsHex (dst, part, curDigits, hexDigitChars);
3460 outputDigits -= curDigits;
3463 if (roundUp) {
3464 char *q = dst;
3466 /* Note that hexDigitChars has a trailing '0'. */
3467 do {
3468 q--;
3469 *q = hexDigitChars[hexDigitValue (*q) + 1];
3470 } while (*q == '0');
3471 assert(q >= p);
3472 } else {
3473 /* Add trailing zeroes. */
3474 memset (dst, '0', outputDigits);
3475 dst += outputDigits;
3478 /* Move the most significant digit to before the point, and if there
3479 is something after the decimal point add it. This must come
3480 after rounding above. */
3481 p[-1] = p[0];
3482 if (dst -1 == p)
3483 dst--;
3484 else
3485 p[0] = '.';
3487 /* Finally output the exponent. */
3488 *dst++ = upperCase ? 'P': 'p';
3490 return writeSignedDecimal (dst, exponent);
3493 hash_code hash_value(const IEEEFloat &Arg) {
3494 if (!Arg.isFiniteNonZero())
3495 return hash_combine((uint8_t)Arg.category,
3496 // NaN has no sign, fix it at zero.
3497 Arg.isNaN() ? (uint8_t)0 : (uint8_t)Arg.sign,
3498 Arg.semantics->precision);
3500 // Normal floats need their exponent and significand hashed.
3501 return hash_combine((uint8_t)Arg.category, (uint8_t)Arg.sign,
3502 Arg.semantics->precision, Arg.exponent,
3503 hash_combine_range(
3504 Arg.significandParts(),
3505 Arg.significandParts() + Arg.partCount()));
3508 // Conversion from APFloat to/from host float/double. It may eventually be
3509 // possible to eliminate these and have everybody deal with APFloats, but that
3510 // will take a while. This approach will not easily extend to long double.
3511 // Current implementation requires integerPartWidth==64, which is correct at
3512 // the moment but could be made more general.
3514 // Denormals have exponent minExponent in APFloat, but minExponent-1 in
3515 // the actual IEEE respresentations. We compensate for that here.
3517 APInt IEEEFloat::convertF80LongDoubleAPFloatToAPInt() const {
3518 assert(semantics == (const llvm::fltSemantics*)&semX87DoubleExtended);
3519 assert(partCount()==2);
3521 uint64_t myexponent, mysignificand;
3523 if (isFiniteNonZero()) {
3524 myexponent = exponent+16383; //bias
3525 mysignificand = significandParts()[0];
3526 if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL))
3527 myexponent = 0; // denormal
3528 } else if (category==fcZero) {
3529 myexponent = 0;
3530 mysignificand = 0;
3531 } else if (category==fcInfinity) {
3532 myexponent = 0x7fff;
3533 mysignificand = 0x8000000000000000ULL;
3534 } else {
3535 assert(category == fcNaN && "Unknown category");
3536 myexponent = 0x7fff;
3537 mysignificand = significandParts()[0];
3540 uint64_t words[2];
3541 words[0] = mysignificand;
3542 words[1] = ((uint64_t)(sign & 1) << 15) |
3543 (myexponent & 0x7fffLL);
3544 return APInt(80, words);
3547 APInt IEEEFloat::convertPPCDoubleDoubleLegacyAPFloatToAPInt() const {
3548 assert(semantics == (const llvm::fltSemantics *)&semPPCDoubleDoubleLegacy);
3549 assert(partCount()==2);
3551 uint64_t words[2];
3552 opStatus fs;
3553 bool losesInfo;
3555 // Convert number to double. To avoid spurious underflows, we re-
3556 // normalize against the "double" minExponent first, and only *then*
3557 // truncate the mantissa. The result of that second conversion
3558 // may be inexact, but should never underflow.
3559 // Declare fltSemantics before APFloat that uses it (and
3560 // saves pointer to it) to ensure correct destruction order.
3561 fltSemantics extendedSemantics = *semantics;
3562 extendedSemantics.minExponent = semIEEEdouble.minExponent;
3563 IEEEFloat extended(*this);
3564 fs = extended.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo);
3565 assert(fs == opOK && !losesInfo);
3566 (void)fs;
3568 IEEEFloat u(extended);
3569 fs = u.convert(semIEEEdouble, rmNearestTiesToEven, &losesInfo);
3570 assert(fs == opOK || fs == opInexact);
3571 (void)fs;
3572 words[0] = *u.convertDoubleAPFloatToAPInt().getRawData();
3574 // If conversion was exact or resulted in a special case, we're done;
3575 // just set the second double to zero. Otherwise, re-convert back to
3576 // the extended format and compute the difference. This now should
3577 // convert exactly to double.
3578 if (u.isFiniteNonZero() && losesInfo) {
3579 fs = u.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo);
3580 assert(fs == opOK && !losesInfo);
3581 (void)fs;
3583 IEEEFloat v(extended);
3584 v.subtract(u, rmNearestTiesToEven);
3585 fs = v.convert(semIEEEdouble, rmNearestTiesToEven, &losesInfo);
3586 assert(fs == opOK && !losesInfo);
3587 (void)fs;
3588 words[1] = *v.convertDoubleAPFloatToAPInt().getRawData();
3589 } else {
3590 words[1] = 0;
3593 return APInt(128, words);
3596 template <const fltSemantics &S>
3597 APInt IEEEFloat::convertIEEEFloatToAPInt() const {
3598 assert(semantics == &S);
3599 const int bias =
3600 (semantics == &semFloat8E8M0FNU) ? -S.minExponent : -(S.minExponent - 1);
3601 constexpr unsigned int trailing_significand_bits = S.precision - 1;
3602 constexpr int integer_bit_part = trailing_significand_bits / integerPartWidth;
3603 constexpr integerPart integer_bit =
3604 integerPart{1} << (trailing_significand_bits % integerPartWidth);
3605 constexpr uint64_t significand_mask = integer_bit - 1;
3606 constexpr unsigned int exponent_bits =
3607 trailing_significand_bits ? (S.sizeInBits - 1 - trailing_significand_bits)
3608 : S.sizeInBits;
3609 static_assert(exponent_bits < 64);
3610 constexpr uint64_t exponent_mask = (uint64_t{1} << exponent_bits) - 1;
3612 uint64_t myexponent;
3613 std::array<integerPart, partCountForBits(trailing_significand_bits)>
3614 mysignificand;
3616 if (isFiniteNonZero()) {
3617 myexponent = exponent + bias;
3618 std::copy_n(significandParts(), mysignificand.size(),
3619 mysignificand.begin());
3620 if (myexponent == 1 &&
3621 !(significandParts()[integer_bit_part] & integer_bit))
3622 myexponent = 0; // denormal
3623 } else if (category == fcZero) {
3624 if (!S.hasZero)
3625 llvm_unreachable("semantics does not support zero!");
3626 myexponent = ::exponentZero(S) + bias;
3627 mysignificand.fill(0);
3628 } else if (category == fcInfinity) {
3629 if (S.nonFiniteBehavior == fltNonfiniteBehavior::NanOnly ||
3630 S.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)
3631 llvm_unreachable("semantics don't support inf!");
3632 myexponent = ::exponentInf(S) + bias;
3633 mysignificand.fill(0);
3634 } else {
3635 assert(category == fcNaN && "Unknown category!");
3636 if (S.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)
3637 llvm_unreachable("semantics don't support NaN!");
3638 myexponent = ::exponentNaN(S) + bias;
3639 std::copy_n(significandParts(), mysignificand.size(),
3640 mysignificand.begin());
3642 std::array<uint64_t, (S.sizeInBits + 63) / 64> words;
3643 auto words_iter =
3644 std::copy_n(mysignificand.begin(), mysignificand.size(), words.begin());
3645 if constexpr (significand_mask != 0 || trailing_significand_bits == 0) {
3646 // Clear the integer bit.
3647 words[mysignificand.size() - 1] &= significand_mask;
3649 std::fill(words_iter, words.end(), uint64_t{0});
3650 constexpr size_t last_word = words.size() - 1;
3651 uint64_t shifted_sign = static_cast<uint64_t>(sign & 1)
3652 << ((S.sizeInBits - 1) % 64);
3653 words[last_word] |= shifted_sign;
3654 uint64_t shifted_exponent = (myexponent & exponent_mask)
3655 << (trailing_significand_bits % 64);
3656 words[last_word] |= shifted_exponent;
3657 if constexpr (last_word == 0) {
3658 return APInt(S.sizeInBits, words[0]);
3660 return APInt(S.sizeInBits, words);
3663 APInt IEEEFloat::convertQuadrupleAPFloatToAPInt() const {
3664 assert(partCount() == 2);
3665 return convertIEEEFloatToAPInt<semIEEEquad>();
3668 APInt IEEEFloat::convertDoubleAPFloatToAPInt() const {
3669 assert(partCount()==1);
3670 return convertIEEEFloatToAPInt<semIEEEdouble>();
3673 APInt IEEEFloat::convertFloatAPFloatToAPInt() const {
3674 assert(partCount()==1);
3675 return convertIEEEFloatToAPInt<semIEEEsingle>();
3678 APInt IEEEFloat::convertBFloatAPFloatToAPInt() const {
3679 assert(partCount() == 1);
3680 return convertIEEEFloatToAPInt<semBFloat>();
3683 APInt IEEEFloat::convertHalfAPFloatToAPInt() const {
3684 assert(partCount()==1);
3685 return convertIEEEFloatToAPInt<semIEEEhalf>();
3688 APInt IEEEFloat::convertFloat8E5M2APFloatToAPInt() const {
3689 assert(partCount() == 1);
3690 return convertIEEEFloatToAPInt<semFloat8E5M2>();
3693 APInt IEEEFloat::convertFloat8E5M2FNUZAPFloatToAPInt() const {
3694 assert(partCount() == 1);
3695 return convertIEEEFloatToAPInt<semFloat8E5M2FNUZ>();
3698 APInt IEEEFloat::convertFloat8E4M3APFloatToAPInt() const {
3699 assert(partCount() == 1);
3700 return convertIEEEFloatToAPInt<semFloat8E4M3>();
3703 APInt IEEEFloat::convertFloat8E4M3FNAPFloatToAPInt() const {
3704 assert(partCount() == 1);
3705 return convertIEEEFloatToAPInt<semFloat8E4M3FN>();
3708 APInt IEEEFloat::convertFloat8E4M3FNUZAPFloatToAPInt() const {
3709 assert(partCount() == 1);
3710 return convertIEEEFloatToAPInt<semFloat8E4M3FNUZ>();
3713 APInt IEEEFloat::convertFloat8E4M3B11FNUZAPFloatToAPInt() const {
3714 assert(partCount() == 1);
3715 return convertIEEEFloatToAPInt<semFloat8E4M3B11FNUZ>();
3718 APInt IEEEFloat::convertFloat8E3M4APFloatToAPInt() const {
3719 assert(partCount() == 1);
3720 return convertIEEEFloatToAPInt<semFloat8E3M4>();
3723 APInt IEEEFloat::convertFloatTF32APFloatToAPInt() const {
3724 assert(partCount() == 1);
3725 return convertIEEEFloatToAPInt<semFloatTF32>();
3728 APInt IEEEFloat::convertFloat8E8M0FNUAPFloatToAPInt() const {
3729 assert(partCount() == 1);
3730 return convertIEEEFloatToAPInt<semFloat8E8M0FNU>();
3733 APInt IEEEFloat::convertFloat6E3M2FNAPFloatToAPInt() const {
3734 assert(partCount() == 1);
3735 return convertIEEEFloatToAPInt<semFloat6E3M2FN>();
3738 APInt IEEEFloat::convertFloat6E2M3FNAPFloatToAPInt() const {
3739 assert(partCount() == 1);
3740 return convertIEEEFloatToAPInt<semFloat6E2M3FN>();
3743 APInt IEEEFloat::convertFloat4E2M1FNAPFloatToAPInt() const {
3744 assert(partCount() == 1);
3745 return convertIEEEFloatToAPInt<semFloat4E2M1FN>();
3748 // This function creates an APInt that is just a bit map of the floating
3749 // point constant as it would appear in memory. It is not a conversion,
3750 // and treating the result as a normal integer is unlikely to be useful.
