/src/keystone/llvm/lib/Support/APFloat.cpp
Line | Count | Source |
1 | | //===-- APFloat.cpp - Implement APFloat class -----------------------------===// |
2 | | // |
3 | | // The LLVM Compiler Infrastructure |
4 | | // |
5 | | // This file is distributed under the University of Illinois Open Source |
6 | | // License. See LICENSE.TXT for details. |
7 | | // |
8 | | //===----------------------------------------------------------------------===// |
9 | | // |
10 | | // This file implements a class to represent arbitrary precision floating |
11 | | // point values and provide a variety of arithmetic operations on them. |
12 | | // |
13 | | //===----------------------------------------------------------------------===// |
14 | | |
15 | | #include "llvm/ADT/APFloat.h" |
16 | | #include "llvm/ADT/APSInt.h" |
17 | | #include "llvm/ADT/FoldingSet.h" |
18 | | #include "llvm/ADT/Hashing.h" |
19 | | #include "llvm/ADT/StringExtras.h" |
20 | | #include "llvm/ADT/StringRef.h" |
21 | | #include "llvm/Support/ErrorHandling.h" |
22 | | #include "llvm/Support/MathExtras.h" |
23 | | #include <cstring> |
24 | | #include <limits.h> |
25 | | |
26 | | using namespace llvm_ks; |
27 | | |
28 | | /// A macro used to combine two fcCategory enums into one key which can be used |
29 | | /// in a switch statement to classify how the interaction of two APFloat's |
30 | | /// categories affects an operation. |
31 | | /// |
32 | | /// TODO: If clang source code is ever allowed to use constexpr in its own |
33 | | /// codebase, change this into a static inline function. |
34 | 0 | #define PackCategoriesIntoKey(_lhs, _rhs) ((_lhs) * 4 + (_rhs)) |
35 | | |
36 | | /* Assumed in hexadecimal significand parsing, and conversion to |
37 | | hexadecimal strings. */ |
38 | | static_assert(integerPartWidth % 4 == 0, "Part width must be divisible by 4!"); |
39 | | |
40 | | namespace llvm_ks { |
41 | | |
42 | | /* Represents floating point arithmetic semantics. */ |
43 | | struct fltSemantics { |
44 | | /* The largest E such that 2^E is representable; this matches the |
45 | | definition of IEEE 754. */ |
46 | | APFloat::ExponentType maxExponent; |
47 | | |
48 | | /* The smallest E such that 2^E is a normalized number; this |
49 | | matches the definition of IEEE 754. */ |
50 | | APFloat::ExponentType minExponent; |
51 | | |
52 | | /* Number of bits in the significand. This includes the integer |
53 | | bit. */ |
54 | | unsigned int precision; |
55 | | |
56 | | /* Number of bits actually used in the semantics. */ |
57 | | unsigned int sizeInBits; |
58 | | }; |
59 | | |
60 | | const fltSemantics APFloat::IEEEhalf = { 15, -14, 11, 16 }; |
61 | | const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, 32 }; |
62 | | const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, 64 }; |
63 | | const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, 128 }; |
64 | | const fltSemantics APFloat::x87DoubleExtended = { 16383, -16382, 64, 80 }; |
65 | | const fltSemantics APFloat::Bogus = { 0, 0, 0, 0 }; |
66 | | |
67 | | /* The PowerPC format consists of two doubles. It does not map cleanly |
68 | | onto the usual format above. It is approximated using twice the |
69 | | mantissa bits. Note that for exponents near the double minimum, |
70 | | we no longer can represent the full 106 mantissa bits, so those |
71 | | will be treated as denormal numbers. |
72 | | |
73 | | FIXME: While this approximation is equivalent to what GCC uses for |
74 | | compile-time arithmetic on PPC double-double numbers, it is not able |
75 | | to represent all possible values held by a PPC double-double number, |
76 | | for example: (long double) 1.0 + (long double) 0x1p-106 |
77 | | Should this be replaced by a full emulation of PPC double-double? */ |
78 | | const fltSemantics APFloat::PPCDoubleDouble = { 1023, -1022 + 53, 53 + 53, 128 }; |
79 | | |
80 | | /* A tight upper bound on number of parts required to hold the value |
81 | | pow(5, power) is |
82 | | |
83 | | power * 815 / (351 * integerPartWidth) + 1 |
84 | | |
85 | | However, whilst the result may require only this many parts, |
86 | | because we are multiplying two values to get it, the |
87 | | multiplication may require an extra part with the excess part |
88 | | being zero (consider the trivial case of 1 * 1, tcFullMultiply |
89 | | requires two parts to hold the single-part result). So we add an |
90 | | extra one to guarantee enough space whilst multiplying. */ |
91 | | const unsigned int maxExponent = 16383; |
92 | | const unsigned int maxPrecision = 113; |
93 | | const unsigned int maxPowerOfFiveExponent = maxExponent + maxPrecision - 1; |
94 | | const unsigned int maxPowerOfFiveParts = 2 + ((maxPowerOfFiveExponent * 815) |
95 | | / (351 * integerPartWidth)); |
96 | | } |
97 | | |
98 | | /* A bunch of private, handy routines. */ |
99 | | |
100 | | static inline unsigned int |
101 | | partCountForBits(unsigned int bits) |
102 | 32.7M | { |
103 | 32.7M | return ((bits) + integerPartWidth - 1) / integerPartWidth; |
104 | 32.7M | } |
105 | | |
106 | | /* Returns 0U-9U. Return values >= 10U are not digits. */ |
107 | | static inline unsigned int |
108 | | decDigitValue(unsigned int c) |
109 | 12.2M | { |
110 | 12.2M | return c - '0'; |
111 | 12.2M | } |
112 | | |
113 | | /* Return the value of a decimal exponent of the form |
114 | | [+-]ddddddd. |
115 | | |
116 | | If the exponent overflows, returns a large exponent with the |
117 | | appropriate sign. */ |
118 | | static int |
119 | | readExponent(StringRef::iterator begin, StringRef::iterator end, APFloat::opStatus &fp) |
120 | 114k | { |
121 | 114k | bool isNegative; |
122 | 114k | unsigned int absExponent; |
123 | 114k | const unsigned int overlargeExponent = 24000; /* FIXME. */ |
124 | 114k | StringRef::iterator p = begin; |
125 | | |
126 | 114k | fp = APFloat::opOK; |
127 | | |
128 | | //assert(p != end && "Exponent has no digits"); // qq |
129 | 114k | if (p == end) { |
130 | 7.27k | fp = APFloat::opInvalidOp; |
131 | 7.27k | return 0; |
132 | 7.27k | } |
133 | | |
134 | 106k | isNegative = (*p == '-'); |
135 | 106k | if (*p == '-' || *p == '+') { |
136 | 38.8k | p++; |
137 | | //assert(p != end && "Exponent has no digits"); |
138 | 38.8k | if (p == end) { |
139 | 9.34k | fp = APFloat::opInvalidOp; |
140 | 9.34k | return 0; |
141 | 9.34k | } |
142 | 38.8k | } |
143 | | |
144 | 97.6k | absExponent = decDigitValue(*p++); |
145 | | //assert(absExponent < 10U && "Invalid character in exponent"); |
146 | 97.6k | if (absExponent >= 10U) { |
147 | 18 | fp = APFloat::opInvalidOp; |
148 | 18 | return 0; |
149 | 18 | } |
150 | | |
151 | 289k | for (; p != end; ++p) { |
152 | 202k | unsigned int value; |
153 | | |
154 | 202k | value = decDigitValue(*p); |
155 | | //assert(value < 10U && "Invalid character in exponent"); |
156 | 202k | if (value >= 10U) { |
157 | 8 | fp = APFloat::opInvalidOp; |
158 | 8 | return 0; |
159 | 8 | } |
160 | | |
161 | 202k | value += absExponent * 10; |
162 | 202k | if (absExponent >= overlargeExponent) { |
163 | 10.9k | absExponent = overlargeExponent; |
164 | 10.9k | p = end; /* outwit assert below */ |
165 | 10.9k | break; |
166 | 10.9k | } |
167 | 191k | absExponent = value; |
168 | 191k | } |
169 | | |
170 | | //assert(p == end && "Invalid exponent in exponent"); |
171 | 97.6k | if (p != end) { |
172 | 0 | fp = APFloat::opInvalidOp; |
173 | 0 | return 0; |
174 | 0 | } |
175 | | |
176 | 97.6k | if (isNegative) |
177 | 28.1k | return -(int) absExponent; |
178 | 69.4k | else |
179 | 69.4k | return (int) absExponent; |
180 | 97.6k | } |
181 | | |
182 | | /* This is ugly and needs cleaning up, but I don't immediately see |
183 | | how whilst remaining safe. */ |
184 | | static int |
185 | | totalExponent(StringRef::iterator p, StringRef::iterator end, |
186 | | int exponentAdjustment) |
187 | 127k | { |
188 | 127k | int unsignedExponent; |
189 | 127k | bool negative, overflow; |
190 | 127k | int exponent = 0; |
191 | | |
192 | 127k | assert(p != end && "Exponent has no digits"); |
193 | | |
194 | 127k | negative = *p == '-'; |
195 | 127k | if (*p == '-' || *p == '+') { |
196 | 36.1k | p++; |
197 | 36.1k | assert(p != end && "Exponent has no digits"); |
198 | 36.1k | } |
199 | | |
200 | 127k | unsignedExponent = 0; |
201 | 127k | overflow = false; |
202 | 432k | for (; p != end; ++p) { |
203 | 317k | unsigned int value; |
204 | | |
205 | 317k | value = decDigitValue(*p); |
206 | 317k | assert(value < 10U && "Invalid character in exponent"); |
207 | | |
208 | 317k | unsignedExponent = unsignedExponent * 10 + value; |
209 | 317k | if (unsignedExponent > 32767) { |
210 | 13.0k | overflow = true; |
211 | 13.0k | break; |
212 | 13.0k | } |
213 | 317k | } |
214 | | |
215 | 127k | if (exponentAdjustment > 32767 || exponentAdjustment < -32768) |
216 | 0 | overflow = true; |
217 | | |
218 | 127k | if (!overflow) { |
219 | 114k | exponent = unsignedExponent; |
220 | 114k | if (negative) |
221 | 17.5k | exponent = -exponent; |
222 | 114k | exponent += exponentAdjustment; |
223 | 114k | if (exponent > 32767 || exponent < -32768) |
224 | 7.68k | overflow = true; |
225 | 114k | } |
226 | | |
227 | 127k | if (overflow) |
228 | 20.7k | exponent = negative ? -32768: 32767; |
229 | | |
230 | 127k | return exponent; |
231 | 127k | } |
232 | | |
233 | | static StringRef::iterator |
234 | | skipLeadingZeroesAndAnyDot(StringRef::iterator begin, StringRef::iterator end, |
235 | | StringRef::iterator *dot) |
236 | 960k | { |
237 | 960k | StringRef::iterator p = begin; |
238 | 960k | *dot = end; |
239 | 1.03M | while (p != end && *p == '0') |
240 | 78.2k | p++; |
241 | | |
242 | 960k | if (p != end && *p == '.') { |
243 | 831k | *dot = p++; |
244 | | |
245 | 831k | assert(end - begin != 1 && "Significand has no digits"); |
246 | | |
247 | 1.82M | while (p != end && *p == '0') |
248 | 992k | p++; |
249 | 831k | } |
250 | | |
251 | 960k | return p; |
252 | 960k | } |
253 | | |
254 | | /* Given a normal decimal floating point number of the form |
255 | | |
256 | | dddd.dddd[eE][+-]ddd |
257 | | |
258 | | where the decimal point and exponent are optional, fill out the |
259 | | structure D. Exponent is appropriate if the significand is |
260 | | treated as an integer, and normalizedExponent if the significand |
261 | | is taken to have the decimal point after a single leading |
262 | | non-zero digit. |
263 | | |
264 | | If the value is zero, V->firstSigDigit points to a non-digit, and |
265 | | the return exponent is zero. |
266 | | */ |
267 | | struct decimalInfo { |
268 | | const char *firstSigDigit; |
269 | | const char *lastSigDigit; |
270 | | int exponent; |
271 | | int normalizedExponent; |
272 | | }; |
273 | | |
274 | | APFloat::opStatus |
275 | | interpretDecimal(StringRef::iterator begin, StringRef::iterator end, |
276 | | decimalInfo *D) |
277 | 828k | { |
278 | 828k | StringRef::iterator dot = end; |
279 | 828k | StringRef::iterator p = skipLeadingZeroesAndAnyDot (begin, end, &dot); |
280 | 828k | APFloat::opStatus fp; |
281 | | |
282 | 828k | D->firstSigDigit = p; |
283 | 828k | D->exponent = 0; |
284 | 828k | D->normalizedExponent = 0; |
285 | | |
286 | 5.92M | for (; p != end; ++p) { |
287 | 5.21M | if (*p == '.') { |
288 | | //assert(dot == end && "String contains multiple dots"); |
289 | 0 | if (dot != end) |
290 | 0 | return APFloat::opInvalidOp; |
291 | 0 | dot = p++; |
292 | 0 | if (p == end) |
293 | 0 | break; |
294 | 0 | } |
295 | 5.21M | if (decDigitValue(*p) >= 10U) |
296 | 114k | break; |
297 | 5.21M | } |
298 | | |
299 | 828k | if (p != end) { |
300 | | //assert((*p == 'e' || *p == 'E') && "Invalid character in significand"); |
301 | 114k | if (*p != 'e' && *p != 'E') |
302 | 18 | return APFloat::opInvalidOp; |
303 | | //assert(p != begin && "Significand has no digits"); |
304 | 114k | if (p == begin) |
305 | 0 | return APFloat::opInvalidOp; |
306 | | //assert((dot == end || p - begin != 1) && "Significand has no digits"); |
307 | 114k | if (dot != end && p - begin == 1) |
308 | 0 | return APFloat::opInvalidOp; |
309 | | |
310 | | /* p points to the first non-digit in the string */ |
311 | 114k | D->exponent = readExponent(p + 1, end, fp); // qq |
312 | 114k | if (fp) |
313 | 16.6k | return fp; |
314 | | |
315 | | /* Implied decimal point? */ |
316 | 97.6k | if (dot == end) |
317 | 28 | dot = p; |
318 | 97.6k | } |
319 | | |
320 | | /* If number is all zeroes accept any exponent. */ |
321 | 811k | if (p != D->firstSigDigit) { |
322 | | /* Drop insignificant trailing zeroes. */ |
323 | 756k | if (p != begin) { |
324 | 756k | do |
325 | 756k | do |
326 | 867k | p--; |
327 | 867k | while (p != begin && *p == '0'); |