3752 APInt IEEEFloat::bitcastToAPInt() const {
3753 if (semantics == (const llvm::fltSemantics*)&semIEEEhalf)
3754 return convertHalfAPFloatToAPInt();
3756 if (semantics == (const llvm::fltSemantics *)&semBFloat)
3757 return convertBFloatAPFloatToAPInt();
3759 if (semantics == (const llvm::fltSemantics*)&semIEEEsingle)
3760 return convertFloatAPFloatToAPInt();
3762 if (semantics == (const llvm::fltSemantics*)&semIEEEdouble)
3763 return convertDoubleAPFloatToAPInt();
3765 if (semantics == (const llvm::fltSemantics*)&semIEEEquad)
3766 return convertQuadrupleAPFloatToAPInt();
3768 if (semantics == (const llvm::fltSemantics *)&semPPCDoubleDoubleLegacy)
3769 return convertPPCDoubleDoubleLegacyAPFloatToAPInt();
3771 if (semantics == (const llvm::fltSemantics *)&semFloat8E5M2)
3772 return convertFloat8E5M2APFloatToAPInt();
3774 if (semantics == (const llvm::fltSemantics *)&semFloat8E5M2FNUZ)
3775 return convertFloat8E5M2FNUZAPFloatToAPInt();
3777 if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3)
3778 return convertFloat8E4M3APFloatToAPInt();
3780 if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3FN)
3781 return convertFloat8E4M3FNAPFloatToAPInt();
3783 if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3FNUZ)
3784 return convertFloat8E4M3FNUZAPFloatToAPInt();
3786 if (semantics == (const llvm::fltSemantics *)&semFloat8E4M3B11FNUZ)
3787 return convertFloat8E4M3B11FNUZAPFloatToAPInt();
3789 if (semantics == (const llvm::fltSemantics *)&semFloat8E3M4)
3790 return convertFloat8E3M4APFloatToAPInt();
3792 if (semantics == (const llvm::fltSemantics *)&semFloatTF32)
3793 return convertFloatTF32APFloatToAPInt();
3795 if (semantics == (const llvm::fltSemantics *)&semFloat8E8M0FNU)
3796 return convertFloat8E8M0FNUAPFloatToAPInt();
3798 if (semantics == (const llvm::fltSemantics *)&semFloat6E3M2FN)
3799 return convertFloat6E3M2FNAPFloatToAPInt();
3801 if (semantics == (const llvm::fltSemantics *)&semFloat6E2M3FN)
3802 return convertFloat6E2M3FNAPFloatToAPInt();
3804 if (semantics == (const llvm::fltSemantics *)&semFloat4E2M1FN)
3805 return convertFloat4E2M1FNAPFloatToAPInt();
3807 assert(semantics == (const llvm::fltSemantics*)&semX87DoubleExtended &&
3808 "unknown format!");
3809 return convertF80LongDoubleAPFloatToAPInt();
3812 float IEEEFloat::convertToFloat() const {
3813 assert(semantics == (const llvm::fltSemantics*)&semIEEEsingle &&
3814 "Float semantics are not IEEEsingle");
3815 APInt api = bitcastToAPInt();
3816 return api.bitsToFloat();
3819 double IEEEFloat::convertToDouble() const {
3820 assert(semantics == (const llvm::fltSemantics*)&semIEEEdouble &&
3821 "Float semantics are not IEEEdouble");
3822 APInt api = bitcastToAPInt();
3823 return api.bitsToDouble();
3826 #ifdef HAS_IEE754_FLOAT128
3827 float128 IEEEFloat::convertToQuad() const {
3828 assert(semantics == (const llvm::fltSemantics *)&semIEEEquad &&
3829 "Float semantics are not IEEEquads");
3830 APInt api = bitcastToAPInt();
3831 return api.bitsToQuad();
3833 #endif
3835 /// Integer bit is explicit in this format. Intel hardware (387 and later)
3836 /// does not support these bit patterns:
3837 /// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity")
3838 /// exponent = all 1's, integer bit 0, significand nonzero ("pseudoNaN")
3839 /// exponent!=0 nor all 1's, integer bit 0 ("unnormal")
3840 /// exponent = 0, integer bit 1 ("pseudodenormal")
3841 /// At the moment, the first three are treated as NaNs, the last one as Normal.
3842 void IEEEFloat::initFromF80LongDoubleAPInt(const APInt &api) {
3843 uint64_t i1 = api.getRawData()[0];
3844 uint64_t i2 = api.getRawData()[1];
3845 uint64_t myexponent = (i2 & 0x7fff);
3846 uint64_t mysignificand = i1;
3847 uint8_t myintegerbit = mysignificand >> 63;
3849 initialize(&semX87DoubleExtended);
3850 assert(partCount()==2);
3852 sign = static_cast<unsigned int>(i2>>15);
3853 if (myexponent == 0 && mysignificand == 0) {
3854 makeZero(sign);
3855 } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) {
3856 makeInf(sign);
3857 } else if ((myexponent == 0x7fff && mysignificand != 0x8000000000000000ULL) ||
3858 (myexponent != 0x7fff && myexponent != 0 && myintegerbit == 0)) {
3859 category = fcNaN;
3860 exponent = exponentNaN();
3861 significandParts()[0] = mysignificand;
3862 significandParts()[1] = 0;
3863 } else {
3864 category = fcNormal;
3865 exponent = myexponent - 16383;
3866 significandParts()[0] = mysignificand;
3867 significandParts()[1] = 0;
3868 if (myexponent==0) // denormal
3869 exponent = -16382;
3873 void IEEEFloat::initFromPPCDoubleDoubleLegacyAPInt(const APInt &api) {
3874 uint64_t i1 = api.getRawData()[0];
3875 uint64_t i2 = api.getRawData()[1];
3876 opStatus fs;
3877 bool losesInfo;
3879 // Get the first double and convert to our format.
3880 initFromDoubleAPInt(APInt(64, i1));
3881 fs = convert(semPPCDoubleDoubleLegacy, rmNearestTiesToEven, &losesInfo);
3882 assert(fs == opOK && !losesInfo);
3883 (void)fs;
3885 // Unless we have a special case, add in second double.
3886 if (isFiniteNonZero()) {
3887 IEEEFloat v(semIEEEdouble, APInt(64, i2));
3888 fs = v.convert(semPPCDoubleDoubleLegacy, rmNearestTiesToEven, &losesInfo);
3889 assert(fs == opOK && !losesInfo);
3890 (void)fs;
3892 add(v, rmNearestTiesToEven);
3896 // The E8M0 format has the following characteristics:
3897 // It is an 8-bit unsigned format with only exponents (no actual significand).
3898 // No encodings for {zero, infinities or denorms}.
3899 // NaN is represented by all 1's.
3900 // Bias is 127.
3901 void IEEEFloat::initFromFloat8E8M0FNUAPInt(const APInt &api) {
3902 const uint64_t exponent_mask = 0xff;
3903 uint64_t val = api.getRawData()[0];
3904 uint64_t myexponent = (val & exponent_mask);
3906 initialize(&semFloat8E8M0FNU);
3907 assert(partCount() == 1);
3909 // This format has unsigned representation only
3910 sign = 0;
3912 // Set the significand
3913 // This format does not have any significand but the 'Pth' precision bit is
3914 // always set to 1 for consistency in APFloat's internal representation.
3915 uint64_t mysignificand = 1;
3916 significandParts()[0] = mysignificand;
3918 // This format can either have a NaN or fcNormal
3919 // All 1's i.e. 255 is a NaN
3920 if (val == exponent_mask) {
3921 category = fcNaN;
3922 exponent = exponentNaN();
3923 return;
3925 // Handle fcNormal...
3926 category = fcNormal;
3927 exponent = myexponent - 127; // 127 is bias
3929 template <const fltSemantics &S>
3930 void IEEEFloat::initFromIEEEAPInt(const APInt &api) {
3931 assert(api.getBitWidth() == S.sizeInBits);
3932 constexpr integerPart integer_bit = integerPart{1}
3933 << ((S.precision - 1) % integerPartWidth);
3934 constexpr uint64_t significand_mask = integer_bit - 1;
3935 constexpr unsigned int trailing_significand_bits = S.precision - 1;
3936 constexpr unsigned int stored_significand_parts =
3937 partCountForBits(trailing_significand_bits);
3938 constexpr unsigned int exponent_bits =
3939 S.sizeInBits - 1 - trailing_significand_bits;
3940 static_assert(exponent_bits < 64);
3941 constexpr uint64_t exponent_mask = (uint64_t{1} << exponent_bits) - 1;
3942 constexpr int bias = -(S.minExponent - 1);
3944 // Copy the bits of the significand. We need to clear out the exponent and
3945 // sign bit in the last word.
3946 std::array<integerPart, stored_significand_parts> mysignificand;
3947 std::copy_n(api.getRawData(), mysignificand.size(), mysignificand.begin());
3948 if constexpr (significand_mask != 0) {
3949 mysignificand[mysignificand.size() - 1] &= significand_mask;
3952 // We assume the last word holds the sign bit, the exponent, and potentially
3953 // some of the trailing significand field.
3954 uint64_t last_word = api.getRawData()[api.getNumWords() - 1];
3955 uint64_t myexponent =
3956 (last_word >> (trailing_significand_bits % 64)) & exponent_mask;
3958 initialize(&S);
3959 assert(partCount() == mysignificand.size());
3961 sign = static_cast<unsigned int>(last_word >> ((S.sizeInBits - 1) % 64));
3963 bool all_zero_significand =
3964 llvm::all_of(mysignificand, [](integerPart bits) { return bits == 0; });
3966 bool is_zero = myexponent == 0 && all_zero_significand;
3968 if constexpr (S.nonFiniteBehavior == fltNonfiniteBehavior::IEEE754) {
3969 if (myexponent - bias == ::exponentInf(S) && all_zero_significand) {
3970 makeInf(sign);
3971 return;
3975 bool is_nan = false;
3977 if constexpr (S.nanEncoding == fltNanEncoding::IEEE) {
3978 is_nan = myexponent - bias == ::exponentNaN(S) && !all_zero_significand;
3979 } else if constexpr (S.nanEncoding == fltNanEncoding::AllOnes) {
3980 bool all_ones_significand =
3981 std::all_of(mysignificand.begin(), mysignificand.end() - 1,
3982 [](integerPart bits) { return bits == ~integerPart{0}; }) &&
3983 (!significand_mask ||
3984 mysignificand[mysignificand.size() - 1] == significand_mask);
3985 is_nan = myexponent - bias == ::exponentNaN(S) && all_ones_significand;
3986 } else if constexpr (S.nanEncoding == fltNanEncoding::NegativeZero) {
3987 is_nan = is_zero && sign;
3990 if (is_nan) {
3991 category = fcNaN;
3992 exponent = ::exponentNaN(S);
3993 std::copy_n(mysignificand.begin(), mysignificand.size(),
3994 significandParts());
3995 return;
3998 if (is_zero) {
3999 makeZero(sign);
4000 return;
4003 category = fcNormal;
4004 exponent = myexponent - bias;
4005 std::copy_n(mysignificand.begin(), mysignificand.size(), significandParts());
4006 if (myexponent == 0) // denormal
4007 exponent = S.minExponent;
4008 else
4009 significandParts()[mysignificand.size()-1] |= integer_bit; // integer bit
4012 void IEEEFloat::initFromQuadrupleAPInt(const APInt &api) {
4013 initFromIEEEAPInt<semIEEEquad>(api);
4016 void IEEEFloat::initFromDoubleAPInt(const APInt &api) {
4017 initFromIEEEAPInt<semIEEEdouble>(api);
4020 void IEEEFloat::initFromFloatAPInt(const APInt &api) {
4021 initFromIEEEAPInt<semIEEEsingle>(api);
4024 void IEEEFloat::initFromBFloatAPInt(const APInt &api) {
4025 initFromIEEEAPInt<semBFloat>(api);
4028 void IEEEFloat::initFromHalfAPInt(const APInt &api) {
4029 initFromIEEEAPInt<semIEEEhalf>(api);
4032 void IEEEFloat::initFromFloat8E5M2APInt(const APInt &api) {
4033 initFromIEEEAPInt<semFloat8E5M2>(api);
4036 void IEEEFloat::initFromFloat8E5M2FNUZAPInt(const APInt &api) {
4037 initFromIEEEAPInt<semFloat8E5M2FNUZ>(api);
4040 void IEEEFloat::initFromFloat8E4M3APInt(const APInt &api) {
4041 initFromIEEEAPInt<semFloat8E4M3>(api);
4044 void IEEEFloat::initFromFloat8E4M3FNAPInt(const APInt &api) {
4045 initFromIEEEAPInt<semFloat8E4M3FN>(api);
4048 void IEEEFloat::initFromFloat8E4M3FNUZAPInt(const APInt &api) {
4049 initFromIEEEAPInt<semFloat8E4M3FNUZ>(api);
4052 void IEEEFloat::initFromFloat8E4M3B11FNUZAPInt(const APInt &api) {
4053 initFromIEEEAPInt<semFloat8E4M3B11FNUZ>(api);
4056 void IEEEFloat::initFromFloat8E3M4APInt(const APInt &api) {
4057 initFromIEEEAPInt<semFloat8E3M4>(api);
4060 void IEEEFloat::initFromFloatTF32APInt(const APInt &api) {
4061 initFromIEEEAPInt<semFloatTF32>(api);
4064 void IEEEFloat::initFromFloat6E3M2FNAPInt(const APInt &api) {
4065 initFromIEEEAPInt<semFloat6E3M2FN>(api);
4068 void IEEEFloat::initFromFloat6E2M3FNAPInt(const APInt &api) {
4069 initFromIEEEAPInt<semFloat6E2M3FN>(api);
4072 void IEEEFloat::initFromFloat4E2M1FNAPInt(const APInt &api) {
4073 initFromIEEEAPInt<semFloat4E2M1FN>(api);
4076 /// Treat api as containing the bits of a floating point number.
4077 void IEEEFloat::initFromAPInt(const fltSemantics *Sem, const APInt &api) {
4078 assert(api.getBitWidth() == Sem->sizeInBits);
4079 if (Sem == &semIEEEhalf)
4080 return initFromHalfAPInt(api);
4081 if (Sem == &semBFloat)
4082 return initFromBFloatAPInt(api);
4083 if (Sem == &semIEEEsingle)
4084 return initFromFloatAPInt(api);
4085 if (Sem == &semIEEEdouble)
4086 return initFromDoubleAPInt(api);
4087 if (Sem == &semX87DoubleExtended)
4088 return initFromF80LongDoubleAPInt(api);
4089 if (Sem == &semIEEEquad)
4090 return initFromQuadrupleAPInt(api);
4091 if (Sem == &semPPCDoubleDoubleLegacy)
4092 return initFromPPCDoubleDoubleLegacyAPInt(api);
4093 if (Sem == &semFloat8E5M2)
4094 return initFromFloat8E5M2APInt(api);
4095 if (Sem == &semFloat8E5M2FNUZ)
4096 return initFromFloat8E5M2FNUZAPInt(api);
4097 if (Sem == &semFloat8E4M3)
4098 return initFromFloat8E4M3APInt(api);
4099 if (Sem == &semFloat8E4M3FN)
4100 return initFromFloat8E4M3FNAPInt(api);
4101 if (Sem == &semFloat8E4M3FNUZ)
4102 return initFromFloat8E4M3FNUZAPInt(api);
4103 if (Sem == &semFloat8E4M3B11FNUZ)
4104 return initFromFloat8E4M3B11FNUZAPInt(api);
4105 if (Sem == &semFloat8E3M4)
4106 return initFromFloat8E3M4APInt(api);
4107 if (Sem == &semFloatTF32)
4108 return initFromFloatTF32APInt(api);
4109 if (Sem == &semFloat8E8M0FNU)
4110 return initFromFloat8E8M0FNUAPInt(api);
4111 if (Sem == &semFloat6E3M2FN)
4112 return initFromFloat6E3M2FNAPInt(api);
4113 if (Sem == &semFloat6E2M3FN)
4114 return initFromFloat6E2M3FNAPInt(api);
4115 if (Sem == &semFloat4E2M1FN)
4116 return initFromFloat4E2M1FNAPInt(api);
4118 llvm_unreachable("unsupported semantics");
4121 /// Make this number the largest magnitude normal number in the given
4122 /// semantics.
4123 void IEEEFloat::makeLargest(bool Negative) {
4124 if (Negative && !semantics->hasSignedRepr)
4125 llvm_unreachable(
4126 "This floating point format does not support signed values");
4127 // We want (in interchange format):
4128 // sign = {Negative}
4129 // exponent = 1..10
4130 // significand = 1..1
4131 category = fcNormal;
4132 sign = Negative;
4133 exponent = semantics->maxExponent;
4135 // Use memset to set all but the highest integerPart to all ones.
4136 integerPart *significand = significandParts();
4137 unsigned PartCount = partCount();
4138 memset(significand, 0xFF, sizeof(integerPart)*(PartCount - 1));
4140 // Set the high integerPart especially setting all unused top bits for
4141 // internal consistency.
4142 const unsigned NumUnusedHighBits =
4143 PartCount*integerPartWidth - semantics->precision;
4144 significand[PartCount - 1] = (NumUnusedHighBits < integerPartWidth)
4145 ? (~integerPart(0) >> NumUnusedHighBits)
4146 : 0;
4147 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly &&
4148 semantics->nanEncoding == fltNanEncoding::AllOnes &&
4149 (semantics->precision > 1))
4150 significand[0] &= ~integerPart(1);
4153 /// Make this number the smallest magnitude denormal number in the given
4154 /// semantics.
4155 void IEEEFloat::makeSmallest(bool Negative) {
4156 if (Negative && !semantics->hasSignedRepr)
4157 llvm_unreachable(
4158 "This floating point format does not support signed values");
4159 // We want (in interchange format):
4160 // sign = {Negative}
4161 // exponent = 0..0
4162 // significand = 0..01
4163 category = fcNormal;
4164 sign = Negative;
4165 exponent = semantics->minExponent;
4166 APInt::tcSet(significandParts(), 1, partCount());
4169 void IEEEFloat::makeSmallestNormalized(bool Negative) {
4170 if (Negative && !semantics->hasSignedRepr)
4171 llvm_unreachable(
4172 "This floating point format does not support signed values");
4173 // We want (in interchange format):
4174 // sign = {Negative}
4175 // exponent = 0..0
4176 // significand = 10..0
4178 category = fcNormal;
4179 zeroSignificand();
4180 sign = Negative;
4181 exponent = semantics->minExponent;
4182 APInt::tcSetBit(significandParts(), semantics->precision - 1);
4185 IEEEFloat::IEEEFloat(const fltSemantics &Sem, const APInt &API) {
4186 initFromAPInt(&Sem, API);
4189 IEEEFloat::IEEEFloat(float f) {
4190 initFromAPInt(&semIEEEsingle, APInt::floatToBits(f));
4193 IEEEFloat::IEEEFloat(double d) {
4194 initFromAPInt(&semIEEEdouble, APInt::doubleToBits(d));
4197 namespace {
4198 void append(SmallVectorImpl<char> &Buffer, StringRef Str) {
4199 Buffer.append(Str.begin(), Str.end());
4202 /// Removes data from the given significand until it is no more
4203 /// precise than is required for the desired precision.