328 | 756k | while (p != begin && *p == '.'); |
329 | 756k | } |
330 | | |
331 | | /* Adjust the exponents for any decimal point. */ |
332 | 756k | D->exponent += static_cast<APFloat::ExponentType>((dot - p) - (dot > p)); |
333 | 756k | D->normalizedExponent = (D->exponent + |
334 | 756k | static_cast<APFloat::ExponentType>((p - D->firstSigDigit) |
335 | 756k | - (dot > D->firstSigDigit && dot < p))); |
336 | 756k | } |
337 | | |
338 | 811k | D->lastSigDigit = p; |
339 | | |
340 | 811k | return APFloat::opOK; |
341 | 828k | } |
342 | | |
343 | | /* Return the trailing fraction of a hexadecimal number. |
344 | | DIGITVALUE is the first hex digit of the fraction, P points to |
345 | | the next digit. */ |
346 | | static lostFraction |
347 | | trailingHexadecimalFraction(StringRef::iterator p, StringRef::iterator end, |
348 | | unsigned int digitValue) |
349 | 25.8k | { |
350 | 25.8k | unsigned int hexDigit; |
351 | | |
352 | | /* If the first trailing digit isn't 0 or 8 we can work out the |
353 | | fraction immediately. */ |
354 | 25.8k | if (digitValue > 8) |
355 | 7.35k | return lfMoreThanHalf; |
356 | 18.5k | else if (digitValue < 8 && digitValue > 0) |
357 | 4.75k | return lfLessThanHalf; |
358 | | |
359 | | // Otherwise we need to find the first non-zero digit. |
360 | 62.1k | while (p != end && (*p == '0' || *p == '.')) |
361 | 48.3k | p++; |
362 | | |
363 | 13.7k | assert(p != end && "Invalid trailing hexadecimal fraction!"); |
364 | | |
365 | 13.7k | hexDigit = hexDigitValue(*p); |
366 | | |
367 | | /* If we ran off the end it is exactly zero or one-half, otherwise |
368 | | a little more. */ |
369 | 13.7k | if (hexDigit == -1U) |
370 | 2.71k | return digitValue == 0 ? lfExactlyZero: lfExactlyHalf; |
371 | 11.0k | else |
372 | 11.0k | return digitValue == 0 ? lfLessThanHalf: lfMoreThanHalf; |
373 | 13.7k | } |
374 | | |
375 | | /* Return the fraction lost were a bignum truncated losing the least |
376 | | significant BITS bits. */ |
377 | | static lostFraction |
378 | | lostFractionThroughTruncation(const integerPart *parts, |
379 | | unsigned int partCount, |
380 | | unsigned int bits) |
381 | 1.15M | { |
382 | 1.15M | unsigned int lsb; |
383 | | |
384 | 1.15M | lsb = APInt::tcLSB(parts, partCount); |
385 | | |
386 | | /* Note this is guaranteed true if bits == 0, or LSB == -1U. */ |
387 | 1.15M | if (bits <= lsb) |
388 | 267k | return lfExactlyZero; |
389 | 892k | if (bits == lsb + 1) |
390 | 6.67k | return lfExactlyHalf; |
391 | 885k | if (bits <= partCount * integerPartWidth && |
392 | 865k | APInt::tcExtractBit(parts, bits - 1)) |
393 | 600k | return lfMoreThanHalf; |
394 | | |
395 | 285k | return lfLessThanHalf; |
396 | 885k | } |
397 | | |
398 | | /* Shift DST right BITS bits noting lost fraction. */ |
399 | | static lostFraction |
400 | | shiftRight(integerPart *dst, unsigned int parts, unsigned int bits) |
401 | 232k | { |
402 | 232k | lostFraction lost_fraction; |
403 | | |
404 | 232k | lost_fraction = lostFractionThroughTruncation(dst, parts, bits); |
405 | | |
406 | 232k | APInt::tcShiftRight(dst, parts, bits); |
407 | | |
408 | 232k | return lost_fraction; |
409 | 232k | } |
410 | | |
411 | | /* Combine the effect of two lost fractions. */ |
412 | | static lostFraction |
413 | | combineLostFractions(lostFraction moreSignificant, |
414 | | lostFraction lessSignificant) |
415 | 209k | { |
416 | 209k | if (lessSignificant != lfExactlyZero) { |
417 | 21.9k | if (moreSignificant == lfExactlyZero) |
418 | 8.30k | moreSignificant = lfLessThanHalf; |
419 | 13.6k | else if (moreSignificant == lfExactlyHalf) |
420 | 1.22k | moreSignificant = lfMoreThanHalf; |
421 | 21.9k | } |
422 | | |
423 | 209k | return moreSignificant; |
424 | 209k | } |
425 | | |
426 | | /* The error from the true value, in half-ulps, on multiplying two |
427 | | floating point numbers, which differ from the value they |
428 | | approximate by at most HUE1 and HUE2 half-ulps, is strictly less |
429 | | than the returned value. |
430 | | |
431 | | See "How to Read Floating Point Numbers Accurately" by William D |
432 | | Clinger. */ |
433 | | static unsigned int |
434 | | HUerrBound(bool inexactMultiply, unsigned int HUerr1, unsigned int HUerr2) |
435 | 767k | { |
436 | 767k | assert(HUerr1 < 2 || HUerr2 < 2 || (HUerr1 + HUerr2 < 8)); |
437 | | |
438 | 767k | if (HUerr1 + HUerr2 == 0) |
439 | 72.8k | return inexactMultiply * 2; /* <= inexactMultiply half-ulps. */ |
440 | 694k | else |
441 | 694k | return inexactMultiply + 2 * (HUerr1 + HUerr2); |
442 | 767k | } |
443 | | |
444 | | /* The number of ulps from the boundary (zero, or half if ISNEAREST) |
445 | | when the least significant BITS are truncated. BITS cannot be |
446 | | zero. */ |
447 | | static integerPart |
448 | | ulpsFromBoundary(const integerPart *parts, unsigned int bits, bool isNearest) |
449 | 767k | { |
450 | 767k | unsigned int count, partBits; |
451 | 767k | integerPart part, boundary; |
452 | | |
453 | 767k | assert(bits != 0); |
454 | | |
455 | 767k | bits--; |
456 | 767k | count = bits / integerPartWidth; |
457 | 767k | partBits = bits % integerPartWidth + 1; |
458 | | |
459 | 767k | part = parts[count] & (~(integerPart) 0 >> (integerPartWidth - partBits)); |
460 | | |
461 | 767k | if (isNearest) |
462 | 767k | boundary = (integerPart) 1 << (partBits - 1); |
463 | 0 | else |
464 | 0 | boundary = 0; |
465 | | |
466 | 767k | if (count == 0) { |
467 | 730k | if (part - boundary <= boundary - part) |
468 | 492k | return part - boundary; |
469 | 237k | else |
470 | 237k | return boundary - part; |
471 | 730k | } |
472 | | |
473 | 36.8k | if (part == boundary) { |
474 | 61.5k | while (--count) |
475 | 54.2k | if (parts[count]) |
476 | 643 | return ~(integerPart) 0; /* A lot. */ |
477 | | |
478 | 7.39k | return parts[0]; |
479 | 28.7k | } else if (part == boundary - 1) { |
480 | 65.3k | while (--count) |
481 | 48.6k | if (~parts[count]) |
482 | 1.51k | return ~(integerPart) 0; /* A lot. */ |
483 | | |
484 | 16.7k | return -parts[0]; |
485 | 18.2k | } |
486 | | |
487 | 10.5k | return ~(integerPart) 0; /* A lot. */ |
488 | 36.8k | } |
489 | | |
490 | | /* Place pow(5, power) in DST, and return the number of parts used. |
491 | | DST must be at least one part larger than size of the answer. */ |
492 | | static unsigned int |
493 | | powerOf5(integerPart *dst, unsigned int power) |
494 | 730k | { |
495 | 730k | static const integerPart firstEightPowers[] = { 1, 5, 25, 125, 625, 3125, |
496 | 730k | 15625, 78125 }; |
497 | 730k | integerPart pow5s[maxPowerOfFiveParts * 2 + 5]; |
498 | 730k | pow5s[0] = 78125 * 5; |
499 | | |
500 | 730k | unsigned int partsCount[16] = { 1 }; |
501 | 730k | integerPart scratch[maxPowerOfFiveParts], *p1, *p2, *pow5; |
502 | 730k | unsigned int result; |
503 | 730k | assert(power <= maxExponent); |
504 | | |
505 | 730k | p1 = dst; |
506 | 730k | p2 = scratch; |
507 | | |
508 | 730k | *p1 = firstEightPowers[power & 7]; |
509 | 730k | power >>= 3; |
510 | | |
511 | 730k | result = 1; |
512 | 730k | pow5 = pow5s; |
513 | | |
514 | 1.20M | for (unsigned int n = 0; power; power >>= 1, n++) { |
515 | 470k | unsigned int pc; |
516 | | |
517 | 470k | pc = partsCount[n]; |
518 | | |
519 | | /* Calculate pow(5,pow(2,n+3)) if we haven't yet. */ |
520 | 470k | if (pc == 0) { |
521 | 314k | pc = partsCount[n - 1]; |
522 | 314k | APInt::tcFullMultiply(pow5, pow5 - pc, pow5 - pc, pc, pc); |
523 | 314k | pc *= 2; |
524 | 314k | if (pow5[pc - 1] == 0) |
525 | 192k | pc--; |
526 | 314k | partsCount[n] = pc; |
527 | 314k | } |
528 | | |
529 | 470k | if (power & 1) { |
530 | 244k | integerPart *tmp; |
531 | | |
532 | 244k | APInt::tcFullMultiply(p2, p1, pow5, result, pc); |
533 | 244k | result += pc; |
534 | 244k | if (p2[result - 1] == 0) |
535 | 216k | result--; |
536 | | |
537 | | /* Now result is in p1 with partsCount parts and p2 is scratch |
538 | | space. */ |
539 | 244k | tmp = p1, p1 = p2, p2 = tmp; |
540 | 244k | } |
541 | | |
542 | 470k | pow5 += pc; |
543 | 470k | } |
544 | | |
545 | 730k | if (p1 != dst) |
546 | 108k | APInt::tcAssign(dst, p1, result); |
547 | | |
548 | 730k | return result; |
549 | 730k | } |
550 | | |
551 | | /* Zero at the end to avoid modular arithmetic when adding one; used |
552 | | when rounding up during hexadecimal output. */ |
553 | | static const char hexDigitsLower[] = "0123456789abcdef0"; |
554 | | static const char hexDigitsUpper[] = "0123456789ABCDEF0"; |
555 | | static const char infinityL[] = "infinity"; |
556 | | static const char infinityU[] = "INFINITY"; |
557 | | static const char NaNL[] = "nan"; |
558 | | static const char NaNU[] = "NAN"; |
559 | | |
560 | | /* Write out an integerPart in hexadecimal, starting with the most |
561 | | significant nibble. Write out exactly COUNT hexdigits, return |
562 | | COUNT. */ |
563 | | static unsigned int |
564 | | partAsHex (char *dst, integerPart part, unsigned int count, |
565 | | const char *hexDigitChars) |
566 | 0 | { |
567 | 0 | unsigned int result = count; |
568 | |
|
569 | 0 | assert(count != 0 && count <= integerPartWidth / 4); |
570 | | |
571 | 0 | part >>= (integerPartWidth - 4 * count); |
572 | 0 | while (count--) { |
573 | 0 | dst[count] = hexDigitChars[part & 0xf]; |
574 | 0 | part >>= 4; |
575 | 0 | } |
576 | |
|
577 | 0 | return result; |
578 | 0 | } |
579 | | |
580 | | /* Write out an unsigned decimal integer. */ |
581 | | static char * |
582 | | writeUnsignedDecimal (char *dst, unsigned int n) |
583 | 0 | { |
584 | 0 | char buff[40], *p; |
585 | |
|
586 | 0 | p = buff; |
587 | 0 | do |
588 | 0 | *p++ = '0' + n % 10; |
589 | 0 | while (n /= 10); |
590 | |
|
591 | 0 | do |
592 | 0 | *dst++ = *--p; |
593 | 0 | while (p != buff); |
594 | |
|
595 | 0 | return dst; |
596 | 0 | } |
597 | | |
598 | | /* Write out a signed decimal integer. */ |
599 | | static char * |
600 | | writeSignedDecimal (char *dst, int value) |
601 | 0 | { |
602 | 0 | if (value < 0) { |
603 | 0 | *dst++ = '-'; |
604 | 0 | dst = writeUnsignedDecimal(dst, -(unsigned) value); |
605 | 0 | } else |
606 | 0 | dst = writeUnsignedDecimal(dst, value); |
607 | |
|
608 | 0 | return dst; |
609 | 0 | } |
610 | | |
611 | | /* Constructors. */ |
612 | | void |
613 | | APFloat::initialize(const fltSemantics *ourSemantics) |
614 | 2.54M | { |
615 | 2.54M | unsigned int count; |
616 | | |
617 | 2.54M | semantics = ourSemantics; |
618 | 2.54M | count = partCount(); |
619 | 2.54M | if (count > 1) |
620 | 73.6k | significand.parts = new integerPart[count]; |
621 | 2.54M | } |
622 | | |
623 | | void |
624 | | APFloat::freeSignificand() |
625 | 2.56M | { |
626 | 2.56M | if (needsCleanup()) |
627 | 73.6k | delete [] significand.parts; |
628 | 2.56M | } |
629 | | |
630 | | void |
631 | | APFloat::assign(const APFloat &rhs) |
632 | 0 | { |
633 | 0 | assert(semantics == rhs.semantics); |
634 | | |
635 | 0 | sign = rhs.sign; |
636 | 0 | category = rhs.category; |
637 | 0 | exponent = rhs.exponent; |
638 | 0 | if (isFiniteNonZero() || category == fcNaN) |
639 | 0 | copySignificand(rhs); |
640 | 0 | } |
641 | | |
642 | | void |
643 | | APFloat::copySignificand(const APFloat &rhs) |
644 | 0 | { |
645 | 0 | assert(isFiniteNonZero() || category == fcNaN); |
646 | 0 | assert(rhs.partCount() >= partCount()); |
647 | | |
648 | 0 | APInt::tcAssign(significandParts(), rhs.significandParts(), |
649 | 0 | partCount()); |
650 | 0 | } |
651 | | |
652 | | /* Make this number a NaN, with an arbitrary but deterministic value |
653 | | for the significand. If double or longer, this is a signalling NaN, |
654 | | which may not be ideal. If float, this is QNaN(0). */ |
655 | | void APFloat::makeNaN(bool SNaN, bool Negative, const APInt *fill) |
656 | 4.88k | { |
657 | 4.88k | category = fcNaN; |
658 | 4.88k | sign = Negative; |
659 | | |
660 | 4.88k | integerPart *significand = significandParts(); |
661 | 4.88k | unsigned numParts = partCount(); |
662 | | |
663 | | // Set the significand bits to the fill. |
664 | 4.88k | if (!fill || fill->getNumWords() < numParts) |
665 | 0 | APInt::tcSet(significand, 0, numParts); |
666 | 4.88k | if (fill) { |
667 | 4.88k | APInt::tcAssign(significand, fill->getRawData(), |
668 | 4.88k | std::min(fill->getNumWords(), numParts)); |
669 | | |
670 | | // Zero out the excess bits of the significand. |
671 | 4.88k | unsigned bitsToPreserve = semantics->precision - 1; |
672 | 4.88k | unsigned part = bitsToPreserve / 64; |
673 | 4.88k | bitsToPreserve %= 64; |
674 | 4.88k | significand[part] &= ((1ULL << bitsToPreserve) - 1); |
675 | 4.88k | for (part++; part != numParts; ++part) |
676 | 0 | significand[part] = 0; |
677 | 4.88k | } |
678 | | |
679 | 4.88k | unsigned QNaNBit = semantics->precision - 2; |
680 | | |
681 | 4.88k | if (SNaN) { |
682 | | // We always have to clear the QNaN bit to make it an SNaN. |
683 | 0 | APInt::tcClearBit(significand, QNaNBit); |
684 | | |
685 | | // If there are no bits set in the payload, we have to set |
686 | | // *something* to make it a NaN instead of an infinity; |
687 | | // conventionally, this is the next bit down from the QNaN bit. |
688 | 0 | if (APInt::tcIsZero(significand, numParts)) |
689 | 0 | APInt::tcSetBit(significand, QNaNBit - 1); |
690 | 4.88k | } else { |
691 | | // We always have to set the QNaN bit to make it a QNaN. |
692 | 4.88k | APInt::tcSetBit(significand, QNaNBit); |
693 | 4.88k | } |
694 | | |
695 | | // For x87 extended precision, we want to make a NaN, not a |
696 | | // pseudo-NaN. Maybe we should expose the ability to make |
697 | | // pseudo-NaNs? |
698 | 4.88k | if (semantics == &APFloat::x87DoubleExtended) |
699 | 0 | APInt::tcSetBit(significand, QNaNBit + 1); |
700 | 4.88k | } |
701 | | |
702 | | APFloat APFloat::makeNaN(const fltSemantics &Sem, bool SNaN, bool Negative, |
703 | 4.88k | const APInt *fill) { |
704 | 4.88k | APFloat value(Sem, uninitialized); |
705 | 4.88k | value.makeNaN(SNaN, Negative, fill); |
706 | 4.88k | return value; |
707 | 4.88k | } |
708 | | |
709 | | APFloat & |
710 | | APFloat::operator=(const APFloat &rhs) |
711 | 0 | { |
712 | 0 | if (this != &rhs) { |
713 | 0 | if (semantics != rhs.semantics) { |
714 | 0 | freeSignificand(); |
715 | 0 | initialize(rhs.semantics); |
716 | 0 | } |
717 | 0 | assign(rhs); |
718 | 0 | } |
719 | |
|
720 | 0 | return *this; |
721 | 0 | } |
722 | | |
723 | | APFloat & |
724 | 16.4k | APFloat::operator=(APFloat &&rhs) { |
725 | 16.4k | freeSignificand(); |
726 | | |
727 | 16.4k | semantics = rhs.semantics; |
728 | 16.4k | significand = rhs.significand; |
729 | 16.4k | exponent = rhs.exponent; |
730 | 16.4k | category = rhs.category; |
731 | 16.4k | sign = rhs.sign; |
732 | | |
733 | 16.4k | rhs.semantics = &Bogus; |
734 | 16.4k | return *this; |
735 | 16.4k | } |
736 | | |
737 | | bool |
738 | 0 | APFloat::isDenormal() const { |
739 | 0 | return isFiniteNonZero() && (exponent == semantics->minExponent) && |
740 | 0 | (APInt::tcExtractBit(significandParts(), |
741 | 0 | semantics->precision - 1) == 0); |
742 | 0 | } |
743 | | |
744 | | bool |
745 | 0 | APFloat::isSmallest() const { |
746 | | // The smallest number by magnitude in our format will be the smallest |
747 | | // denormal, i.e. the floating point number with exponent being minimum |
748 | | // exponent and significand bitwise equal to 1 (i.e. with MSB equal to 0). |
749 | 0 | return isFiniteNonZero() && exponent == semantics->minExponent && |
750 | 0 | significandMSB() == 0; |
751 | 0 | } |
752 | | |
753 | 0 | bool APFloat::isSignificandAllOnes() const { |
754 | | // Test if the significand excluding the integral bit is all ones. This allows |
755 | | // us to test for binade boundaries. |
756 | 0 | const integerPart *Parts = significandParts(); |
757 | 0 | const unsigned PartCount = partCount(); |
758 | 0 | for (unsigned i = 0; i < PartCount - 1; i++) |
759 | 0 | if (~Parts[i]) |
760 | 0 | return false; |
761 | | |
762 | | // Set the unused high bits to all ones when we compare. |
763 | 0 | const unsigned NumHighBits = |
764 | 0 | PartCount*integerPartWidth - semantics->precision + 1; |
765 | 0 | assert(NumHighBits <= integerPartWidth && "Can not have more high bits to " |
766 | 0 | "fill than integerPartWidth"); |
767 | 0 | const integerPart HighBitFill = |
768 | 0 | ~integerPart(0) << (integerPartWidth - NumHighBits); |
769 | 0 | if (~(Parts[PartCount - 1] | HighBitFill)) |
770 | 0 | return false; |
771 | | |
772 | 0 | return true; |
773 | 0 | } |
774 | | |
775 | 0 | bool APFloat::isSignificandAllZeros() const { |
776 | | // Test if the significand excluding the integral bit is all zeros. This |
777 | | // allows us to test for binade boundaries. |
778 | 0 | const integerPart *Parts = significandParts(); |
779 | 0 | const unsigned PartCount = partCount(); |
780 | |
|
781 | 0 | for (unsigned i = 0; i < PartCount - 1; i++) |
782 | 0 | if (Parts[i]) |
783 | 0 | return false; |
784 | | |
785 | 0 | const unsigned NumHighBits = |
786 | 0 | PartCount*integerPartWidth - semantics->precision + 1; |
787 | 0 | assert(NumHighBits <= integerPartWidth && "Can not have more high bits to " |
788 | 0 | "clear than integerPartWidth"); |
789 | 0 | const integerPart HighBitMask = ~integerPart(0) >> NumHighBits; |
790 | |
|
791 | 0 | if (Parts[PartCount - 1] & HighBitMask) |
792 | 0 | return false; |
793 | | |
794 | 0 | return true; |
795 | 0 | } |
796 | | |
797 | | bool |
798 | 0 | APFloat::isLargest() const { |
799 | | // The largest number by magnitude in our format will be the floating point |
800 | | // number with maximum exponent and with significand that is all ones. |
801 | 0 | return isFiniteNonZero() && exponent == semantics->maxExponent |
802 | 0 | && isSignificandAllOnes(); |
803 | 0 | } |
804 | | |
805 | | bool |
806 | 0 | APFloat::isInteger() const { |
807 | | // This could be made more efficient; I'm going for obviously correct. |
808 | 0 | if (!isFinite()) return false; |
809 | 0 | APFloat truncated = *this; |
810 | 0 | truncated.roundToIntegral(rmTowardZero); |
811 | 0 | return compare(truncated) == cmpEqual; |
812 | 0 | } |
813 | | |
814 | | bool |
815 | 0 | APFloat::bitwiseIsEqual(const APFloat &rhs) const { |
816 | 0 | if (this == &rhs) |
817 | 0 | return true; |
818 | 0 | if (semantics != rhs.semantics || |
819 | 0 | category != rhs.category || |
820 | 0 | sign != rhs.sign) |
821 | 0 | return false; |
822 | 0 | if (category==fcZero || category==fcInfinity) |
823 | 0 | return true; |
824 | | |
825 | 0 | if (isFiniteNonZero() && exponent != rhs.exponent) |
826 | 0 | return false; |
827 | | |
828 | 0 | return std::equal(significandParts(), significandParts() + partCount(), |
829 | 0 | rhs.significandParts()); |
830 | 0 | } |
831 | | |
832 | 0 | APFloat::APFloat(const fltSemantics &ourSemantics, integerPart value) { |
833 | 0 | initialize(&ourSemantics); |
834 | 0 | sign = 0; |
835 | 0 | category = fcNormal; |
836 | 0 | zeroSignificand(); |
837 | 0 | exponent = ourSemantics.precision - 1; |
838 | 0 | significandParts()[0] = value; |
839 | 0 | normalize(rmNearestTiesToEven, lfExactlyZero); |
840 | 0 | } |
841 | | |
842 | 880k | APFloat::APFloat(const fltSemantics &ourSemantics) { |
843 | 880k | initialize(&ourSemantics); |
844 | 880k | category = fcZero; |
845 | 880k | sign = false; |
846 | 880k | } |
847 | | |
848 | 783k | APFloat::APFloat(const fltSemantics &ourSemantics, uninitializedTag tag) { |
849 | | // Allocates storage if necessary but does not initialize it. |
850 | 783k | initialize(&ourSemantics); |
851 | 783k | } |
852 | | |
853 | 884k | APFloat::APFloat(const fltSemantics &ourSemantics, StringRef text) { |
854 | 884k | initialize(&ourSemantics); |
855 | 884k | convertFromString(text, rmNearestTiesToEven); |
856 | 884k | } |
857 | | |
858 | 0 | APFloat::APFloat(const APFloat &rhs) { |
859 | 0 | initialize(rhs.semantics); |
860 | 0 | assign(rhs); |
861 | 0 | } |
862 | | |
863 | 0 | APFloat::APFloat(APFloat &&rhs) : semantics(&Bogus) { |
864 | 0 | *this = std::move(rhs); |
865 | 0 | } |
866 | | |
867 | | APFloat::~APFloat() |
868 | 2.54M | { |
869 | 2.54M | freeSignificand(); |
870 | 2.54M | } |
871 | | |
872 | | // Profile - This method 'profiles' an APFloat for use with FoldingSet. |
873 | 0 | void APFloat::Profile(FoldingSetNodeID& ID) const { |
874 | 0 | ID.Add(bitcastToAPInt()); |
875 | 0 | } |
876 | | |
877 | | unsigned int |
878 | | APFloat::partCount() const |
879 | 31.0M | { |
880 | 31.0M | return partCountForBits(semantics->precision + 1); |
881 | 31.0M | } |
882 | | |
883 | | unsigned int |
884 | | APFloat::semanticsPrecision(const fltSemantics &semantics) |
885 | 0 | { |
886 | 0 | return semantics.precision; |
887 | 0 | } |
888 | | APFloat::ExponentType |
889 | | APFloat::semanticsMaxExponent(const fltSemantics &semantics) |
890 | 0 | { |
891 | 0 | return semantics.maxExponent; |
892 | 0 | } |
893 | | APFloat::ExponentType |
894 | | APFloat::semanticsMinExponent(const fltSemantics &semantics) |
895 | 0 | { |
896 | 0 | return semantics.minExponent; |
897 | 0 | } |
898 | | unsigned int |
899 | | APFloat::semanticsSizeInBits(const fltSemantics &semantics) |
900 | 0 | { |
901 | 0 | return semantics.sizeInBits; |
902 | 0 | } |
903 | | |
904 | | const integerPart * |
905 | | APFloat::significandParts() const |
906 | 4.60M | { |
907 | 4.60M | return const_cast<APFloat *>(this)->significandParts(); |
908 | 4.60M | } |
909 | | |
910 | | integerPart * |
911 | | APFloat::significandParts() |
912 | 15.0M | { |
913 | 15.0M | if (partCount() > 1) |
914 | 506k | return significand.parts; |
915 | 14.5M | else |
916 | 14.5M | return &significand.part; |
917 | 15.0M | } |
918 | | |
919 | | void |
920 | | APFloat::zeroSignificand() |
921 | 143k | { |
922 | 143k | APInt::tcSet(significandParts(), 0, partCount()); |
923 | 143k | } |
924 | | |
925 | | /* Increment an fcNormal floating point number's significand. */ |
926 | | void |
927 | | APFloat::incrementSignificand() |
928 | 583k | { |
929 | 583k | integerPart carry; |
930 | | |
931 | 583k | carry = APInt::tcIncrement(significandParts(), partCount()); |
932 | | |
933 | | /* Our callers should never cause us to overflow. */ |
934 | 583k | assert(carry == 0); |
935 | 583k | (void)carry; |
936 | 583k | } |
937 | | |
938 | | /* Add the significand of the RHS. Returns the carry flag. */ |
939 | | integerPart |
940 | | APFloat::addSignificand(const APFloat &rhs) |
941 | 0 | { |
942 | 0 | integerPart *parts; |
943 | |
|
944 | 0 | parts = significandParts(); |
945 | |
|
946 | 0 | assert(semantics == rhs.semantics); |
947 | 0 | assert(exponent == rhs.exponent); |
948 | | |
949 | 0 | return APInt::tcAdd(parts, rhs.significandParts(), 0, partCount()); |
950 | 0 | } |
951 | | |
952 | | /* Subtract the significand of the RHS with a borrow flag. Returns |
953 | | the borrow flag. */ |
954 | | integerPart |
955 | | APFloat::subtractSignificand(const APFloat &rhs, integerPart borrow) |
956 | 0 | { |
957 | 0 | integerPart *parts; |
958 | |
|
959 | 0 | parts = significandParts(); |
960 | |
|
961 | 0 | assert(semantics == rhs.semantics); |
962 | 0 | assert(exponent == rhs.exponent); |
963 | | |
964 | 0 | return APInt::tcSubtract(parts, rhs.significandParts(), borrow, |
965 | 0 | partCount()); |
966 | 0 | } |
967 | | |
968 | | /* Multiply the significand of the RHS. If ADDEND is non-NULL, add it |
969 | | on to the full-precision result of the multiplication. Returns the |
970 | | lost fraction. */ |
971 | | lostFraction |
972 | | APFloat::multiplySignificand(const APFloat &rhs, const APFloat *addend) |
973 | 91.9k | { |
974 | 91.9k | unsigned int omsb; // One, not zero, based MSB. |
975 | 91.9k | unsigned int partsCount, newPartsCount, precision; |
976 | 91.9k | integerPart *lhsSignificand; |
977 | 91.9k | integerPart scratch[4]; |
978 | 91.9k | integerPart *fullSignificand; |
979 | 91.9k | lostFraction lost_fraction; |
980 | 91.9k | bool ignored; |
981 | | |
982 | 91.9k | assert(semantics == rhs.semantics); |
983 | | |
984 | 91.9k | precision = semantics->precision; |
985 | | |
986 | | // Allocate space for twice as many bits as the original significand, plus one |
987 | | // extra bit for the addition to overflow into. |
988 | 91.9k | newPartsCount = partCountForBits(precision * 2 + 1); |
989 | | |
990 | 91.9k | if (newPartsCount > 4) |
991 | 0 | fullSignificand = new integerPart[newPartsCount]; |
992 | 91.9k | else |
993 | 91.9k | fullSignificand = scratch; |
994 | | |
995 | 91.9k | lhsSignificand = significandParts(); |
996 | 91.9k | partsCount = partCount(); |
997 | | |
998 | 91.9k | APInt::tcFullMultiply(fullSignificand, lhsSignificand, |
999 | 91.9k | rhs.significandParts(), partsCount, partsCount); |
1000 | | |
1001 | 91.9k | lost_fraction = lfExactlyZero; |
1002 | 91.9k | omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1; |
1003 | 91.9k | exponent += rhs.exponent; |
1004 | | |
1005 | | // Assume the operands involved in the multiplication are single-precision |
1006 | | // FP, and the two multiplicants are: |
1007 | | // *this = a23 . a22 ... a0 * 2^e1 |
1008 | | // rhs = b23 . b22 ... b0 * 2^e2 |
1009 | | // the result of multiplication is: |
1010 | | // *this = c48 c47 c46 . c45 ... c0 * 2^(e1+e2) |
1011 | | // Note that there are three significant bits at the left-hand side of the |
1012 | | // radix point: two for the multiplication, and an overflow bit for the |