4204 void AdjustToPrecision(APInt &significand,
4205 int &exp, unsigned FormatPrecision) {
4206 unsigned bits = significand.getActiveBits();
4208 // 196/59 is a very slight overestimate of lg_2(10).
4209 unsigned bitsRequired = (FormatPrecision * 196 + 58) / 59;
4211 if (bits <= bitsRequired) return;
4213 unsigned tensRemovable = (bits - bitsRequired) * 59 / 196;
4214 if (!tensRemovable) return;
4216 exp += tensRemovable;
4218 APInt divisor(significand.getBitWidth(), 1);
4219 APInt powten(significand.getBitWidth(), 10);
4220 while (true) {
4221 if (tensRemovable & 1)
4222 divisor *= powten;
4223 tensRemovable >>= 1;
4224 if (!tensRemovable) break;
4225 powten *= powten;
4228 significand = significand.udiv(divisor);
4230 // Truncate the significand down to its active bit count.
4231 significand = significand.trunc(significand.getActiveBits());
4235 void AdjustToPrecision(SmallVectorImpl<char> &buffer,
4236 int &exp, unsigned FormatPrecision) {
4237 unsigned N = buffer.size();
4238 if (N <= FormatPrecision) return;
4240 // The most significant figures are the last ones in the buffer.
4241 unsigned FirstSignificant = N - FormatPrecision;
4243 // Round.
4244 // FIXME: this probably shouldn't use 'round half up'.
4246 // Rounding down is just a truncation, except we also want to drop
4247 // trailing zeros from the new result.
4248 if (buffer[FirstSignificant - 1] < '5') {
4249 while (FirstSignificant < N && buffer[FirstSignificant] == '0')
4250 FirstSignificant++;
4252 exp += FirstSignificant;
4253 buffer.erase(&buffer[0], &buffer[FirstSignificant]);
4254 return;
4257 // Rounding up requires a decimal add-with-carry. If we continue
4258 // the carry, the newly-introduced zeros will just be truncated.
4259 for (unsigned I = FirstSignificant; I != N; ++I) {
4260 if (buffer[I] == '9') {
4261 FirstSignificant++;
4262 } else {
4263 buffer[I]++;
4264 break;
4268 // If we carried through, we have exactly one digit of precision.
4269 if (FirstSignificant == N) {
4270 exp += FirstSignificant;
4271 buffer.clear();
4272 buffer.push_back('1');
4273 return;
4276 exp += FirstSignificant;
4277 buffer.erase(&buffer[0], &buffer[FirstSignificant]);
4280 void toStringImpl(SmallVectorImpl<char> &Str, const bool isNeg, int exp,
4281 APInt significand, unsigned FormatPrecision,
4282 unsigned FormatMaxPadding, bool TruncateZero) {
4283 const int semanticsPrecision = significand.getBitWidth();
4285 if (isNeg)
4286 Str.push_back('-');
4288 // Set FormatPrecision if zero. We want to do this before we
4289 // truncate trailing zeros, as those are part of the precision.
4290 if (!FormatPrecision) {
4291 // We use enough digits so the number can be round-tripped back to an
4292 // APFloat. The formula comes from "How to Print Floating-Point Numbers
4293 // Accurately" by Steele and White.
4294 // FIXME: Using a formula based purely on the precision is conservative;
4295 // we can print fewer digits depending on the actual value being printed.
4297 // FormatPrecision = 2 + floor(significandBits / lg_2(10))
4298 FormatPrecision = 2 + semanticsPrecision * 59 / 196;
4301 // Ignore trailing binary zeros.
4302 int trailingZeros = significand.countr_zero();
4303 exp += trailingZeros;
4304 significand.lshrInPlace(trailingZeros);
4306 // Change the exponent from 2^e to 10^e.
4307 if (exp == 0) {
4308 // Nothing to do.
4309 } else if (exp > 0) {
4310 // Just shift left.
4311 significand = significand.zext(semanticsPrecision + exp);
4312 significand <<= exp;
4313 exp = 0;
4314 } else { /* exp < 0 */
4315 int texp = -exp;
4317 // We transform this using the identity:
4318 // (N)(2^-e) == (N)(5^e)(10^-e)
4319 // This means we have to multiply N (the significand) by 5^e.
4320 // To avoid overflow, we have to operate on numbers large
4321 // enough to store N * 5^e:
4322 // log2(N * 5^e) == log2(N) + e * log2(5)
4323 // <= semantics->precision + e * 137 / 59
4324 // (log_2(5) ~ 2.321928 < 2.322034 ~ 137/59)
4326 unsigned precision = semanticsPrecision + (137 * texp + 136) / 59;
4328 // Multiply significand by 5^e.
4329 // N * 5^0101 == N * 5^(1*1) * 5^(0*2) * 5^(1*4) * 5^(0*8)
4330 significand = significand.zext(precision);
4331 APInt five_to_the_i(precision, 5);
4332 while (true) {
4333 if (texp & 1)
4334 significand *= five_to_the_i;
4336 texp >>= 1;
4337 if (!texp)
4338 break;
4339 five_to_the_i *= five_to_the_i;
4343 AdjustToPrecision(significand, exp, FormatPrecision);
4345 SmallVector<char, 256> buffer;
4347 // Fill the buffer.
4348 unsigned precision = significand.getBitWidth();
4349 if (precision < 4) {
4350 // We need enough precision to store the value 10.
4351 precision = 4;
4352 significand = significand.zext(precision);
4354 APInt ten(precision, 10);
4355 APInt digit(precision, 0);
4357 bool inTrail = true;
4358 while (significand != 0) {
4359 // digit <- significand % 10
4360 // significand <- significand / 10
4361 APInt::udivrem(significand, ten, significand, digit);
4363 unsigned d = digit.getZExtValue();
4365 // Drop trailing zeros.
4366 if (inTrail && !d)
4367 exp++;
4368 else {
4369 buffer.push_back((char) ('0' + d));
4370 inTrail = false;
4374 assert(!buffer.empty() && "no characters in buffer!");
4376 // Drop down to FormatPrecision.
4377 // TODO: don't do more precise calculations above than are required.
4378 AdjustToPrecision(buffer, exp, FormatPrecision);
4380 unsigned NDigits = buffer.size();
4382 // Check whether we should use scientific notation.
4383 bool FormatScientific;
4384 if (!FormatMaxPadding)
4385 FormatScientific = true;
4386 else {
4387 if (exp >= 0) {
4388 // 765e3 --> 765000
4389 // ^^^
4390 // But we shouldn't make the number look more precise than it is.
4391 FormatScientific = ((unsigned) exp > FormatMaxPadding ||
4392 NDigits + (unsigned) exp > FormatPrecision);
4393 } else {
4394 // Power of the most significant digit.
4395 int MSD = exp + (int) (NDigits - 1);
4396 if (MSD >= 0) {
4397 // 765e-2 == 7.65
4398 FormatScientific = false;
4399 } else {
4400 // 765e-5 == 0.00765
4401 // ^ ^^
4402 FormatScientific = ((unsigned) -MSD) > FormatMaxPadding;
4407 // Scientific formatting is pretty straightforward.
4408 if (FormatScientific) {
4409 exp += (NDigits - 1);
4411 Str.push_back(buffer[NDigits-1]);
4412 Str.push_back('.');
4413 if (NDigits == 1 && TruncateZero)
4414 Str.push_back('0');
4415 else
4416 for (unsigned I = 1; I != NDigits; ++I)
4417 Str.push_back(buffer[NDigits-1-I]);
4418 // Fill with zeros up to FormatPrecision.
4419 if (!TruncateZero && FormatPrecision > NDigits - 1)
4420 Str.append(FormatPrecision - NDigits + 1, '0');
4421 // For !TruncateZero we use lower 'e'.
4422 Str.push_back(TruncateZero ? 'E' : 'e');
4424 Str.push_back(exp >= 0 ? '+' : '-');
4425 if (exp < 0)
4426 exp = -exp;
4427 SmallVector<char, 6> expbuf;
4428 do {
4429 expbuf.push_back((char) ('0' + (exp % 10)));
4430 exp /= 10;
4431 } while (exp);
4432 // Exponent always at least two digits if we do not truncate zeros.
4433 if (!TruncateZero && expbuf.size() < 2)
4434 expbuf.push_back('0');
4435 for (unsigned I = 0, E = expbuf.size(); I != E; ++I)
4436 Str.push_back(expbuf[E-1-I]);
4437 return;
4440 // Non-scientific, positive exponents.
4441 if (exp >= 0) {
4442 for (unsigned I = 0; I != NDigits; ++I)
4443 Str.push_back(buffer[NDigits-1-I]);
4444 for (unsigned I = 0; I != (unsigned) exp; ++I)
4445 Str.push_back('0');
4446 return;
4449 // Non-scientific, negative exponents.
4451 // The number of digits to the left of the decimal point.
4452 int NWholeDigits = exp + (int) NDigits;
4454 unsigned I = 0;
4455 if (NWholeDigits > 0) {
4456 for (; I != (unsigned) NWholeDigits; ++I)
4457 Str.push_back(buffer[NDigits-I-1]);
4458 Str.push_back('.');
4459 } else {
4460 unsigned NZeros = 1 + (unsigned) -NWholeDigits;
4462 Str.push_back('0');
4463 Str.push_back('.');
4464 for (unsigned Z = 1; Z != NZeros; ++Z)
4465 Str.push_back('0');
4468 for (; I != NDigits; ++I)
4469 Str.push_back(buffer[NDigits-I-1]);
4472 } // namespace
4474 void IEEEFloat::toString(SmallVectorImpl<char> &Str, unsigned FormatPrecision,
4475 unsigned FormatMaxPadding, bool TruncateZero) const {
4476 switch (category) {
4477 case fcInfinity:
4478 if (isNegative())
4479 return append(Str, "-Inf");
4480 else
4481 return append(Str, "+Inf");
4483 case fcNaN: return append(Str, "NaN");
4485 case fcZero:
4486 if (isNegative())
4487 Str.push_back('-');
4489 if (!FormatMaxPadding) {
4490 if (TruncateZero)
4491 append(Str, "0.0E+0");
4492 else {
4493 append(Str, "0.0");
4494 if (FormatPrecision > 1)
4495 Str.append(FormatPrecision - 1, '0');
4496 append(Str, "e+00");
4498 } else
4499 Str.push_back('0');
4500 return;
4502 case fcNormal:
4503 break;
4506 // Decompose the number into an APInt and an exponent.