1013 | | // addition (that will always be zero at this point). Move the radix point |
1014 | | // toward left by two bits, and adjust exponent accordingly. |
1015 | 91.9k | exponent += 2; |
1016 | | |
1017 | 91.9k | if (addend && addend->isNonZero()) { |
1018 | | // The intermediate result of the multiplication has "2 * precision" |
1019 | | // signicant bit; adjust the addend to be consistent with mul result. |
1020 | | // |
1021 | 0 | Significand savedSignificand = significand; |
1022 | 0 | const fltSemantics *savedSemantics = semantics; |
1023 | 0 | fltSemantics extendedSemantics; |
1024 | 0 | opStatus status; |
1025 | 0 | unsigned int extendedPrecision; |
1026 | | |
1027 | | // Normalize our MSB to one below the top bit to allow for overflow. |
1028 | 0 | extendedPrecision = 2 * precision + 1; |
1029 | 0 | if (omsb != extendedPrecision - 1) { |
1030 | 0 | assert(extendedPrecision > omsb); |
1031 | 0 | APInt::tcShiftLeft(fullSignificand, newPartsCount, |
1032 | 0 | (extendedPrecision - 1) - omsb); |
1033 | 0 | exponent -= (extendedPrecision - 1) - omsb; |
1034 | 0 | } |
1035 | | |
1036 | | /* Create new semantics. */ |
1037 | 0 | extendedSemantics = *semantics; |
1038 | 0 | extendedSemantics.precision = extendedPrecision; |
1039 | |
|
1040 | 0 | if (newPartsCount == 1) |
1041 | 0 | significand.part = fullSignificand[0]; |
1042 | 0 | else |
1043 | 0 | significand.parts = fullSignificand; |
1044 | 0 | semantics = &extendedSemantics; |
1045 | |
|
1046 | 0 | APFloat extendedAddend(*addend); |
1047 | 0 | status = extendedAddend.convert(extendedSemantics, rmTowardZero, &ignored); |
1048 | 0 | assert(status == opOK); |
1049 | 0 | (void)status; |
1050 | | |
1051 | | // Shift the significand of the addend right by one bit. This guarantees |
1052 | | // that the high bit of the significand is zero (same as fullSignificand), |
1053 | | // so the addition will overflow (if it does overflow at all) into the top bit. |
1054 | 0 | lost_fraction = extendedAddend.shiftSignificandRight(1); |
1055 | 0 | assert(lost_fraction == lfExactlyZero && |
1056 | 0 | "Lost precision while shifting addend for fused-multiply-add."); |
1057 | | |
1058 | 0 | lost_fraction = addOrSubtractSignificand(extendedAddend, false); |
1059 | | |
1060 | | /* Restore our state. */ |
1061 | 0 | if (newPartsCount == 1) |
1062 | 0 | fullSignificand[0] = significand.part; |
1063 | 0 | significand = savedSignificand; |
1064 | 0 | semantics = savedSemantics; |
1065 | |
|
1066 | 0 | omsb = APInt::tcMSB(fullSignificand, newPartsCount) + 1; |
1067 | 0 | } |
1068 | | |
1069 | | // Convert the result having "2 * precision" significant-bits back to the one |
1070 | | // having "precision" significant-bits. First, move the radix point from |
1071 | | // poision "2*precision - 1" to "precision - 1". The exponent need to be |
1072 | | // adjusted by "2*precision - 1" - "precision - 1" = "precision". |
1073 | 91.9k | exponent -= precision + 1; |
1074 | | |
1075 | | // In case MSB resides at the left-hand side of radix point, shift the |
1076 | | // mantissa right by some amount to make sure the MSB reside right before |
1077 | | // the radix point (i.e. "MSB . rest-significant-bits"). |
1078 | | // |
1079 | | // Note that the result is not normalized when "omsb < precision". So, the |
1080 | | // caller needs to call APFloat::normalize() if normalized value is expected. |
1081 | 91.9k | if (omsb > precision) { |
1082 | 91.9k | unsigned int bits, significantParts; |
1083 | 91.9k | lostFraction lf; |
1084 | | |
1085 | 91.9k | bits = omsb - precision; |
1086 | 91.9k | significantParts = partCountForBits(omsb); |
1087 | 91.9k | lf = shiftRight(fullSignificand, significantParts, bits); |
1088 | 91.9k | lost_fraction = combineLostFractions(lf, lost_fraction); |
1089 | 91.9k | exponent += bits; |
1090 | 91.9k | } |
1091 | | |
1092 | 91.9k | APInt::tcAssign(lhsSignificand, fullSignificand, partsCount); |
1093 | | |
1094 | 91.9k | if (newPartsCount > 4) |
1095 | 0 | delete [] fullSignificand; |
1096 | | |
1097 | 91.9k | return lost_fraction; |
1098 | 91.9k | } |
1099 | | |
1100 | | /* Multiply the significands of LHS and RHS to DST. */ |
1101 | | lostFraction |
1102 | | APFloat::divideSignificand(const APFloat &rhs) |
1103 | 675k | { |
1104 | 675k | unsigned int bit, i, partsCount; |
1105 | 675k | const integerPart *rhsSignificand; |
1106 | 675k | integerPart *lhsSignificand, *dividend, *divisor; |
1107 | 675k | integerPart scratch[4]; |
1108 | 675k | lostFraction lost_fraction; |
1109 | | |
1110 | 675k | assert(semantics == rhs.semantics); |
1111 | | |
1112 | 675k | lhsSignificand = significandParts(); |
1113 | 675k | rhsSignificand = rhs.significandParts(); |
1114 | 675k | partsCount = partCount(); |
1115 | | |
1116 | 675k | if (partsCount > 2) |
1117 | 7.46k | dividend = new integerPart[partsCount * 2]; |
1118 | 667k | else |
1119 | 667k | dividend = scratch; |
1120 | | |
1121 | 675k | divisor = dividend + partsCount; |
1122 | | |
1123 | | /* Copy the dividend and divisor as they will be modified in-place. */ |
1124 | 1.50M | for (i = 0; i < partsCount; i++) { |
1125 | 830k | dividend[i] = lhsSignificand[i]; |
1126 | 830k | divisor[i] = rhsSignificand[i]; |
1127 | 830k | lhsSignificand[i] = 0; |
1128 | 830k | } |
1129 | | |
1130 | 675k | exponent -= rhs.exponent; |
1131 | | |
1132 | 675k | unsigned int precision = semantics->precision; |
1133 | | |
1134 | | /* Normalize the divisor. */ |
1135 | 675k | bit = precision - APInt::tcMSB(divisor, partsCount) - 1; |
1136 | 675k | if (bit) { |
1137 | 0 | exponent += bit; |
1138 | 0 | APInt::tcShiftLeft(divisor, partsCount, bit); |
1139 | 0 | } |
1140 | | |
1141 | | /* Normalize the dividend. */ |
1142 | 675k | bit = precision - APInt::tcMSB(dividend, partsCount) - 1; |
1143 | 675k | if (bit) { |
1144 | 0 | exponent -= bit; |
1145 | 0 | APInt::tcShiftLeft(dividend, partsCount, bit); |
1146 | 0 | } |
1147 | | |
1148 | | /* Ensure the dividend >= divisor initially for the loop below. |
1149 | | Incidentally, this means that the division loop below is |
1150 | | guaranteed to set the integer bit to one. */ |
1151 | 675k | if (APInt::tcCompare(dividend, divisor, partsCount) < 0) { |
1152 | 520k | exponent--; |
1153 | 520k | APInt::tcShiftLeft(dividend, partsCount, 1); |
1154 | 520k | assert(APInt::tcCompare(dividend, divisor, partsCount) >= 0); |
1155 | 520k | } |
1156 | | |
1157 | | /* Long division. */ |
1158 | 53.1M | for (bit = precision; bit; bit -= 1) { |
1159 | 52.4M | if (APInt::tcCompare(dividend, divisor, partsCount) >= 0) { |
1160 | 26.0M | APInt::tcSubtract(dividend, divisor, 0, partsCount); |
1161 | 26.0M | APInt::tcSetBit(lhsSignificand, bit - 1); |
1162 | 26.0M | } |
1163 | | |
1164 | 52.4M | APInt::tcShiftLeft(dividend, partsCount, 1); |
1165 | 52.4M | } |
1166 | | |
1167 | | /* Figure out the lost fraction. */ |
1168 | 675k | int cmp = APInt::tcCompare(dividend, divisor, partsCount); |
1169 | | |
1170 | 675k | if (cmp > 0) |
1171 | 179k | lost_fraction = lfMoreThanHalf; |
1172 | 495k | else if (cmp == 0) |
1173 | 0 | lost_fraction = lfExactlyHalf; |
1174 | 495k | else if (APInt::tcIsZero(dividend, partsCount)) |
1175 | 41.6k | lost_fraction = lfExactlyZero; |
1176 | 454k | else |
1177 | 454k | lost_fraction = lfLessThanHalf; |
1178 | | |
1179 | 675k | if (partsCount > 2) |
1180 | 7.46k | delete [] dividend; |
1181 | | |
1182 | 675k | return lost_fraction; |
1183 | 675k | } |
1184 | | |
1185 | | unsigned int |
1186 | | APFloat::significandMSB() const |
1187 | 2.99M | { |
1188 | 2.99M | return APInt::tcMSB(significandParts(), partCount()); |
1189 | 2.99M | } |
1190 | | |
1191 | | unsigned int |
1192 | | APFloat::significandLSB() const |
1193 | 0 | { |
1194 | 0 | return APInt::tcLSB(significandParts(), partCount()); |
1195 | 0 | } |
1196 | | |
1197 | | /* Note that a zero result is NOT normalized to fcZero. */ |
1198 | | lostFraction |
1199 | | APFloat::shiftSignificandRight(unsigned int bits) |
1200 | 140k | { |
1201 | | /* Our exponent should not overflow. */ |
1202 | 140k | assert((ExponentType) (exponent + bits) >= exponent); |
1203 | | |
1204 | 140k | exponent += bits; |
1205 | | |
1206 | 140k | return shiftRight(significandParts(), partCount(), bits); |
1207 | 140k | } |
1208 | | |
1209 | | /* Shift the significand left BITS bits, subtract BITS from its exponent. */ |
1210 | | void |
1211 | | APFloat::shiftSignificandLeft(unsigned int bits) |
1212 | 1.33M | { |
1213 | 1.33M | assert(bits < semantics->precision); |
1214 | | |
1215 | 1.33M | if (bits) { |
1216 | 1.33M | unsigned int partsCount = partCount(); |
1217 | | |
1218 | 1.33M | APInt::tcShiftLeft(significandParts(), partsCount, bits); |
1219 | 1.33M | exponent -= bits; |
1220 | | |
1221 | 1.33M | assert(!APInt::tcIsZero(significandParts(), partsCount)); |
1222 | 1.33M | } |
1223 | 1.33M | } |
1224 | | |
1225 | | APFloat::cmpResult |
1226 | | APFloat::compareAbsoluteValue(const APFloat &rhs) const |
1227 | 0 | { |
1228 | 0 | int compare; |
1229 | |
|
1230 | 0 | assert(semantics == rhs.semantics); |
1231 | 0 | assert(isFiniteNonZero()); |
1232 | 0 | assert(rhs.isFiniteNonZero()); |
1233 | | |
1234 | 0 | compare = exponent - rhs.exponent; |
1235 | | |
1236 | | /* If exponents are equal, do an unsigned bignum comparison of the |
1237 | | significands. */ |
1238 | 0 | if (compare == 0) |
1239 | 0 | compare = APInt::tcCompare(significandParts(), rhs.significandParts(), |
1240 | 0 | partCount()); |
1241 | |
|
1242 | 0 | if (compare > 0) |
1243 | 0 | return cmpGreaterThan; |
1244 | 0 | else if (compare < 0) |
1245 | 0 | return cmpLessThan; |
1246 | 0 | else |
1247 | 0 | return cmpEqual; |
1248 | 0 | } |
1249 | | |
1250 | | /* Handle overflow. Sign is preserved. We either become infinity or |
1251 | | the largest finite number. */ |
1252 | | APFloat::opStatus |
1253 | | APFloat::handleOverflow(roundingMode rounding_mode) |
1254 | 24.4k | { |
1255 | | /* Infinity? */ |
1256 | 24.4k | if (rounding_mode == rmNearestTiesToEven || |
1257 | 0 | rounding_mode == rmNearestTiesToAway || |
1258 | 0 | (rounding_mode == rmTowardPositive && !sign) || |
1259 | 24.4k | (rounding_mode == rmTowardNegative && sign)) { |
1260 | 24.4k | category = fcInfinity; |
1261 | 24.4k | return (opStatus) (opOverflow | opInexact); |
1262 | 24.4k | } |
1263 | | |
1264 | | /* Otherwise we become the largest finite number. */ |
1265 | 0 | category = fcNormal; |
1266 | 0 | exponent = semantics->maxExponent; |
1267 | 0 | APInt::tcSetLeastSignificantBits(significandParts(), partCount(), |
1268 | 0 | semantics->precision); |
1269 | |
|
1270 | 0 | return opInexact; |
1271 | 24.4k | } |
1272 | | |
1273 | | /* Returns TRUE if, when truncating the current number, with BIT the |
1274 | | new LSB, with the given lost fraction and rounding mode, the result |
1275 | | would need to be rounded away from zero (i.e., by increasing the |
1276 | | signficand). This routine must work for fcZero of both signs, and |
1277 | | fcNormal numbers. */ |
1278 | | bool |
1279 | | APFloat::roundAwayFromZero(roundingMode rounding_mode, |
1280 | | lostFraction lost_fraction, |
1281 | | unsigned int bit) const |
1282 | 882k | { |
1283 | | /* NaNs and infinities should not have lost fractions. */ |
1284 | 882k | assert(isFiniteNonZero() || category == fcZero); |
1285 | | |
1286 | | /* Current callers never pass this so we don't handle it. */ |
1287 | 882k | assert(lost_fraction != lfExactlyZero); |
1288 | | |
1289 | 882k | switch (rounding_mode) { |
1290 | 0 | case rmNearestTiesToAway: |
1291 | 0 | return lost_fraction == lfExactlyHalf || lost_fraction == lfMoreThanHalf; |
1292 | | |
1293 | 882k | case rmNearestTiesToEven: |
1294 | 882k | if (lost_fraction == lfMoreThanHalf) |
1295 | 580k | return true; |
1296 | | |
1297 | | /* Our zeroes don't have a significand to test. */ |
1298 | 302k | if (lost_fraction == lfExactlyHalf && category != fcZero) |
1299 | 3.20k | return APInt::tcExtractBit(significandParts(), bit); |
1300 | | |
1301 | 298k | return false; |
1302 | | |
1303 | 0 | case rmTowardZero: |
1304 | 0 | return false; |
1305 | | |
1306 | 0 | case rmTowardPositive: |
1307 | 0 | return !sign; |
1308 | | |
1309 | 0 | case rmTowardNegative: |
1310 | 0 | return sign; |
1311 | 882k | } |
1312 | 882k | llvm_unreachable("Invalid rounding mode found"); |
1313 | 882k | } |
1314 | | |
1315 | | APFloat::opStatus |
1316 | | APFloat::normalize(roundingMode rounding_mode, |
1317 | | lostFraction lost_fraction) |
1318 | 2.40M | { |
1319 | 2.40M | unsigned int omsb; /* One, not zero, based MSB. */ |
1320 | 2.40M | int exponentChange; |
1321 | | |
1322 | 2.40M | if (!isFiniteNonZero()) |
1323 | 0 | return opOK; |
1324 | | |