4507 int exp = exponent - ((int) semantics->precision - 1);
4508 APInt significand(
4509 semantics->precision,
4510 ArrayRef(significandParts(), partCountForBits(semantics->precision)));
4512 toStringImpl(Str, isNegative(), exp, significand, FormatPrecision,
4513 FormatMaxPadding, TruncateZero);
4517 bool IEEEFloat::getExactInverse(APFloat *inv) const {
4518 // Special floats and denormals have no exact inverse.
4519 if (!isFiniteNonZero())
4520 return false;
4522 // Check that the number is a power of two by making sure that only the
4523 // integer bit is set in the significand.
4524 if (significandLSB() != semantics->precision - 1)
4525 return false;
4527 // Get the inverse.
4528 IEEEFloat reciprocal(*semantics, 1ULL);
4529 if (reciprocal.divide(*this, rmNearestTiesToEven) != opOK)
4530 return false;
4532 // Avoid multiplication with a denormal, it is not safe on all platforms and
4533 // may be slower than a normal division.
4534 if (reciprocal.isDenormal())
4535 return false;
4537 assert(reciprocal.isFiniteNonZero() &&
4538 reciprocal.significandLSB() == reciprocal.semantics->precision - 1);
4540 if (inv)
4541 *inv = APFloat(reciprocal, *semantics);
4543 return true;
4546 int IEEEFloat::getExactLog2Abs() const {
4547 if (!isFinite() || isZero())
4548 return INT_MIN;
4550 const integerPart *Parts = significandParts();
4551 const int PartCount = partCountForBits(semantics->precision);
4553 int PopCount = 0;
4554 for (int i = 0; i < PartCount; ++i) {
4555 PopCount += llvm::popcount(Parts[i]);
4556 if (PopCount > 1)
4557 return INT_MIN;
4560 if (exponent != semantics->minExponent)
4561 return exponent;
4563 int CountrParts = 0;
4564 for (int i = 0; i < PartCount;
4565 ++i, CountrParts += APInt::APINT_BITS_PER_WORD) {
4566 if (Parts[i] != 0) {
4567 return exponent - semantics->precision + CountrParts +
4568 llvm::countr_zero(Parts[i]) + 1;
4572 llvm_unreachable("didn't find the set bit");
4575 bool IEEEFloat::isSignaling() const {
4576 if (!isNaN())
4577 return false;
4578 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly ||
4579 semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)
4580 return false;
4582 // IEEE-754R 2008 6.2.1: A signaling NaN bit string should be encoded with the
4583 // first bit of the trailing significand being 0.
4584 return !APInt::tcExtractBit(significandParts(), semantics->precision - 2);
4587 /// IEEE-754R 2008 5.3.1: nextUp/nextDown.
4589 /// *NOTE* since nextDown(x) = -nextUp(-x), we only implement nextUp with
4590 /// appropriate sign switching before/after the computation.
4591 APFloat::opStatus IEEEFloat::next(bool nextDown) {
4592 // If we are performing nextDown, swap sign so we have -x.
4593 if (nextDown)
4594 changeSign();
4596 // Compute nextUp(x)
4597 opStatus result = opOK;
4599 // Handle each float category separately.
4600 switch (category) {
4601 case fcInfinity:
4602 // nextUp(+inf) = +inf
4603 if (!isNegative())
4604 break;
4605 // nextUp(-inf) = -getLargest()
4606 makeLargest(true);
4607 break;
4608 case fcNaN:
4609 // IEEE-754R 2008 6.2 Par 2: nextUp(sNaN) = qNaN. Set Invalid flag.
4610 // IEEE-754R 2008 6.2: nextUp(qNaN) = qNaN. Must be identity so we do not
4611 // change the payload.
4612 if (isSignaling()) {
4613 result = opInvalidOp;
4614 // For consistency, propagate the sign of the sNaN to the qNaN.
4615 makeNaN(false, isNegative(), nullptr);
4617 break;
4618 case fcZero:
4619 // nextUp(pm 0) = +getSmallest()
4620 makeSmallest(false);
4621 break;
4622 case fcNormal:
4623 // nextUp(-getSmallest()) = -0
4624 if (isSmallest() && isNegative()) {
4625 APInt::tcSet(significandParts(), 0, partCount());
4626 category = fcZero;
4627 exponent = 0;
4628 if (semantics->nanEncoding == fltNanEncoding::NegativeZero)
4629 sign = false;
4630 if (!semantics->hasZero)
4631 makeSmallestNormalized(false);
4632 break;
4635 if (isLargest() && !isNegative()) {
4636 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
4637 // nextUp(getLargest()) == NAN
4638 makeNaN();
4639 break;
4640 } else if (semantics->nonFiniteBehavior ==
4641 fltNonfiniteBehavior::FiniteOnly) {
4642 // nextUp(getLargest()) == getLargest()
4643 break;
4644 } else {
4645 // nextUp(getLargest()) == INFINITY
4646 APInt::tcSet(significandParts(), 0, partCount());
4647 category = fcInfinity;
4648 exponent = semantics->maxExponent + 1;
4649 break;
4653 // nextUp(normal) == normal + inc.
4654 if (isNegative()) {
4655 // If we are negative, we need to decrement the significand.
4657 // We only cross a binade boundary that requires adjusting the exponent
4658 // if:
4659 // 1. exponent != semantics->minExponent. This implies we are not in the
4660 // smallest binade or are dealing with denormals.
4661 // 2. Our significand excluding the integral bit is all zeros.
4662 bool WillCrossBinadeBoundary =
4663 exponent != semantics->minExponent && isSignificandAllZeros();
4665 // Decrement the significand.
4667 // We always do this since:
4668 // 1. If we are dealing with a non-binade decrement, by definition we
4669 // just decrement the significand.
4670 // 2. If we are dealing with a normal -> normal binade decrement, since
4671 // we have an explicit integral bit the fact that all bits but the
4672 // integral bit are zero implies that subtracting one will yield a
4673 // significand with 0 integral bit and 1 in all other spots. Thus we
4674 // must just adjust the exponent and set the integral bit to 1.
4675 // 3. If we are dealing with a normal -> denormal binade decrement,
4676 // since we set the integral bit to 0 when we represent denormals, we
4677 // just decrement the significand.
4678 integerPart *Parts = significandParts();
4679 APInt::tcDecrement(Parts, partCount());
4681 if (WillCrossBinadeBoundary) {
4682 // Our result is a normal number. Do the following:
4683 // 1. Set the integral bit to 1.
4684 // 2. Decrement the exponent.
4685 APInt::tcSetBit(Parts, semantics->precision - 1);
4686 exponent--;
4688 } else {
4689 // If we are positive, we need to increment the significand.
4691 // We only cross a binade boundary that requires adjusting the exponent if
4692 // the input is not a denormal and all of said input's significand bits
4693 // are set. If all of said conditions are true: clear the significand, set
4694 // the integral bit to 1, and increment the exponent. If we have a
4695 // denormal always increment since moving denormals and the numbers in the
4696 // smallest normal binade have the same exponent in our representation.
4697 // If there are only exponents, any increment always crosses the
4698 // BinadeBoundary.
4699 bool WillCrossBinadeBoundary = !APFloat::hasSignificand(*semantics) ||
4700 (!isDenormal() && isSignificandAllOnes());
4702 if (WillCrossBinadeBoundary) {
4703 integerPart *Parts = significandParts();
4704 APInt::tcSet(Parts, 0, partCount());
4705 APInt::tcSetBit(Parts, semantics->precision - 1);
4706 assert(exponent != semantics->maxExponent &&
4707 "We can not increment an exponent beyond the maxExponent allowed"
4708 " by the given floating point semantics.");
4709 exponent++;
4710 } else {
4711 incrementSignificand();
4714 break;
4717 // If we are performing nextDown, swap sign so we have -nextUp(-x)
4718 if (nextDown)
4719 changeSign();
4721 return result;
4724 APFloatBase::ExponentType IEEEFloat::exponentNaN() const {
4725 return ::exponentNaN(*semantics);
4728 APFloatBase::ExponentType IEEEFloat::exponentInf() const {
4729 return ::exponentInf(*semantics);
4732 APFloatBase::ExponentType IEEEFloat::exponentZero() const {
4733 return ::exponentZero(*semantics);
4736 void IEEEFloat::makeInf(bool Negative) {
4737 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)
4738 llvm_unreachable("This floating point format does not support Inf");
4740 if (semantics->nonFiniteBehavior == fltNonfiniteBehavior::NanOnly) {
4741 // There is no Inf, so make NaN instead.
4742 makeNaN(false, Negative);
4743 return;
4745 category = fcInfinity;
4746 sign = Negative;
4747 exponent = exponentInf();
4748 APInt::tcSet(significandParts(), 0, partCount());
4751 void IEEEFloat::makeZero(bool Negative) {
4752 if (!semantics->hasZero)
4753 llvm_unreachable("This floating point format does not support Zero");
4755 category = fcZero;
4756 sign = Negative;
4757 if (semantics->nanEncoding == fltNanEncoding::NegativeZero) {
4758 // Merge negative zero to positive because 0b10000...000 is used for NaN
4759 sign = false;
4761 exponent = exponentZero();
4762 APInt::tcSet(significandParts(), 0, partCount());
4765 void IEEEFloat::makeQuiet() {
4766 assert(isNaN());
4767 if (semantics->nonFiniteBehavior != fltNonfiniteBehavior::NanOnly)
4768 APInt::tcSetBit(significandParts(), semantics->precision - 2);
4771 int ilogb(const IEEEFloat &Arg) {
4772 if (Arg.isNaN())
4773 return APFloat::IEK_NaN;
4774 if (Arg.isZero())
4775 return APFloat::IEK_Zero;
4776 if (Arg.isInfinity())
4777 return APFloat::IEK_Inf;
4778 if (!Arg.isDenormal())
4779 return Arg.exponent;
4781 IEEEFloat Normalized(Arg);
4782 int SignificandBits = Arg.getSemantics().precision - 1;
4784 Normalized.exponent += SignificandBits;
4785 Normalized.normalize(APFloat::rmNearestTiesToEven, lfExactlyZero);
4786 return Normalized.exponent - SignificandBits;
4789 IEEEFloat scalbn(IEEEFloat X, int Exp, roundingMode RoundingMode) {
4790 auto MaxExp = X.getSemantics().maxExponent;
4791 auto MinExp = X.getSemantics().minExponent;
4793 // If Exp is wildly out-of-scale, simply adding it to X.exponent will
4794 // overflow; clamp it to a safe range before adding, but ensure that the range
4795 // is large enough that the clamp does not change the result. The range we
4796 // need to support is the difference between the largest possible exponent and
4797 // the normalized exponent of half the smallest denormal.
4799 int SignificandBits = X.getSemantics().precision - 1;
4800 int MaxIncrement = MaxExp - (MinExp - SignificandBits) + 1;
4802 // Clamp to one past the range ends to let normalize handle overlflow.
4803 X.exponent += std::clamp(Exp, -MaxIncrement - 1, MaxIncrement);
4804 X.normalize(RoundingMode, lfExactlyZero);
4805 if (X.isNaN())
4806 X.makeQuiet();
4807 return X;
4810 IEEEFloat frexp(const IEEEFloat &Val, int &Exp, roundingMode RM) {
4811 Exp = ilogb(Val);
4813 // Quiet signalling nans.
4814 if (Exp == APFloat::IEK_NaN) {
4815 IEEEFloat Quiet(Val);
4816 Quiet.makeQuiet();
4817 return Quiet;
4820 if (Exp == APFloat::IEK_Inf)
4821 return Val;
4823 // 1 is added because frexp is defined to return a normalized fraction in
4824 // +/-[0.5, 1.0), rather than the usual +/-[1.0, 2.0).
4825 Exp = Exp == APFloat::IEK_Zero ? 0 : Exp + 1;
4826 return scalbn(Val, -Exp, RM);
4829 DoubleAPFloat::DoubleAPFloat(const fltSemantics &S)
4830 : Semantics(&S),
4831 Floats(new APFloat[2]{APFloat(semIEEEdouble), APFloat(semIEEEdouble)}) {
4832 assert(Semantics == &semPPCDoubleDouble);
4835 DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, uninitializedTag)
4836 : Semantics(&S),
4837 Floats(new APFloat[2]{APFloat(semIEEEdouble, uninitialized),
4838 APFloat(semIEEEdouble, uninitialized)}) {
4839 assert(Semantics == &semPPCDoubleDouble);
4842 DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, integerPart I)
4843 : Semantics(&S), Floats(new APFloat[2]{APFloat(semIEEEdouble, I),
4844 APFloat(semIEEEdouble)}) {
4845 assert(Semantics == &semPPCDoubleDouble);
4848 DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, const APInt &I)
4849 : Semantics(&S),
4850 Floats(new APFloat[2]{
4851 APFloat(semIEEEdouble, APInt(64, I.getRawData()[0])),
4852 APFloat(semIEEEdouble, APInt(64, I.getRawData()[1]))}) {
4853 assert(Semantics == &semPPCDoubleDouble);
4856 DoubleAPFloat::DoubleAPFloat(const fltSemantics &S, APFloat &&First,
4857 APFloat &&Second)
4858 : Semantics(&S),
4859 Floats(new APFloat[2]{std::move(First), std::move(Second)}) {
4860 assert(Semantics == &semPPCDoubleDouble);
4861 assert(&Floats[0].getSemantics() == &semIEEEdouble);
4862 assert(&Floats[1].getSemantics() == &semIEEEdouble);
4865 DoubleAPFloat::DoubleAPFloat(const DoubleAPFloat &RHS)
4866 : Semantics(RHS.Semantics),
4867 Floats(RHS.Floats ? new APFloat[2]{APFloat(RHS.Floats[0]),
4868 APFloat(RHS.Floats[1])}
4869 : nullptr) {
4870 assert(Semantics == &semPPCDoubleDouble);
4873 DoubleAPFloat::DoubleAPFloat(DoubleAPFloat &&RHS)
4874 : Semantics(RHS.Semantics), Floats(std::move(RHS.Floats)) {
4875 RHS.Semantics = &semBogus;
4876 assert(Semantics == &semPPCDoubleDouble);
4879 DoubleAPFloat &DoubleAPFloat::operator=(const DoubleAPFloat &RHS) {
4880 if (Semantics == RHS.Semantics && RHS.Floats) {
4881 Floats[0] = RHS.Floats[0];
4882 Floats[1] = RHS.Floats[1];
4883 } else if (this != &RHS) {
4884 this->~DoubleAPFloat();
4885 new (this) DoubleAPFloat(RHS);
4887 return *this;
4890 // Implement addition, subtraction, multiplication and division based on:
4891 // "Software for Doubled-Precision Floating-Point Computations",
4892 // by Seppo Linnainmaa, ACM TOMS vol 7 no 3, September 1981, pages 272-283.
4893 APFloat::opStatus DoubleAPFloat::addImpl(const APFloat &a, const APFloat &aa,
4894 const APFloat &c, const APFloat &cc,
4895 roundingMode RM) {
4896 int Status = opOK;
4897 APFloat z = a;
4898 Status |= z.add(c, RM);
4899 if (!z.isFinite()) {
4900 if (!z.isInfinity()) {
4901 Floats[0] = std::move(z);
4902 Floats[1].makeZero(/* Neg = */ false);
4903 return (opStatus)Status;
4905 Status = opOK;
4906 auto AComparedToC = a.compareAbsoluteValue(c);
4907 z = cc;
4908 Status |= z.add(aa, RM);
4909 if (AComparedToC == APFloat::cmpGreaterThan) {
4910 // z = cc + aa + c + a;
4911 Status |= z.add(c, RM);
4912 Status |= z.add(a, RM);
4913 } else {
4914 // z = cc + aa + a + c;
4915 Status |= z.add(a, RM);
4916 Status |= z.add(c, RM);
4918 if (!z.isFinite()) {
4919 Floats[0] = std::move(z);
4920 Floats[1].makeZero(/* Neg = */ false);
4921 return (opStatus)Status;
4923 Floats[0] = z;
4924 APFloat zz = aa;
4925 Status |= zz.add(cc, RM);
4926 if (AComparedToC == APFloat::cmpGreaterThan) {
4927 // Floats[1] = a - z + c + zz;
4928 Floats[1] = a;
4929 Status |= Floats[1].subtract(z, RM);
4930 Status |= Floats[1].add(c, RM);
4931 Status |= Floats[1].add(zz, RM);
4932 } else {
4933 // Floats[1] = c - z + a + zz;
4934 Floats[1] = c;
4935 Status |= Floats[1].subtract(z, RM);
4936 Status |= Floats[1].add(a, RM);
4937 Status |= Floats[1].add(zz, RM);
4939 } else {
4940 // q = a - z;
4941 APFloat q = a;
4942 Status |= q.subtract(z, RM);
4944 // zz = q + c + (a - (q + z)) + aa + cc;
4945 // Compute a - (q + z) as -((q + z) - a) to avoid temporary copies.
4946 auto zz = q;
4947 Status |= zz.add(c, RM);
4948 Status |= q.add(z, RM);
4949 Status |= q.subtract(a, RM);
4950 q.changeSign();
4951 Status |= zz.add(q, RM);
4952 Status |= zz.add(aa, RM);
4953 Status |= zz.add(cc, RM);
4954 if (zz.isZero() && !zz.isNegative()) {
4955 Floats[0] = std::move(z);
4956 Floats[1].makeZero(/* Neg = */ false);
4957 return opOK;
4959 Floats[0] = z;
4960 Status |= Floats[0].add(zz, RM);
4961 if (!Floats[0].isFinite()) {
4962 Floats[1].makeZero(/* Neg = */ false);
4963 return (opStatus)Status;
4965 Floats[1] = std::move(z);
4966 Status |= Floats[1].subtract(Floats[0], RM);
4967 Status |= Floats[1].add(zz, RM);
4969 return (opStatus)Status;
4972 APFloat::opStatus DoubleAPFloat::addWithSpecial(const DoubleAPFloat &LHS,
4973 const DoubleAPFloat &RHS,
4974 DoubleAPFloat &Out,
4975 roundingMode RM) {
4976 if (LHS.getCategory() == fcNaN) {
4977 Out = LHS;
4978 return opOK;
4980 if (RHS.getCategory() == fcNaN) {
4981 Out = RHS;
4982 return opOK;
4984 if (LHS.getCategory() == fcZero) {
4985 Out = RHS;
4986 return opOK;
4988 if (RHS.getCategory() == fcZero) {
4989 Out = LHS;
4990 return opOK;
4992 if (LHS.getCategory() == fcInfinity && RHS.getCategory() == fcInfinity &&
4993 LHS.isNegative() != RHS.isNegative()) {
4994 Out.makeNaN(false, Out.isNegative(), nullptr);
4995 return opInvalidOp;
4997 if (LHS.getCategory() == fcInfinity) {
4998 Out = LHS;
4999 return opOK;
5001 if (RHS.getCategory() == fcInfinity) {
5002 Out = RHS;
5003 return opOK;
5005 assert(LHS.getCategory() == fcNormal && RHS.getCategory() == fcNormal);
5007 APFloat A(LHS.Floats[0]), AA(LHS.Floats[1]), C(RHS.Floats[0]),
5008 CC(RHS.Floats[1]);
5009 assert(&A.getSemantics() == &semIEEEdouble);
5010 assert(&AA.getSemantics() == &semIEEEdouble);
5011 assert(&C.getSemantics() == &semIEEEdouble);
5012 assert(&CC.getSemantics() == &semIEEEdouble);
5013 assert(&Out.Floats[0].getSemantics() == &semIEEEdouble);
5014 assert(&Out.Floats[1].getSemantics() == &semIEEEdouble);
5015 return Out.addImpl(A, AA, C, CC, RM);
5018 APFloat::opStatus DoubleAPFloat::add(const DoubleAPFloat &RHS,
5019 roundingMode RM) {
5020 return addWithSpecial(*this, RHS, *this, RM);
5023 APFloat::opStatus DoubleAPFloat::subtract(const DoubleAPFloat &RHS,
5024 roundingMode RM) {
5025 changeSign();
5026 auto Ret = add(RHS, RM);
5027 changeSign();
5028 return Ret;
5031 APFloat::opStatus DoubleAPFloat::multiply(const DoubleAPFloat &RHS,
5032 APFloat::roundingMode RM) {
5033 const auto &LHS = *this;
5034 auto &Out = *this;
5035 /* Interesting observation: For special categories, finding the lowest
5036 common ancestor of the following layered graph gives the correct
5037 return category:
5041 Zero Inf
5043 Normal
5045 e.g. NaN * NaN = NaN
5046 Zero * Inf = NaN
5047 Normal * Zero = Zero
5048 Normal * Inf = Inf
5050 if (LHS.getCategory() == fcNaN) {
5051 Out = LHS;
5052 return opOK;
5054 if (RHS.getCategory() == fcNaN) {
5055 Out = RHS;
5056 return opOK;
5058 if ((LHS.getCategory() == fcZero && RHS.getCategory() == fcInfinity) ||
5059 (LHS.getCategory() == fcInfinity && RHS.getCategory() == fcZero)) {
5060 Out.makeNaN(false, false, nullptr);
5061 return opOK;
5063 if (LHS.getCategory() == fcZero || LHS.getCategory() == fcInfinity) {
5064 Out = LHS;
5065 return opOK;
5067 if (RHS.getCategory() == fcZero || RHS.getCategory() == fcInfinity) {
5068 Out = RHS;
5069 return opOK;
5071 assert(LHS.getCategory() == fcNormal && RHS.getCategory() == fcNormal &&
5072 "Special cases not handled exhaustively");
5074 int Status = opOK;
5075 APFloat A = Floats[0], B = Floats[1], C = RHS.Floats[0], D = RHS.Floats[1];
5076 // t = a * c
5077 APFloat T = A;
5078 Status |= T.multiply(C, RM);
5079 if (!T.isFiniteNonZero()) {
5080 Floats[0] = T;
5081 Floats[1].makeZero(/* Neg = */ false);
5082 return (opStatus)Status;
5085 // tau = fmsub(a, c, t), that is -fmadd(-a, c, t).