1325 | | /* Before rounding normalize the exponent of fcNormal numbers. */ |
1326 | 2.40M | omsb = significandMSB() + 1; |
1327 | | |
1328 | 2.40M | if (omsb) { |
1329 | | /* OMSB is numbered from 1. We want to place it in the integer |
1330 | | bit numbered PRECISION if possible, with a compensating change in |
1331 | | the exponent. */ |
1332 | 2.38M | exponentChange = omsb - semantics->precision; |
1333 | | |
1334 | | /* If the resulting exponent is too high, overflow according to |
1335 | | the rounding mode. */ |
1336 | 2.38M | if (exponent + exponentChange > semantics->maxExponent) |
1337 | 9.63k | return handleOverflow(rounding_mode); |
1338 | | |
1339 | | /* Subnormal numbers have exponent minExponent, and their MSB |
1340 | | is forced based on that. */ |
1341 | 2.37M | if (exponent + exponentChange < semantics->minExponent) |
1342 | 21.9k | exponentChange = semantics->minExponent - exponent; |
1343 | | |
1344 | | /* Shifting left is easy as we don't lose precision. */ |
1345 | 2.37M | if (exponentChange < 0) { |
1346 | 1.33M | assert(lost_fraction == lfExactlyZero); |
1347 | | |
1348 | 1.33M | shiftSignificandLeft(-exponentChange); |
1349 | | |
1350 | 1.33M | return opOK; |
1351 | 1.33M | } |
1352 | | |
1353 | 1.04M | if (exponentChange > 0) { |
1354 | 118k | lostFraction lf; |
1355 | | |
1356 | | /* Shift right and capture any new lost fraction. */ |
1357 | 118k | lf = shiftSignificandRight(exponentChange); |
1358 | | |
1359 | 118k | lost_fraction = combineLostFractions(lf, lost_fraction); |
1360 | | |
1361 | | /* Keep OMSB up-to-date. */ |
1362 | 118k | if (omsb > (unsigned) exponentChange) |
1363 | 98.8k | omsb -= exponentChange; |
1364 | 19.2k | else |
1365 | 19.2k | omsb = 0; |
1366 | 118k | } |
1367 | 1.04M | } |
1368 | | |
1369 | | /* Now round the number according to rounding_mode given the lost |
1370 | | fraction. */ |
1371 | | |
1372 | | /* As specified in IEEE 754, since we do not trap we do not report |
1373 | | underflow for exact results. */ |
1374 | 1.06M | if (lost_fraction == lfExactlyZero) { |
1375 | | /* Canonicalize zeroes. */ |
1376 | 177k | if (omsb == 0) |
1377 | 4.27k | category = fcZero; |
1378 | | |
1379 | 177k | return opOK; |
1380 | 177k | } |
1381 | | |
1382 | | /* Increment the significand if we're rounding away from zero. */ |
1383 | 882k | if (roundAwayFromZero(rounding_mode, lost_fraction, 0)) { |
1384 | 583k | if (omsb == 0) |
1385 | 3.59k | exponent = semantics->minExponent; |
1386 | | |
1387 | 583k | incrementSignificand(); |
1388 | 583k | omsb = significandMSB() + 1; |
1389 | | |
1390 | | /* Did the significand increment overflow? */ |
1391 | 583k | if (omsb == (unsigned) semantics->precision + 1) { |
1392 | | /* Renormalize by incrementing the exponent and shifting our |
1393 | | significand right one. However if we already have the |
1394 | | maximum exponent we overflow to infinity. */ |
1395 | 22.8k | if (exponent == semantics->maxExponent) { |
1396 | 10 | category = fcInfinity; |
1397 | | |
1398 | 10 | return (opStatus) (opOverflow | opInexact); |
1399 | 10 | } |
1400 | | |
1401 | 22.8k | shiftSignificandRight(1); |
1402 | | |
1403 | 22.8k | return opInexact; |
1404 | 22.8k | } |
1405 | 583k | } |
1406 | | |
1407 | | /* The normal case - we were and are not denormal, and any |
1408 | | significand increment above didn't overflow. */ |
1409 | 860k | if (omsb == semantics->precision) |
1410 | 822k | return opInexact; |
1411 | | |
1412 | | /* We have a non-zero denormal. */ |
1413 | 860k | assert(omsb < semantics->precision); |
1414 | | |
1415 | | /* Canonicalize zeroes. */ |
1416 | 37.9k | if (omsb == 0) |
1417 | 31.5k | category = fcZero; |
1418 | | |
1419 | | /* The fcZero case is a denormal that underflowed to zero. */ |
1420 | 37.9k | return (opStatus) (opUnderflow | opInexact); |
1421 | 37.9k | } |
1422 | | |
1423 | | APFloat::opStatus |
1424 | | APFloat::addOrSubtractSpecials(const APFloat &rhs, bool subtract) |
1425 | 0 | { |
1426 | 0 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1427 | 0 | default: |
1428 | 0 | llvm_unreachable(nullptr); |
1429 | | |
1430 | 0 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1431 | 0 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1432 | 0 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1433 | 0 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1434 | 0 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1435 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1436 | 0 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1437 | 0 | return opOK; |
1438 | | |
1439 | 0 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1440 | 0 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1441 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1442 | | // We need to be sure to flip the sign here for subtraction because we |
1443 | | // don't have a separate negate operation so -NaN becomes 0 - NaN here. |
1444 | 0 | sign = rhs.sign ^ subtract; |
1445 | 0 | category = fcNaN; |
1446 | 0 | copySignificand(rhs); |
1447 | 0 | return opOK; |
1448 | | |
1449 | 0 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1450 | 0 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1451 | 0 | category = fcInfinity; |
1452 | 0 | sign = rhs.sign ^ subtract; |
1453 | 0 | return opOK; |
1454 | | |
1455 | 0 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1456 | 0 | assign(rhs); |
1457 | 0 | sign = rhs.sign ^ subtract; |
1458 | 0 | return opOK; |
1459 | | |
1460 | 0 | case PackCategoriesIntoKey(fcZero, fcZero): |
1461 | | /* Sign depends on rounding mode; handled by caller. */ |
1462 | 0 | return opOK; |
1463 | | |
1464 | 0 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1465 | | /* Differently signed infinities can only be validly |
1466 | | subtracted. */ |
1467 | 0 | if (((sign ^ rhs.sign)!=0) != subtract) { |
1468 | 0 | makeNaN(); |
1469 | 0 | return APFloat::opInvalidOp; |
1470 | 0 | } |
1471 | | |
1472 | 0 | return opOK; |
1473 | | |
1474 | 0 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1475 | 0 | return opDivByZero; |
1476 | 0 | } |
1477 | 0 | } |
1478 | | |
1479 | | /* Add or subtract two normal numbers. */ |
1480 | | lostFraction |
1481 | | APFloat::addOrSubtractSignificand(const APFloat &rhs, bool subtract) |
1482 | 0 | { |
1483 | 0 | integerPart carry; |
1484 | 0 | lostFraction lost_fraction; |
1485 | 0 | int bits; |
1486 | | |
1487 | | /* Determine if the operation on the absolute values is effectively |
1488 | | an addition or subtraction. */ |
1489 | 0 | subtract ^= static_cast<bool>(sign ^ rhs.sign); |
1490 | | |
1491 | | /* Are we bigger exponent-wise than the RHS? */ |
1492 | 0 | bits = exponent - rhs.exponent; |
1493 | | |
1494 | | /* Subtraction is more subtle than one might naively expect. */ |
1495 | 0 | if (subtract) { |
1496 | 0 | APFloat temp_rhs(rhs); |
1497 | 0 | bool reverse; |
1498 | |
|
1499 | 0 | if (bits == 0) { |
1500 | 0 | reverse = compareAbsoluteValue(temp_rhs) == cmpLessThan; |
1501 | 0 | lost_fraction = lfExactlyZero; |
1502 | 0 | } else if (bits > 0) { |
1503 | 0 | lost_fraction = temp_rhs.shiftSignificandRight(bits - 1); |
1504 | 0 | shiftSignificandLeft(1); |
1505 | 0 | reverse = false; |
1506 | 0 | } else { |
1507 | 0 | lost_fraction = shiftSignificandRight(-bits - 1); |
1508 | 0 | temp_rhs.shiftSignificandLeft(1); |
1509 | 0 | reverse = true; |
1510 | 0 | } |
1511 | |
|
1512 | 0 | if (reverse) { |
1513 | 0 | carry = temp_rhs.subtractSignificand |
1514 | 0 | (*this, lost_fraction != lfExactlyZero); |
1515 | 0 | copySignificand(temp_rhs); |
1516 | 0 | sign = !sign; |
1517 | 0 | } else { |
1518 | 0 | carry = subtractSignificand |
1519 | 0 | (temp_rhs, lost_fraction != lfExactlyZero); |
1520 | 0 | } |
1521 | | |
1522 | | /* Invert the lost fraction - it was on the RHS and |
1523 | | subtracted. */ |
1524 | 0 | if (lost_fraction == lfLessThanHalf) |
1525 | 0 | lost_fraction = lfMoreThanHalf; |
1526 | 0 | else if (lost_fraction == lfMoreThanHalf) |
1527 | 0 | lost_fraction = lfLessThanHalf; |
1528 | | |
1529 | | /* The code above is intended to ensure that no borrow is |
1530 | | necessary. */ |
1531 | 0 | assert(!carry); |
1532 | 0 | (void)carry; |
1533 | 0 | } else { |
1534 | 0 | if (bits > 0) { |
1535 | 0 | APFloat temp_rhs(rhs); |
1536 | |
|
1537 | 0 | lost_fraction = temp_rhs.shiftSignificandRight(bits); |
1538 | 0 | carry = addSignificand(temp_rhs); |
1539 | 0 | } else { |
1540 | 0 | lost_fraction = shiftSignificandRight(-bits); |
1541 | 0 | carry = addSignificand(rhs); |
1542 | 0 | } |
1543 | | |
1544 | | /* We have a guard bit; generating a carry cannot happen. */ |
1545 | 0 | assert(!carry); |
1546 | 0 | (void)carry; |
1547 | 0 | } |
1548 | | |
1549 | 0 | return lost_fraction; |
1550 | 0 | } |
1551 | | |
1552 | | APFloat::opStatus |
1553 | | APFloat::multiplySpecials(const APFloat &rhs) |
1554 | 0 | { |
1555 | 0 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1556 | 0 | default: |
1557 | 0 | llvm_unreachable(nullptr); |
1558 | | |
1559 | 0 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1560 | 0 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1561 | 0 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1562 | 0 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1563 | 0 | sign = false; |
1564 | 0 | return opOK; |
1565 | | |
1566 | 0 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1567 | 0 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1568 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1569 | 0 | sign = false; |
1570 | 0 | category = fcNaN; |
1571 | 0 | copySignificand(rhs); |
1572 | 0 | return opOK; |
1573 | | |
1574 | 0 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1575 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1576 | 0 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1577 | 0 | category = fcInfinity; |
1578 | 0 | return opOK; |
1579 | | |
1580 | 0 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1581 | 0 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1582 | 0 | case PackCategoriesIntoKey(fcZero, fcZero): |
1583 | 0 | category = fcZero; |
1584 | 0 | return opOK; |
1585 | | |
1586 | 0 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1587 | 0 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1588 | 0 | makeNaN(); |
1589 | 0 | return opInvalidOp; |
1590 | | |
1591 | 0 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1592 | 0 | return opOK; |
1593 | 0 | } |
1594 | 0 | } |
1595 | | |
1596 | | APFloat::opStatus |
1597 | | APFloat::divideSpecials(const APFloat &rhs) |
1598 | 0 | { |
1599 | 0 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1600 | 0 | default: |
1601 | 0 | llvm_unreachable(nullptr); |
1602 | | |
1603 | 0 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1604 | 0 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1605 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1606 | 0 | category = fcNaN; |
1607 | 0 | copySignificand(rhs); |
1608 | 0 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1609 | 0 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1610 | 0 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1611 | 0 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1612 | 0 | sign = false; |
1613 | 0 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1614 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1615 | 0 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1616 | 0 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1617 | 0 | return opOK; |
1618 | | |
1619 | 0 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1620 | 0 | category = fcZero; |
1621 | 0 | return opOK; |
1622 | | |
1623 | 0 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1624 | 0 | category = fcInfinity; |
1625 | 0 | return opDivByZero; |
1626 | | |
1627 | 0 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1628 | 0 | case PackCategoriesIntoKey(fcZero, fcZero): |
1629 | 0 | makeNaN(); |
1630 | 0 | return opInvalidOp; |
1631 | | |
1632 | 0 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1633 | 0 | return opOK; |
1634 | 0 | } |
1635 | 0 | } |
1636 | | |
1637 | | APFloat::opStatus |
1638 | | APFloat::modSpecials(const APFloat &rhs) |
1639 | 0 | { |
1640 | 0 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1641 | 0 | default: |
1642 | 0 | llvm_unreachable(nullptr); |
1643 | | |
1644 | 0 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1645 | 0 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1646 | 0 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1647 | 0 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1648 | 0 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1649 | 0 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1650 | 0 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1651 | 0 | return opOK; |
1652 | | |
1653 | 0 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1654 | 0 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1655 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1656 | 0 | sign = false; |
1657 | 0 | category = fcNaN; |
1658 | 0 | copySignificand(rhs); |
1659 | 0 | return opOK; |
1660 | | |
1661 | 0 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1662 | 0 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1663 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1664 | 0 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1665 | 0 | case PackCategoriesIntoKey(fcZero, fcZero): |
1666 | 0 | makeNaN(); |
1667 | 0 | return opInvalidOp; |
1668 | | |
1669 | 0 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1670 | 0 | return opOK; |
1671 | 0 | } |
1672 | 0 | } |
1673 | | |
1674 | | /* Change sign. */ |
1675 | | void |
1676 | | APFloat::changeSign() |
1677 | 16.5k | { |
1678 | | /* Look mummy, this one's easy. */ |
1679 | 16.5k | sign = !sign; |
1680 | 16.5k | } |
1681 | | |
1682 | | void |
1683 | | APFloat::clearSign() |
1684 | 0 | { |
1685 | | /* So is this one. */ |
1686 | 0 | sign = 0; |
1687 | 0 | } |
1688 | | |
1689 | | void |
1690 | | APFloat::copySign(const APFloat &rhs) |
1691 | 0 | { |
1692 | | /* And this one. */ |
1693 | 0 | sign = rhs.sign; |
1694 | 0 | } |
1695 | | |
1696 | | /* Normalized addition or subtraction. */ |
1697 | | APFloat::opStatus |
1698 | | APFloat::addOrSubtract(const APFloat &rhs, roundingMode rounding_mode, |
1699 | | bool subtract) |
1700 | 0 | { |
1701 | 0 | opStatus fs; |
1702 | |
|
1703 | 0 | fs = addOrSubtractSpecials(rhs, subtract); |
1704 | | |
1705 | | /* This return code means it was not a simple case. */ |
1706 | 0 | if (fs == opDivByZero) { |
1707 | 0 | lostFraction lost_fraction; |
1708 | |
|
1709 | 0 | lost_fraction = addOrSubtractSignificand(rhs, subtract); |
1710 | 0 | fs = normalize(rounding_mode, lost_fraction); |
1711 | | |
1712 | | /* Can only be zero if we lost no fraction. */ |
1713 | 0 | assert(category != fcZero || lost_fraction == lfExactlyZero); |
1714 | 0 | } |
1715 | | |
1716 | | /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a |
1717 | | positive zero unless rounding to minus infinity, except that |
1718 | | adding two like-signed zeroes gives that zero. */ |
1719 | 0 | if (category == fcZero) { |
1720 | 0 | if (rhs.category != fcZero || (sign == rhs.sign) == subtract) |
1721 | 0 | sign = (rounding_mode == rmTowardNegative); |
1722 | 0 | } |
1723 | |
|
1724 | 0 | return fs; |
1725 | 0 | } |
1726 | | |
1727 | | /* Normalized addition. */ |
1728 | | APFloat::opStatus |
1729 | | APFloat::add(const APFloat &rhs, roundingMode rounding_mode) |
1730 | 0 | { |
1731 | 0 | return addOrSubtract(rhs, rounding_mode, false); |
1732 | 0 | } |
1733 | | |
1734 | | /* Normalized subtraction. */ |
1735 | | APFloat::opStatus |
1736 | | APFloat::subtract(const APFloat &rhs, roundingMode rounding_mode) |
1737 | 0 | { |
1738 | 0 | return addOrSubtract(rhs, rounding_mode, true); |
1739 | 0 | } |
1740 | | |
1741 | | /* Normalized multiply. */ |
1742 | | APFloat::opStatus |
1743 | | APFloat::multiply(const APFloat &rhs, roundingMode rounding_mode) |
1744 | 0 | { |
1745 | 0 | opStatus fs; |
1746 | |
|
1747 | 0 | sign ^= rhs.sign; |
1748 | 0 | fs = multiplySpecials(rhs); |
1749 | |
|
1750 | 0 | if (isFiniteNonZero()) { |
1751 | 0 | lostFraction lost_fraction = multiplySignificand(rhs, nullptr); |
1752 | 0 | fs = normalize(rounding_mode, lost_fraction); |
1753 | 0 | if (lost_fraction != lfExactlyZero) |
1754 | 0 | fs = (opStatus) (fs | opInexact); |
1755 | 0 | } |
1756 | |
|
1757 | 0 | return fs; |
1758 | 0 | } |
1759 | | |
1760 | | /* Normalized divide. */ |
1761 | | APFloat::opStatus |
1762 | | APFloat::divide(const APFloat &rhs, roundingMode rounding_mode) |
1763 | 0 | { |
1764 | 0 | opStatus fs; |
1765 | |
|
1766 | 0 | sign ^= rhs.sign; |
1767 | 0 | fs = divideSpecials(rhs); |
1768 | |
|
1769 | 0 | if (isFiniteNonZero()) { |
1770 | 0 | lostFraction lost_fraction = divideSignificand(rhs); |
1771 | 0 | fs = normalize(rounding_mode, lost_fraction); |
1772 | 0 | if (lost_fraction != lfExactlyZero) |
1773 | 0 | fs = (opStatus) (fs | opInexact); |
1774 | 0 | } |
1775 | |
|
1776 | 0 | return fs; |
1777 | 0 | } |
1778 | | |
1779 | | /* Normalized remainder. This is not currently correct in all cases. */ |
1780 | | APFloat::opStatus |
1781 | | APFloat::remainder(const APFloat &rhs) |
1782 | 0 | { |
1783 | 0 | opStatus fs; |
1784 | 0 | APFloat V = *this; |
1785 | 0 | unsigned int origSign = sign; |
1786 | |
|
1787 | 0 | fs = V.divide(rhs, rmNearestTiesToEven); |
1788 | 0 | if (fs == opDivByZero) |
1789 | 0 | return fs; |
1790 | | |
1791 | 0 | int parts = partCount(); |
1792 | 0 | integerPart *x = new integerPart[parts]; |
1793 | 0 | bool ignored; |
1794 | 0 | fs = V.convertToInteger(x, parts * integerPartWidth, true, |
1795 | 0 | rmNearestTiesToEven, &ignored); |
1796 | 0 | if (fs==opInvalidOp) |
1797 | 0 | return fs; |
1798 | | |
1799 | 0 | fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true, |
1800 | 0 | rmNearestTiesToEven); |
1801 | 0 | assert(fs==opOK); // should always work |
1802 | | |
1803 | 0 | fs = V.multiply(rhs, rmNearestTiesToEven); |
1804 | 0 | assert(fs==opOK || fs==opInexact); // should not overflow or underflow |
1805 | | |
1806 | 0 | fs = subtract(V, rmNearestTiesToEven); |
1807 | 0 | assert(fs==opOK || fs==opInexact); // likewise |
1808 | | |
1809 | 0 | if (isZero()) |
1810 | 0 | sign = origSign; // IEEE754 requires this |
1811 | 0 | delete[] x; |
1812 | 0 | return fs; |
1813 | 0 | } |
1814 | | |
1815 | | /* Normalized llvm frem (C fmod). |
1816 | | This is not currently correct in all cases. */ |
1817 | | APFloat::opStatus |
1818 | | APFloat::mod(const APFloat &rhs) |
1819 | 0 | { |
1820 | 0 | opStatus fs; |
1821 | 0 | fs = modSpecials(rhs); |
1822 | |
|
1823 | 0 | if (isFiniteNonZero() && rhs.isFiniteNonZero()) { |
1824 | 0 | APFloat V = *this; |
1825 | 0 | unsigned int origSign = sign; |
1826 | |
|
1827 | 0 | fs = V.divide(rhs, rmNearestTiesToEven); |
1828 | 0 | if (fs == opDivByZero) |
1829 | 0 | return fs; |
1830 | | |
1831 | 0 | int parts = partCount(); |
1832 | 0 | integerPart *x = new integerPart[parts]; |
1833 | 0 | bool ignored; |
1834 | 0 | fs = V.convertToInteger(x, parts * integerPartWidth, true, |
1835 | 0 | rmTowardZero, &ignored); |
1836 | 0 | if (fs==opInvalidOp) |
1837 | 0 | return fs; |
1838 | | |
1839 | 0 | fs = V.convertFromZeroExtendedInteger(x, parts * integerPartWidth, true, |
1840 | 0 | rmNearestTiesToEven); |
1841 | 0 | assert(fs==opOK); // should always work |
1842 | | |
1843 | 0 | fs = V.multiply(rhs, rmNearestTiesToEven); |
1844 | 0 | assert(fs==opOK || fs==opInexact); // should not overflow or underflow |
1845 | | |
1846 | 0 | fs = subtract(V, rmNearestTiesToEven); |
1847 | 0 | assert(fs==opOK || fs==opInexact); // likewise |
1848 | | |
1849 | 0 | if (isZero()) |
1850 | 0 | sign = origSign; // IEEE754 requires this |
1851 | 0 | delete[] x; |
1852 | 0 | } |
1853 | 0 | return fs; |
1854 | 0 | } |
1855 | | |
1856 | | /* Normalized fused-multiply-add. */ |
1857 | | APFloat::opStatus |
1858 | | APFloat::fusedMultiplyAdd(const APFloat &multiplicand, |
1859 | | const APFloat &addend, |
1860 | | roundingMode rounding_mode) |
1861 | 0 | { |
1862 | 0 | opStatus fs; |
1863 | | |
1864 | | /* Post-multiplication sign, before addition. */ |
1865 | 0 | sign ^= multiplicand.sign; |
1866 | | |
1867 | | /* If and only if all arguments are normal do we need to do an |
1868 | | extended-precision calculation. */ |
1869 | 0 | if (isFiniteNonZero() && |
1870 | 0 | multiplicand.isFiniteNonZero() && |
1871 | 0 | addend.isFinite()) { |
1872 | 0 | lostFraction lost_fraction; |
1873 | |
|
1874 | 0 | lost_fraction = multiplySignificand(multiplicand, &addend); |
1875 | 0 | fs = normalize(rounding_mode, lost_fraction); |
1876 | 0 | if (lost_fraction != lfExactlyZero) |
1877 | 0 | fs = (opStatus) (fs | opInexact); |
1878 | | |
1879 | | /* If two numbers add (exactly) to zero, IEEE 754 decrees it is a |
1880 | | positive zero unless rounding to minus infinity, except that |
1881 | | adding two like-signed zeroes gives that zero. */ |
1882 | 0 | if (category == fcZero && !(fs & opUnderflow) && sign != addend.sign) |
1883 | 0 | sign = (rounding_mode == rmTowardNegative); |
1884 | 0 | } else { |
1885 | 0 | fs = multiplySpecials(multiplicand); |
1886 | | |
1887 | | /* FS can only be opOK or opInvalidOp. There is no more work |
1888 | | to do in the latter case. The IEEE-754R standard says it is |
1889 | | implementation-defined in this case whether, if ADDEND is a |
1890 | | quiet NaN, we raise invalid op; this implementation does so. |
1891 | | |
1892 | | If we need to do the addition we can do so with normal |
1893 | | precision. */ |
1894 | 0 | if (fs == opOK) |
1895 | 0 | fs = addOrSubtract(addend, rounding_mode, false); |
1896 | 0 | } |
1897 | |
|
1898 | 0 | return fs; |
1899 | 0 | } |
1900 | | |
1901 | | /* Rounding-mode corrrect round to integral value. */ |
1902 | 0 | APFloat::opStatus APFloat::roundToIntegral(roundingMode rounding_mode) { |
1903 | 0 | opStatus fs; |
1904 | | |
1905 | | // If the exponent is large enough, we know that this value is already |
1906 | | // integral, and the arithmetic below would potentially cause it to saturate |
1907 | | // to +/-Inf. Bail out early instead. |
1908 | 0 | if (isFiniteNonZero() && exponent+1 >= (int)semanticsPrecision(*semantics)) |
1909 | 0 | return opOK; |
1910 | | |
1911 | | // The algorithm here is quite simple: we add 2^(p-1), where p is the |
1912 | | // precision of our format, and then subtract it back off again. The choice |
1913 | | // of rounding modes for the addition/subtraction determines the rounding mode |
1914 | | // for our integral rounding as well. |
1915 | | // NOTE: When the input value is negative, we do subtraction followed by |
1916 | | // addition instead. |
1917 | 0 | APInt IntegerConstant(NextPowerOf2(semanticsPrecision(*semantics)), 1); |
1918 | 0 | IntegerConstant <<= semanticsPrecision(*semantics)-1; |
1919 | 0 | APFloat MagicConstant(*semantics); |
1920 | 0 | fs = MagicConstant.convertFromAPInt(IntegerConstant, false, |
1921 | 0 | rmNearestTiesToEven); |
1922 | 0 | MagicConstant.copySign(*this); |
1923 | |
|
1924 | 0 | if (fs != opOK) |
1925 | 0 | return fs; |
1926 | | |
1927 | | // Preserve the input sign so that we can handle 0.0/-0.0 cases correctly. |
1928 | 0 | bool inputSign = isNegative(); |
1929 | |
|
1930 | 0 | fs = add(MagicConstant, rounding_mode); |
1931 | 0 | if (fs != opOK && fs != opInexact) |
1932 | 0 | return fs; |
1933 | | |
1934 | 0 | fs = subtract(MagicConstant, rounding_mode); |
1935 | | |
1936 | | // Restore the input sign. |
1937 | 0 | if (inputSign != isNegative()) |
1938 | 0 | changeSign(); |
1939 | |
|
1940 | 0 | return fs; |
1941 | 0 | } |
1942 | | |
1943 | | |
1944 | | /* Comparison requires normalized numbers. */ |
1945 | | APFloat::cmpResult |
1946 | | APFloat::compare(const APFloat &rhs) const |
1947 | 0 | { |
1948 | 0 | cmpResult result; |
1949 | |
|
1950 | 0 | assert(semantics == rhs.semantics); |
1951 | | |
1952 | 0 | switch (PackCategoriesIntoKey(category, rhs.category)) { |
1953 | 0 | default: |
1954 | 0 | llvm_unreachable(nullptr); |
1955 | | |
1956 | 0 | case PackCategoriesIntoKey(fcNaN, fcZero): |
1957 | 0 | case PackCategoriesIntoKey(fcNaN, fcNormal): |
1958 | 0 | case PackCategoriesIntoKey(fcNaN, fcInfinity): |
1959 | 0 | case PackCategoriesIntoKey(fcNaN, fcNaN): |
1960 | 0 | case PackCategoriesIntoKey(fcZero, fcNaN): |
1961 | 0 | case PackCategoriesIntoKey(fcNormal, fcNaN): |
1962 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNaN): |
1963 | 0 | return cmpUnordered; |
1964 | | |
1965 | 0 | case PackCategoriesIntoKey(fcInfinity, fcNormal): |
1966 | 0 | case PackCategoriesIntoKey(fcInfinity, fcZero): |
1967 | 0 | case PackCategoriesIntoKey(fcNormal, fcZero): |
1968 | 0 | if (sign) |
1969 | 0 | return cmpLessThan; |
1970 | 0 | else |
1971 | 0 | return cmpGreaterThan; |
1972 | | |
1973 | 0 | case PackCategoriesIntoKey(fcNormal, fcInfinity): |
1974 | 0 | case PackCategoriesIntoKey(fcZero, fcInfinity): |
1975 | 0 | case PackCategoriesIntoKey(fcZero, fcNormal): |
1976 | 0 | if (rhs.sign) |
1977 | 0 | return cmpGreaterThan; |
1978 | 0 | else |
1979 | 0 | return cmpLessThan; |
1980 | | |
1981 | 0 | case PackCategoriesIntoKey(fcInfinity, fcInfinity): |
1982 | 0 | if (sign == rhs.sign) |
1983 | 0 | return cmpEqual; |
1984 | 0 | else if (sign) |
1985 | 0 | return cmpLessThan; |