5086 APFloat Tau = A;
5087 T.changeSign();
5088 Status |= Tau.fusedMultiplyAdd(C, T, RM);
5089 T.changeSign();
5091 // v = a * d
5092 APFloat V = A;
5093 Status |= V.multiply(D, RM);
5094 // w = b * c
5095 APFloat W = B;
5096 Status |= W.multiply(C, RM);
5097 Status |= V.add(W, RM);
5098 // tau += v + w
5099 Status |= Tau.add(V, RM);
5101 // u = t + tau
5102 APFloat U = T;
5103 Status |= U.add(Tau, RM);
5105 Floats[0] = U;
5106 if (!U.isFinite()) {
5107 Floats[1].makeZero(/* Neg = */ false);
5108 } else {
5109 // Floats[1] = (t - u) + tau
5110 Status |= T.subtract(U, RM);
5111 Status |= T.add(Tau, RM);
5112 Floats[1] = T;
5114 return (opStatus)Status;
5117 APFloat::opStatus DoubleAPFloat::divide(const DoubleAPFloat &RHS,
5118 APFloat::roundingMode RM) {
5119 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5120 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5121 auto Ret =
5122 Tmp.divide(APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt()), RM);
5123 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5124 return Ret;
5127 APFloat::opStatus DoubleAPFloat::remainder(const DoubleAPFloat &RHS) {
5128 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5129 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5130 auto Ret =
5131 Tmp.remainder(APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt()));
5132 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5133 return Ret;
5136 APFloat::opStatus DoubleAPFloat::mod(const DoubleAPFloat &RHS) {
5137 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5138 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5139 auto Ret = Tmp.mod(APFloat(semPPCDoubleDoubleLegacy, RHS.bitcastToAPInt()));
5140 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5141 return Ret;
5144 APFloat::opStatus
5145 DoubleAPFloat::fusedMultiplyAdd(const DoubleAPFloat &Multiplicand,
5146 const DoubleAPFloat &Addend,
5147 APFloat::roundingMode RM) {
5148 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5149 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5150 auto Ret = Tmp.fusedMultiplyAdd(
5151 APFloat(semPPCDoubleDoubleLegacy, Multiplicand.bitcastToAPInt()),
5152 APFloat(semPPCDoubleDoubleLegacy, Addend.bitcastToAPInt()), RM);
5153 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5154 return Ret;
5157 APFloat::opStatus DoubleAPFloat::roundToIntegral(APFloat::roundingMode RM) {
5158 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5159 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5160 auto Ret = Tmp.roundToIntegral(RM);
5161 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5162 return Ret;
5165 void DoubleAPFloat::changeSign() {
5166 Floats[0].changeSign();
5167 Floats[1].changeSign();
5170 APFloat::cmpResult
5171 DoubleAPFloat::compareAbsoluteValue(const DoubleAPFloat &RHS) const {
5172 auto Result = Floats[0].compareAbsoluteValue(RHS.Floats[0]);
5173 if (Result != cmpEqual)
5174 return Result;
5175 Result = Floats[1].compareAbsoluteValue(RHS.Floats[1]);
5176 if (Result == cmpLessThan || Result == cmpGreaterThan) {
5177 auto Against = Floats[0].isNegative() ^ Floats[1].isNegative();
5178 auto RHSAgainst = RHS.Floats[0].isNegative() ^ RHS.Floats[1].isNegative();
5179 if (Against && !RHSAgainst)
5180 return cmpLessThan;
5181 if (!Against && RHSAgainst)
5182 return cmpGreaterThan;
5183 if (!Against && !RHSAgainst)
5184 return Result;
5185 if (Against && RHSAgainst)
5186 return (cmpResult)(cmpLessThan + cmpGreaterThan - Result);
5188 return Result;
5191 APFloat::fltCategory DoubleAPFloat::getCategory() const {
5192 return Floats[0].getCategory();
5195 bool DoubleAPFloat::isNegative() const { return Floats[0].isNegative(); }
5197 void DoubleAPFloat::makeInf(bool Neg) {
5198 Floats[0].makeInf(Neg);
5199 Floats[1].makeZero(/* Neg = */ false);
5202 void DoubleAPFloat::makeZero(bool Neg) {
5203 Floats[0].makeZero(Neg);
5204 Floats[1].makeZero(/* Neg = */ false);
5207 void DoubleAPFloat::makeLargest(bool Neg) {
5208 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5209 Floats[0] = APFloat(semIEEEdouble, APInt(64, 0x7fefffffffffffffull));
5210 Floats[1] = APFloat(semIEEEdouble, APInt(64, 0x7c8ffffffffffffeull));
5211 if (Neg)
5212 changeSign();
5215 void DoubleAPFloat::makeSmallest(bool Neg) {
5216 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5217 Floats[0].makeSmallest(Neg);
5218 Floats[1].makeZero(/* Neg = */ false);
5221 void DoubleAPFloat::makeSmallestNormalized(bool Neg) {
5222 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5223 Floats[0] = APFloat(semIEEEdouble, APInt(64, 0x0360000000000000ull));
5224 if (Neg)
5225 Floats[0].changeSign();
5226 Floats[1].makeZero(/* Neg = */ false);
5229 void DoubleAPFloat::makeNaN(bool SNaN, bool Neg, const APInt *fill) {
5230 Floats[0].makeNaN(SNaN, Neg, fill);
5231 Floats[1].makeZero(/* Neg = */ false);
5234 APFloat::cmpResult DoubleAPFloat::compare(const DoubleAPFloat &RHS) const {
5235 auto Result = Floats[0].compare(RHS.Floats[0]);
5236 // |Float[0]| > |Float[1]|
5237 if (Result == APFloat::cmpEqual)
5238 return Floats[1].compare(RHS.Floats[1]);
5239 return Result;
5242 bool DoubleAPFloat::bitwiseIsEqual(const DoubleAPFloat &RHS) const {
5243 return Floats[0].bitwiseIsEqual(RHS.Floats[0]) &&
5244 Floats[1].bitwiseIsEqual(RHS.Floats[1]);
5247 hash_code hash_value(const DoubleAPFloat &Arg) {
5248 if (Arg.Floats)
5249 return hash_combine(hash_value(Arg.Floats[0]), hash_value(Arg.Floats[1]));
5250 return hash_combine(Arg.Semantics);
5253 APInt DoubleAPFloat::bitcastToAPInt() const {
5254 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5255 uint64_t Data[] = {
5256 Floats[0].bitcastToAPInt().getRawData()[0],
5257 Floats[1].bitcastToAPInt().getRawData()[0],
5259 return APInt(128, 2, Data);
5262 Expected<APFloat::opStatus> DoubleAPFloat::convertFromString(StringRef S,
5263 roundingMode RM) {
5264 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5265 APFloat Tmp(semPPCDoubleDoubleLegacy);
5266 auto Ret = Tmp.convertFromString(S, RM);
5267 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5268 return Ret;
5271 APFloat::opStatus DoubleAPFloat::next(bool nextDown) {
5272 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5273 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5274 auto Ret = Tmp.next(nextDown);
5275 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5276 return Ret;
5279 APFloat::opStatus
5280 DoubleAPFloat::convertToInteger(MutableArrayRef<integerPart> Input,
5281 unsigned int Width, bool IsSigned,
5282 roundingMode RM, bool *IsExact) const {
5283 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5284 return APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt())
5285 .convertToInteger(Input, Width, IsSigned, RM, IsExact);
5288 APFloat::opStatus DoubleAPFloat::convertFromAPInt(const APInt &Input,
5289 bool IsSigned,
5290 roundingMode RM) {
5291 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5292 APFloat Tmp(semPPCDoubleDoubleLegacy);
5293 auto Ret = Tmp.convertFromAPInt(Input, IsSigned, RM);
5294 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5295 return Ret;
5298 APFloat::opStatus
5299 DoubleAPFloat::convertFromSignExtendedInteger(const integerPart *Input,
5300 unsigned int InputSize,
5301 bool IsSigned, roundingMode RM) {
5302 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5303 APFloat Tmp(semPPCDoubleDoubleLegacy);
5304 auto Ret = Tmp.convertFromSignExtendedInteger(Input, InputSize, IsSigned, RM);
5305 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5306 return Ret;
5309 APFloat::opStatus
5310 DoubleAPFloat::convertFromZeroExtendedInteger(const integerPart *Input,
5311 unsigned int InputSize,
5312 bool IsSigned, roundingMode RM) {
5313 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5314 APFloat Tmp(semPPCDoubleDoubleLegacy);
5315 auto Ret = Tmp.convertFromZeroExtendedInteger(Input, InputSize, IsSigned, RM);
5316 *this = DoubleAPFloat(semPPCDoubleDouble, Tmp.bitcastToAPInt());
5317 return Ret;
5320 unsigned int DoubleAPFloat::convertToHexString(char *DST,
5321 unsigned int HexDigits,
5322 bool UpperCase,
5323 roundingMode RM) const {
5324 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5325 return APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt())
5326 .convertToHexString(DST, HexDigits, UpperCase, RM);
5329 bool DoubleAPFloat::isDenormal() const {
5330 return getCategory() == fcNormal &&
5331 (Floats[0].isDenormal() || Floats[1].isDenormal() ||
5332 // (double)(Hi + Lo) == Hi defines a normal number.