1986 | 0 | else |
1987 | 0 | return cmpGreaterThan; |
1988 | | |
1989 | 0 | case PackCategoriesIntoKey(fcZero, fcZero): |
1990 | 0 | return cmpEqual; |
1991 | | |
1992 | 0 | case PackCategoriesIntoKey(fcNormal, fcNormal): |
1993 | 0 | break; |
1994 | 0 | } |
1995 | | |
1996 | | /* Two normal numbers. Do they have the same sign? */ |
1997 | 0 | if (sign != rhs.sign) { |
1998 | 0 | if (sign) |
1999 | 0 | result = cmpLessThan; |
2000 | 0 | else |
2001 | 0 | result = cmpGreaterThan; |
2002 | 0 | } else { |
2003 | | /* Compare absolute values; invert result if negative. */ |
2004 | 0 | result = compareAbsoluteValue(rhs); |
2005 | |
|
2006 | 0 | if (sign) { |
2007 | 0 | if (result == cmpLessThan) |
2008 | 0 | result = cmpGreaterThan; |
2009 | 0 | else if (result == cmpGreaterThan) |
2010 | 0 | result = cmpLessThan; |
2011 | 0 | } |
2012 | 0 | } |
2013 | |
|
2014 | 0 | return result; |
2015 | 0 | } |
2016 | | |
2017 | | /// APFloat::convert - convert a value of one floating point type to another. |
2018 | | /// The return value corresponds to the IEEE754 exceptions. *losesInfo |
2019 | | /// records whether the transformation lost information, i.e. whether |
2020 | | /// converting the result back to the original type will produce the |
2021 | | /// original value (this is almost the same as return value==fsOK, but there |
2022 | | /// are edge cases where this is not so). |
2023 | | |
2024 | | APFloat::opStatus |
2025 | | APFloat::convert(const fltSemantics &toSemantics, |
2026 | | roundingMode rounding_mode, bool *losesInfo) |
2027 | 0 | { |
2028 | 0 | lostFraction lostFraction; |
2029 | 0 | unsigned int newPartCount, oldPartCount; |
2030 | 0 | opStatus fs; |
2031 | 0 | int shift; |
2032 | 0 | const fltSemantics &fromSemantics = *semantics; |
2033 | |
|
2034 | 0 | lostFraction = lfExactlyZero; |
2035 | 0 | newPartCount = partCountForBits(toSemantics.precision + 1); |
2036 | 0 | oldPartCount = partCount(); |
2037 | 0 | shift = toSemantics.precision - fromSemantics.precision; |
2038 | |
|
2039 | 0 | bool X86SpecialNan = false; |
2040 | 0 | if (&fromSemantics == &APFloat::x87DoubleExtended && |
2041 | 0 | &toSemantics != &APFloat::x87DoubleExtended && category == fcNaN && |
2042 | 0 | (!(*significandParts() & 0x8000000000000000ULL) || |
2043 | 0 | !(*significandParts() & 0x4000000000000000ULL))) { |
2044 | | // x86 has some unusual NaNs which cannot be represented in any other |
2045 | | // format; note them here. |
2046 | 0 | X86SpecialNan = true; |
2047 | 0 | } |
2048 | | |
2049 | | // If this is a truncation of a denormal number, and the target semantics |
2050 | | // has larger exponent range than the source semantics (this can happen |
2051 | | // when truncating from PowerPC double-double to double format), the |
2052 | | // right shift could lose result mantissa bits. Adjust exponent instead |
2053 | | // of performing excessive shift. |
2054 | 0 | if (shift < 0 && isFiniteNonZero()) { |
2055 | 0 | int exponentChange = significandMSB() + 1 - fromSemantics.precision; |
2056 | 0 | if (exponent + exponentChange < toSemantics.minExponent) |
2057 | 0 | exponentChange = toSemantics.minExponent - exponent; |
2058 | 0 | if (exponentChange < shift) |
2059 | 0 | exponentChange = shift; |
2060 | 0 | if (exponentChange < 0) { |
2061 | 0 | shift -= exponentChange; |
2062 | 0 | exponent += exponentChange; |
2063 | 0 | } |
2064 | 0 | } |
2065 | | |
2066 | | // If this is a truncation, perform the shift before we narrow the storage. |
2067 | 0 | if (shift < 0 && (isFiniteNonZero() || category==fcNaN)) |
2068 | 0 | lostFraction = shiftRight(significandParts(), oldPartCount, -shift); |
2069 | | |
2070 | | // Fix the storage so it can hold to new value. |
2071 | 0 | if (newPartCount > oldPartCount) { |
2072 | | // The new type requires more storage; make it available. |
2073 | 0 | integerPart *newParts; |
2074 | 0 | newParts = new integerPart[newPartCount]; |
2075 | 0 | APInt::tcSet(newParts, 0, newPartCount); |
2076 | 0 | if (isFiniteNonZero() || category==fcNaN) |
2077 | 0 | APInt::tcAssign(newParts, significandParts(), oldPartCount); |
2078 | 0 | freeSignificand(); |
2079 | 0 | significand.parts = newParts; |
2080 | 0 | } else if (newPartCount == 1 && oldPartCount != 1) { |
2081 | | // Switch to built-in storage for a single part. |
2082 | 0 | integerPart newPart = 0; |
2083 | 0 | if (isFiniteNonZero() || category==fcNaN) |
2084 | 0 | newPart = significandParts()[0]; |
2085 | 0 | freeSignificand(); |
2086 | 0 | significand.part = newPart; |
2087 | 0 | } |
2088 | | |
2089 | | // Now that we have the right storage, switch the semantics. |
2090 | 0 | semantics = &toSemantics; |
2091 | | |
2092 | | // If this is an extension, perform the shift now that the storage is |
2093 | | // available. |
2094 | 0 | if (shift > 0 && (isFiniteNonZero() || category==fcNaN)) |
2095 | 0 | APInt::tcShiftLeft(significandParts(), newPartCount, shift); |
2096 | |
|
2097 | 0 | if (isFiniteNonZero()) { |
2098 | 0 | fs = normalize(rounding_mode, lostFraction); |
2099 | 0 | *losesInfo = (fs != opOK); |
2100 | 0 | } else if (category == fcNaN) { |
2101 | 0 | *losesInfo = lostFraction != lfExactlyZero || X86SpecialNan; |
2102 | | |
2103 | | // For x87 extended precision, we want to make a NaN, not a special NaN if |
2104 | | // the input wasn't special either. |
2105 | 0 | if (!X86SpecialNan && semantics == &APFloat::x87DoubleExtended) |
2106 | 0 | APInt::tcSetBit(significandParts(), semantics->precision - 1); |
2107 | | |
2108 | | // gcc forces the Quiet bit on, which means (float)(double)(float_sNan) |
2109 | | // does not give you back the same bits. This is dubious, and we |
2110 | | // don't currently do it. You're really supposed to get |
2111 | | // an invalid operation signal at runtime, but nobody does that. |
2112 | 0 | fs = opOK; |
2113 | 0 | } else { |
2114 | 0 | *losesInfo = false; |
2115 | 0 | fs = opOK; |
2116 | 0 | } |
2117 | |
|
2118 | 0 | return fs; |
2119 | 0 | } |
2120 | | |
2121 | | /* Convert a floating point number to an integer according to the |
2122 | | rounding mode. If the rounded integer value is out of range this |
2123 | | returns an invalid operation exception and the contents of the |
2124 | | destination parts are unspecified. If the rounded value is in |
2125 | | range but the floating point number is not the exact integer, the C |
2126 | | standard doesn't require an inexact exception to be raised. IEEE |
2127 | | 854 does require it so we do that. |
2128 | | |
2129 | | Note that for conversions to integer type the C standard requires |
2130 | | round-to-zero to always be used. */ |
2131 | | APFloat::opStatus |
2132 | | APFloat::convertToSignExtendedInteger(integerPart *parts, unsigned int width, |
2133 | | bool isSigned, |
2134 | | roundingMode rounding_mode, |
2135 | | bool *isExact) const |
2136 | 0 | { |
2137 | 0 | lostFraction lost_fraction; |
2138 | 0 | const integerPart *src; |
2139 | 0 | unsigned int dstPartsCount, truncatedBits; |
2140 | |
|
2141 | 0 | *isExact = false; |
2142 | | |
2143 | | /* Handle the three special cases first. */ |
2144 | 0 | if (category == fcInfinity || category == fcNaN) |
2145 | 0 | return opInvalidOp; |
2146 | | |
2147 | 0 | dstPartsCount = partCountForBits(width); |
2148 | |
|
2149 | 0 | if (category == fcZero) { |
2150 | 0 | APInt::tcSet(parts, 0, dstPartsCount); |
2151 | | // Negative zero can't be represented as an int. |
2152 | 0 | *isExact = !sign; |
2153 | 0 | return opOK; |
2154 | 0 | } |
2155 | | |
2156 | 0 | src = significandParts(); |
2157 | | |
2158 | | /* Step 1: place our absolute value, with any fraction truncated, in |
2159 | | the destination. */ |
2160 | 0 | if (exponent < 0) { |
2161 | | /* Our absolute value is less than one; truncate everything. */ |
2162 | 0 | APInt::tcSet(parts, 0, dstPartsCount); |
2163 | | /* For exponent -1 the integer bit represents .5, look at that. |
2164 | | For smaller exponents leftmost truncated bit is 0. */ |
2165 | 0 | truncatedBits = semantics->precision -1U - exponent; |
2166 | 0 | } else { |
2167 | | /* We want the most significant (exponent + 1) bits; the rest are |
2168 | | truncated. */ |
2169 | 0 | unsigned int bits = exponent + 1U; |
2170 | | |
2171 | | /* Hopelessly large in magnitude? */ |
2172 | 0 | if (bits > width) |
2173 | 0 | return opInvalidOp; |
2174 | | |
2175 | 0 | if (bits < semantics->precision) { |
2176 | | /* We truncate (semantics->precision - bits) bits. */ |
2177 | 0 | truncatedBits = semantics->precision - bits; |
2178 | 0 | APInt::tcExtract(parts, dstPartsCount, src, bits, truncatedBits); |
2179 | 0 | } else { |
2180 | | /* We want at least as many bits as are available. */ |
2181 | 0 | APInt::tcExtract(parts, dstPartsCount, src, semantics->precision, 0); |
2182 | 0 | APInt::tcShiftLeft(parts, dstPartsCount, bits - semantics->precision); |
2183 | 0 | truncatedBits = 0; |
2184 | 0 | } |
2185 | 0 | } |
2186 | | |
2187 | | /* Step 2: work out any lost fraction, and increment the absolute |
2188 | | value if we would round away from zero. */ |
2189 | 0 | if (truncatedBits) { |
2190 | 0 | lost_fraction = lostFractionThroughTruncation(src, partCount(), |
2191 | 0 | truncatedBits); |
2192 | 0 | if (lost_fraction != lfExactlyZero && |
2193 | 0 | roundAwayFromZero(rounding_mode, lost_fraction, truncatedBits)) { |
2194 | 0 | if (APInt::tcIncrement(parts, dstPartsCount)) |
2195 | 0 | return opInvalidOp; /* Overflow. */ |
2196 | 0 | } |
2197 | 0 | } else { |
2198 | 0 | lost_fraction = lfExactlyZero; |
2199 | 0 | } |
2200 | | |
2201 | | /* Step 3: check if we fit in the destination. */ |
2202 | 0 | unsigned int omsb = APInt::tcMSB(parts, dstPartsCount) + 1; |
2203 | |
|
2204 | 0 | if (sign) { |
2205 | 0 | if (!isSigned) { |
2206 | | /* Negative numbers cannot be represented as unsigned. */ |
2207 | 0 | if (omsb != 0) |
2208 | 0 | return opInvalidOp; |
2209 | 0 | } else { |
2210 | | /* It takes omsb bits to represent the unsigned integer value. |
2211 | | We lose a bit for the sign, but care is needed as the |
2212 | | maximally negative integer is a special case. */ |
2213 | 0 | if (omsb == width && APInt::tcLSB(parts, dstPartsCount) + 1 != omsb) |
2214 | 0 | return opInvalidOp; |
2215 | | |
2216 | | /* This case can happen because of rounding. */ |
2217 | 0 | if (omsb > width) |
2218 | 0 | return opInvalidOp; |
2219 | 0 | } |
2220 | | |
2221 | 0 | APInt::tcNegate (parts, dstPartsCount); |
2222 | 0 | } else { |
2223 | 0 | if (omsb >= width + !isSigned) |
2224 | 0 | return opInvalidOp; |
2225 | 0 | } |
2226 | | |
2227 | 0 | if (lost_fraction == lfExactlyZero) { |
2228 | 0 | *isExact = true; |
2229 | 0 | return opOK; |
2230 | 0 | } else |
2231 | 0 | return opInexact; |
2232 | 0 | } |
2233 | | |
2234 | | /* Same as convertToSignExtendedInteger, except we provide |
2235 | | deterministic values in case of an invalid operation exception, |
2236 | | namely zero for NaNs and the minimal or maximal value respectively |
2237 | | for underflow or overflow. |
2238 | | The *isExact output tells whether the result is exact, in the sense |
2239 | | that converting it back to the original floating point type produces |
2240 | | the original value. This is almost equivalent to result==opOK, |
2241 | | except for negative zeroes. |
2242 | | */ |
2243 | | APFloat::opStatus |
2244 | | APFloat::convertToInteger(integerPart *parts, unsigned int width, |
2245 | | bool isSigned, |
2246 | | roundingMode rounding_mode, bool *isExact) const |
2247 | 0 | { |
2248 | 0 | opStatus fs; |
2249 | |
|
2250 | 0 | fs = convertToSignExtendedInteger(parts, width, isSigned, rounding_mode, |
2251 | 0 | isExact); |
2252 | |
|
2253 | 0 | if (fs == opInvalidOp) { |
2254 | 0 | unsigned int bits, dstPartsCount; |
2255 | |
|
2256 | 0 | dstPartsCount = partCountForBits(width); |
2257 | |
|
2258 | 0 | if (category == fcNaN) |
2259 | 0 | bits = 0; |
2260 | 0 | else if (sign) |
2261 | 0 | bits = isSigned; |
2262 | 0 | else |
2263 | 0 | bits = width - isSigned; |
2264 | |
|
2265 | 0 | APInt::tcSetLeastSignificantBits(parts, dstPartsCount, bits); |
2266 | 0 | if (sign && isSigned) |
2267 | 0 | APInt::tcShiftLeft(parts, dstPartsCount, width - 1); |
2268 | 0 | } |
2269 | |
|
2270 | 0 | return fs; |
2271 | 0 | } |
2272 | | |
2273 | | /* Same as convertToInteger(integerPart*, ...), except the result is returned in |
2274 | | an APSInt, whose initial bit-width and signed-ness are used to determine the |
2275 | | precision of the conversion. |
2276 | | */ |
2277 | | APFloat::opStatus |
2278 | | APFloat::convertToInteger(APSInt &result, |
2279 | | roundingMode rounding_mode, bool *isExact) const |
2280 | 0 | { |