5333 Floats[0] != Floats[0] + Floats[1]);
5336 bool DoubleAPFloat::isSmallest() const {
5337 if (getCategory() != fcNormal)
5338 return false;
5339 DoubleAPFloat Tmp(*this);
5340 Tmp.makeSmallest(this->isNegative());
5341 return Tmp.compare(*this) == cmpEqual;
5344 bool DoubleAPFloat::isSmallestNormalized() const {
5345 if (getCategory() != fcNormal)
5346 return false;
5348 DoubleAPFloat Tmp(*this);
5349 Tmp.makeSmallestNormalized(this->isNegative());
5350 return Tmp.compare(*this) == cmpEqual;
5353 bool DoubleAPFloat::isLargest() const {
5354 if (getCategory() != fcNormal)
5355 return false;
5356 DoubleAPFloat Tmp(*this);
5357 Tmp.makeLargest(this->isNegative());
5358 return Tmp.compare(*this) == cmpEqual;
5361 bool DoubleAPFloat::isInteger() const {
5362 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5363 return Floats[0].isInteger() && Floats[1].isInteger();
5366 void DoubleAPFloat::toString(SmallVectorImpl<char> &Str,
5367 unsigned FormatPrecision,
5368 unsigned FormatMaxPadding,
5369 bool TruncateZero) const {
5370 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5371 APFloat(semPPCDoubleDoubleLegacy, bitcastToAPInt())
5372 .toString(Str, FormatPrecision, FormatMaxPadding, TruncateZero);
5375 bool DoubleAPFloat::getExactInverse(APFloat *inv) const {
5376 assert(Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5377 APFloat Tmp(semPPCDoubleDoubleLegacy, bitcastToAPInt());
5378 if (!inv)
5379 return Tmp.getExactInverse(nullptr);
5380 APFloat Inv(semPPCDoubleDoubleLegacy);
5381 auto Ret = Tmp.getExactInverse(&Inv);
5382 *inv = APFloat(semPPCDoubleDouble, Inv.bitcastToAPInt());
5383 return Ret;
5386 int DoubleAPFloat::getExactLog2() const {
5387 // TODO: Implement me
5388 return INT_MIN;
5391 int DoubleAPFloat::getExactLog2Abs() const {
5392 // TODO: Implement me
5393 return INT_MIN;
5396 DoubleAPFloat scalbn(const DoubleAPFloat &Arg, int Exp,
5397 APFloat::roundingMode RM) {
5398 assert(Arg.Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5399 return DoubleAPFloat(semPPCDoubleDouble, scalbn(Arg.Floats[0], Exp, RM),
5400 scalbn(Arg.Floats[1], Exp, RM));
5403 DoubleAPFloat frexp(const DoubleAPFloat &Arg, int &Exp,
5404 APFloat::roundingMode RM) {
5405 assert(Arg.Semantics == &semPPCDoubleDouble && "Unexpected Semantics");
5406 APFloat First = frexp(Arg.Floats[0], Exp, RM);
5407 APFloat Second = Arg.Floats[1];
5408 if (Arg.getCategory() == APFloat::fcNormal)
5409 Second = scalbn(Second, -Exp, RM);
5410 return DoubleAPFloat(semPPCDoubleDouble, std::move(First), std::move(Second));
5413 } // namespace detail
5415 APFloat::Storage::Storage(IEEEFloat F, const fltSemantics &Semantics) {
5416 if (usesLayout<IEEEFloat>(Semantics)) {
5417 new (&IEEE) IEEEFloat(std::move(F));
5418 return;
5420 if (usesLayout<DoubleAPFloat>(Semantics)) {
5421 const fltSemantics& S = F.getSemantics();
5422 new (&Double)
5423 DoubleAPFloat(Semantics, APFloat(std::move(F), S),
5424 APFloat(semIEEEdouble));
5425 return;
5427 llvm_unreachable("Unexpected semantics");
5430 Expected<APFloat::opStatus> APFloat::convertFromString(StringRef Str,
5431 roundingMode RM) {
5432 APFLOAT_DISPATCH_ON_SEMANTICS(convertFromString(Str, RM));
5435 hash_code hash_value(const APFloat &Arg) {
5436 if (APFloat::usesLayout<detail::IEEEFloat>(Arg.getSemantics()))
5437 return hash_value(Arg.U.IEEE);
5438 if (APFloat::usesLayout<detail::DoubleAPFloat>(Arg.getSemantics()))
5439 return hash_value(Arg.U.Double);
5440 llvm_unreachable("Unexpected semantics");
5443 APFloat::APFloat(const fltSemantics &Semantics, StringRef S)
5444 : APFloat(Semantics) {
5445 auto StatusOrErr = convertFromString(S, rmNearestTiesToEven);
5446 assert(StatusOrErr && "Invalid floating point representation");
5447 consumeError(StatusOrErr.takeError());
5450 FPClassTest APFloat::classify() const {
5451 if (isZero())
5452 return isNegative() ? fcNegZero : fcPosZero;
5453 if (isNormal())
5454 return isNegative() ? fcNegNormal : fcPosNormal;
5455 if (isDenormal())
5456 return isNegative() ? fcNegSubnormal : fcPosSubnormal;
5457 if (isInfinity())
5458 return isNegative() ? fcNegInf : fcPosInf;
5459 assert(isNaN() && "Other class of FP constant");
5460 return isSignaling() ? fcSNan : fcQNan;
5463 APFloat::opStatus APFloat::convert(const fltSemantics &ToSemantics,
5464 roundingMode RM, bool *losesInfo) {
5465 if (&getSemantics() == &ToSemantics) {
5466 *losesInfo = false;
5467 return opOK;
5469 if (usesLayout<IEEEFloat>(getSemantics()) &&
5470 usesLayout<IEEEFloat>(ToSemantics))
5471 return U.IEEE.convert(ToSemantics, RM, losesInfo);
5472 if (usesLayout<IEEEFloat>(getSemantics()) &&
5473 usesLayout<DoubleAPFloat>(ToSemantics)) {
5474 assert(&ToSemantics == &semPPCDoubleDouble);
5475 auto Ret = U.IEEE.convert(semPPCDoubleDoubleLegacy, RM, losesInfo);
5476 *this = APFloat(ToSemantics, U.IEEE.bitcastToAPInt());
5477 return Ret;
5479 if (usesLayout<DoubleAPFloat>(getSemantics()) &&
5480 usesLayout<IEEEFloat>(ToSemantics)) {
5481 auto Ret = getIEEE().convert(ToSemantics, RM, losesInfo);
5482 *this = APFloat(std::move(getIEEE()), ToSemantics);
5483 return Ret;
5485 llvm_unreachable("Unexpected semantics");
5488 APFloat APFloat::getAllOnesValue(const fltSemantics &Semantics) {
5489 return APFloat(Semantics, APInt::getAllOnes(Semantics.sizeInBits));
5492 void APFloat::print(raw_ostream &OS) const {
5493 SmallVector<char, 16> Buffer;
5494 toString(Buffer);
5495 OS << Buffer;
5498 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5499 LLVM_DUMP_METHOD void APFloat::dump() const {
5500 print(dbgs());
5501 dbgs() << '\n';
5503 #endif
5505 void APFloat::Profile(FoldingSetNodeID &NID) const {
5506 NID.Add(bitcastToAPInt());
5509 /* Same as convertToInteger(integerPart*, ...), except the result is returned in
5510 an APSInt, whose initial bit-width and signed-ness are used to determine the
5511 precision of the conversion.
5513 APFloat::opStatus APFloat::convertToInteger(APSInt &result,
5514 roundingMode rounding_mode,
5515 bool *isExact) const {
5516 unsigned bitWidth = result.getBitWidth();
5517 SmallVector<uint64_t, 4> parts(result.getNumWords());
5518 opStatus status = convertToInteger(parts, bitWidth, result.isSigned(),
5519 rounding_mode, isExact);
5520 // Keeps the original signed-ness.
5521 result = APInt(bitWidth, parts);
5522 return status;
5525 double APFloat::convertToDouble() const {
5526 if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEdouble)
5527 return getIEEE().convertToDouble();
5528 assert(isRepresentableBy(getSemantics(), semIEEEdouble) &&
5529 "Float semantics is not representable by IEEEdouble");
5530 APFloat Temp = *this;
5531 bool LosesInfo;
5532 opStatus St = Temp.convert(semIEEEdouble, rmNearestTiesToEven, &LosesInfo);
5533 assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
5534 (void)St;
5535 return Temp.getIEEE().convertToDouble();
5538 #ifdef HAS_IEE754_FLOAT128
5539 float128 APFloat::convertToQuad() const {
5540 if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEquad)
5541 return getIEEE().convertToQuad();
5542 assert(isRepresentableBy(getSemantics(), semIEEEquad) &&
5543 "Float semantics is not representable by IEEEquad");
5544 APFloat Temp = *this;
5545 bool LosesInfo;
5546 opStatus St = Temp.convert(semIEEEquad, rmNearestTiesToEven, &LosesInfo);
5547 assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
5548 (void)St;
5549 return Temp.getIEEE().convertToQuad();
5551 #endif
5553 float APFloat::convertToFloat() const {
5554 if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEsingle)
5555 return getIEEE().convertToFloat();
5556 assert(isRepresentableBy(getSemantics(), semIEEEsingle) &&
5557 "Float semantics is not representable by IEEEsingle");
5558 APFloat Temp = *this;
5559 bool LosesInfo;
5560 opStatus St = Temp.convert(semIEEEsingle, rmNearestTiesToEven, &LosesInfo);
5561 assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
5562 (void)St;
5563 return Temp.getIEEE().convertToFloat();
5566 } // namespace llvm
5568 #undef APFLOAT_DISPATCH_ON_SEMANTICS