2281 | 0 | unsigned bitWidth = result.getBitWidth(); |
2282 | 0 | SmallVector<uint64_t, 4> parts(result.getNumWords()); |
2283 | 0 | opStatus status = convertToInteger( |
2284 | 0 | parts.data(), bitWidth, result.isSigned(), rounding_mode, isExact); |
2285 | | // Keeps the original signed-ness. |
2286 | 0 | result = APInt(bitWidth, parts); |
2287 | 0 | return status; |
2288 | 0 | } |
2289 | | |
2290 | | /* Convert an unsigned integer SRC to a floating point number, |
2291 | | rounding according to ROUNDING_MODE. The sign of the floating |
2292 | | point number is not modified. */ |
2293 | | APFloat::opStatus |
2294 | | APFloat::convertFromUnsignedParts(const integerPart *src, |
2295 | | unsigned int srcCount, |
2296 | | roundingMode rounding_mode) |
2297 | 1.53M | { |
2298 | 1.53M | unsigned int omsb, precision, dstCount; |
2299 | 1.53M | integerPart *dst; |
2300 | 1.53M | lostFraction lost_fraction; |
2301 | | |
2302 | 1.53M | category = fcNormal; |
2303 | 1.53M | omsb = APInt::tcMSB(src, srcCount) + 1; |
2304 | 1.53M | dst = significandParts(); |
2305 | 1.53M | dstCount = partCount(); |
2306 | 1.53M | precision = semantics->precision; |
2307 | | |
2308 | | /* We want the most significant PRECISION bits of SRC. There may not |
2309 | | be that many; extract what we can. */ |
2310 | 1.53M | if (precision <= omsb) { |
2311 | 196k | exponent = omsb - 1; |
2312 | 196k | lost_fraction = lostFractionThroughTruncation(src, srcCount, |
2313 | 196k | omsb - precision); |
2314 | 196k | APInt::tcExtract(dst, dstCount, src, precision, omsb - precision); |
2315 | 1.33M | } else { |
2316 | 1.33M | exponent = precision - 1; |
2317 | 1.33M | lost_fraction = lfExactlyZero; |
2318 | 1.33M | APInt::tcExtract(dst, dstCount, src, omsb, 0); |
2319 | 1.33M | } |
2320 | | |
2321 | 1.53M | return normalize(rounding_mode, lost_fraction); |
2322 | 1.53M | } |
2323 | | |
2324 | | APFloat::opStatus |
2325 | | APFloat::convertFromAPInt(const APInt &Val, |
2326 | | bool isSigned, |
2327 | | roundingMode rounding_mode) |
2328 | 0 | { |
2329 | 0 | unsigned int partCount = Val.getNumWords(); |
2330 | 0 | APInt api = Val; |
2331 | |
|
2332 | 0 | sign = false; |
2333 | 0 | if (isSigned && api.isNegative()) { |
2334 | 0 | sign = true; |
2335 | 0 | api = -api; |
2336 | 0 | } |
2337 | |
|
2338 | 0 | return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode); |
2339 | 0 | } |
2340 | | |
2341 | | /* Convert a two's complement integer SRC to a floating point number, |
2342 | | rounding according to ROUNDING_MODE. ISSIGNED is true if the |
2343 | | integer is signed, in which case it must be sign-extended. */ |
2344 | | APFloat::opStatus |
2345 | | APFloat::convertFromSignExtendedInteger(const integerPart *src, |
2346 | | unsigned int srcCount, |
2347 | | bool isSigned, |
2348 | | roundingMode rounding_mode) |
2349 | 0 | { |
2350 | 0 | opStatus status; |
2351 | |
|
2352 | 0 | if (isSigned && |
2353 | 0 | APInt::tcExtractBit(src, srcCount * integerPartWidth - 1)) { |
2354 | 0 | integerPart *copy; |
2355 | | |
2356 | | /* If we're signed and negative negate a copy. */ |
2357 | 0 | sign = true; |
2358 | 0 | copy = new integerPart[srcCount]; |
2359 | 0 | APInt::tcAssign(copy, src, srcCount); |
2360 | 0 | APInt::tcNegate(copy, srcCount); |
2361 | 0 | status = convertFromUnsignedParts(copy, srcCount, rounding_mode); |
2362 | 0 | delete [] copy; |
2363 | 0 | } else { |
2364 | 0 | sign = false; |
2365 | 0 | status = convertFromUnsignedParts(src, srcCount, rounding_mode); |
2366 | 0 | } |
2367 | |
|
2368 | 0 | return status; |
2369 | 0 | } |
2370 | | |
2371 | | /* FIXME: should this just take a const APInt reference? */ |
2372 | | APFloat::opStatus |
2373 | | APFloat::convertFromZeroExtendedInteger(const integerPart *parts, |
2374 | | unsigned int width, bool isSigned, |
2375 | | roundingMode rounding_mode) |
2376 | 0 | { |
2377 | 0 | unsigned int partCount = partCountForBits(width); |
2378 | 0 | APInt api = APInt(width, makeArrayRef(parts, partCount)); |
2379 | |
|
2380 | 0 | sign = false; |
2381 | 0 | if (isSigned && APInt::tcExtractBit(parts, width - 1)) { |
2382 | 0 | sign = true; |
2383 | 0 | api = -api; |
2384 | 0 | } |
2385 | |
|
2386 | 0 | return convertFromUnsignedParts(api.getRawData(), partCount, rounding_mode); |
2387 | 0 | } |
2388 | | |
2389 | | APFloat::opStatus |
2390 | | APFloat::convertFromHexadecimalString(StringRef s, roundingMode rounding_mode) |
2391 | 131k | { |
2392 | 131k | lostFraction lost_fraction = lfExactlyZero; |
2393 | | |
2394 | 131k | category = fcNormal; |
2395 | 131k | zeroSignificand(); |
2396 | 131k | exponent = 0; |
2397 | | |
2398 | 131k | integerPart *significand = significandParts(); |
2399 | 131k | unsigned partsCount = partCount(); |
2400 | 131k | unsigned bitPos = partsCount * integerPartWidth; |
2401 | 131k | bool computedTrailingFraction = false; |
2402 | | |
2403 | | // Skip leading zeroes and any (hexa)decimal point. |
2404 | 131k | StringRef::iterator begin = s.begin(); |
2405 | 131k | StringRef::iterator end = s.end(); |
2406 | 131k | StringRef::iterator dot; |
2407 | 131k | StringRef::iterator p = skipLeadingZeroesAndAnyDot(begin, end, &dot); |
2408 | 131k | StringRef::iterator firstSignificantDigit = p; |
2409 | | |
2410 | 1.11M | while (p != end) { |
2411 | 1.11M | integerPart hex_value; |
2412 | | |
2413 | 1.11M | if (*p == '.') { |
2414 | 7.66k | assert(dot == end && "String contains multiple dots"); |
2415 | 7.66k | dot = p++; |
2416 | 7.66k | continue; |
2417 | 7.66k | } |
2418 | | |
2419 | 1.10M | hex_value = hexDigitValue(*p); |
2420 | 1.10M | if (hex_value == -1U) |
2421 | 131k | break; |
2422 | | |
2423 | 970k | p++; |
2424 | | |
2425 | | // Store the number while we have space. |
2426 | 970k | if (bitPos) { |
2427 | 613k | bitPos -= 4; |
2428 | 613k | hex_value <<= bitPos % integerPartWidth; |
2429 | 613k | significand[bitPos / integerPartWidth] |= hex_value; |
2430 | 613k | } else if (!computedTrailingFraction) { |
2431 | 25.8k | lost_fraction = trailingHexadecimalFraction(p, end, hex_value); |
2432 | 25.8k | computedTrailingFraction = true; |
2433 | 25.8k | } |
2434 | 970k | } |
2435 | | |
2436 | | /* Hex floats require an exponent but not a hexadecimal point. */ |
2437 | 131k | assert(p != end && "Hex strings require an exponent"); |
2438 | 131k | assert((*p == 'p' || *p == 'P') && "Invalid character in significand"); |
2439 | 131k | assert(p != begin && "Significand has no digits"); |
2440 | 131k | assert((dot == end || p - begin != 1) && "Significand has no digits"); |
2441 | | |
2442 | | /* Ignore the exponent if we are zero. */ |
2443 | 131k | if (p != firstSignificantDigit) { |
2444 | 127k | int expAdjustment; |
2445 | | |
2446 | | /* Implicit hexadecimal point? */ |
2447 | 127k | if (dot == end) |
2448 | 68.7k | dot = p; |
2449 | | |
2450 | | /* Calculate the exponent adjustment implicit in the number of |
2451 | | significant digits. */ |
2452 | 127k | expAdjustment = static_cast<int>(dot - firstSignificantDigit); |
2453 | 127k | if (expAdjustment < 0) |
2454 | 51.2k | expAdjustment++; |
2455 | 127k | expAdjustment = expAdjustment * 4 - 1; |
2456 | | |
2457 | | /* Adjust for writing the significand starting at the most |
2458 | | significant nibble. */ |
2459 | 127k | expAdjustment += semantics->precision; |
2460 | 127k | expAdjustment -= partsCount * integerPartWidth; |
2461 | | |
2462 | | /* Adjust for the given exponent. */ |
2463 | 127k | exponent = totalExponent(p + 1, end, expAdjustment); |
2464 | 127k | } |
2465 | | |
2466 | 131k | return normalize(rounding_mode, lost_fraction); |
2467 | 131k | } |
2468 | | |
2469 | | APFloat::opStatus |
2470 | | APFloat::roundSignificandWithExponent(const integerPart *decSigParts, |
2471 | | unsigned sigPartCount, int exp, |
2472 | | roundingMode rounding_mode) |
2473 | 730k | { |
2474 | 730k | unsigned int parts, pow5PartCount; |
2475 | 730k | fltSemantics calcSemantics = { 32767, -32767, 0, 0 }; |
2476 | 730k | integerPart pow5Parts[maxPowerOfFiveParts]; |
2477 | 730k | bool isNearest; |
2478 | | |
2479 | 730k | isNearest = (rounding_mode == rmNearestTiesToEven || |
2480 | 0 | rounding_mode == rmNearestTiesToAway); |
2481 | | |
2482 | 730k | parts = partCountForBits(semantics->precision + 11); |
2483 | | |
2484 | | /* Calculate pow(5, abs(exp)). */ |
2485 | 730k | pow5PartCount = powerOf5(pow5Parts, exp >= 0 ? exp: -exp); |
2486 | | |
2487 | 767k | for (;; parts *= 2) { |
2488 | 767k | opStatus sigStatus, powStatus; |
2489 | 767k | unsigned int excessPrecision, truncatedBits; |
2490 | | |
2491 | 767k | calcSemantics.precision = parts * integerPartWidth - 1; |
2492 | 767k | excessPrecision = calcSemantics.precision - semantics->precision; |
2493 | 767k | truncatedBits = excessPrecision; |
2494 | | |
2495 | 767k | APFloat decSig = APFloat::getZero(calcSemantics, sign); |
2496 | 767k | APFloat pow5(calcSemantics); |
2497 | | |
2498 | 767k | sigStatus = decSig.convertFromUnsignedParts(decSigParts, sigPartCount, |
2499 | 767k | rmNearestTiesToEven); |
2500 | 767k | powStatus = pow5.convertFromUnsignedParts(pow5Parts, pow5PartCount, |
2501 | 767k | rmNearestTiesToEven); |
2502 | | /* Add exp, as 10^n = 5^n * 2^n. */ |
2503 | 767k | decSig.exponent += exp; |
2504 | | |
2505 | 767k | lostFraction calcLostFraction; |
2506 | 767k | integerPart HUerr, HUdistance; |
2507 | 767k | unsigned int powHUerr; |
2508 | | |
2509 | 767k | if (exp >= 0) { |
2510 | | /* multiplySignificand leaves the precision-th bit set to 1. */ |
2511 | 91.9k | calcLostFraction = decSig.multiplySignificand(pow5, nullptr); |
2512 | 91.9k | powHUerr = powStatus != opOK; |
2513 | 675k | } else { |
2514 | 675k | calcLostFraction = decSig.divideSignificand(pow5); |
2515 | | /* Denormal numbers have less precision. */ |
2516 | 675k | if (decSig.exponent < semantics->minExponent) { |
2517 | 7.46k | excessPrecision += (semantics->minExponent - decSig.exponent); |
2518 | 7.46k | truncatedBits = excessPrecision; |
2519 | 7.46k | if (excessPrecision > calcSemantics.precision) |
2520 | 1.09k | excessPrecision = calcSemantics.precision; |
2521 | 7.46k | } |
2522 | | /* Extra half-ulp lost in reciprocal of exponent. */ |
2523 | 675k | powHUerr = (powStatus == opOK && calcLostFraction == lfExactlyZero) ? 0:2; |
2524 | 675k | } |
2525 | | |
2526 | | /* Both multiplySignificand and divideSignificand return the |
2527 | | result with the integer bit set. */ |
2528 | 767k | assert(APInt::tcExtractBit |
2529 | 767k | (decSig.significandParts(), calcSemantics.precision - 1) == 1); |
2530 | | |
2531 | 767k | HUerr = HUerrBound(calcLostFraction != lfExactlyZero, sigStatus != opOK, |
2532 | 767k | powHUerr); |
2533 | 767k | HUdistance = 2 * ulpsFromBoundary(decSig.significandParts(), |
2534 | 767k | excessPrecision, isNearest); |
2535 | | |
2536 | | /* Are we guaranteed to round correctly if we truncate? */ |
2537 | 767k | if (HUdistance >= HUerr) { |
2538 | 730k | APInt::tcExtract(significandParts(), partCount(), decSig.significandParts(), |
2539 | 730k | calcSemantics.precision - excessPrecision, |
2540 | 730k | excessPrecision); |
2541 | | /* Take the exponent of decSig. If we tcExtract-ed less bits |
2542 | | above we must adjust our exponent to compensate for the |
2543 | | implicit right shift. */ |
2544 | 730k | exponent = (decSig.exponent + semantics->precision |
2545 | 730k | - (calcSemantics.precision - excessPrecision)); |
2546 | 730k | calcLostFraction = lostFractionThroughTruncation(decSig.significandParts(), |
2547 | 730k | decSig.partCount(), |
2548 | 730k | truncatedBits); |
2549 | 730k | return normalize(rounding_mode, calcLostFraction); |
2550 | 730k | } |
2551 | 767k | } |
2552 | 730k | } |
2553 | | |
2554 | | APFloat::opStatus |
2555 | | APFloat::convertFromDecimalString(StringRef str, roundingMode rounding_mode) // qq |
2556 | 828k | { |
2557 | 828k | decimalInfo D; |
2558 | 828k | opStatus fs; |
2559 | | |
2560 | | /* Scan the text. */ |
2561 | 828k | StringRef::iterator p = str.begin(); |
2562 | 828k | fs = interpretDecimal(p, str.end(), &D); |
2563 | 828k | if (fs != opOK) |
2564 | 16.6k | return fs; |
2565 | | |
2566 | | /* Handle the quick cases. First the case of no significant digits, |
2567 | | i.e. zero, and then exponents that are obviously too large or too |
2568 | | small. Writing L for log 10 / log 2, a number d.ddddd*10^exp |
2569 | | definitely overflows if |
2570 | | |
2571 | | (exp - 1) * L >= maxExponent |
2572 | | |
2573 | | and definitely underflows to zero where |
2574 | | |
2575 | | (exp + 1) * L <= minExponent - precision |
2576 | | |
2577 | | With integer arithmetic the tightest bounds for L are |
2578 | | |
2579 | | 93/28 < L < 196/59 [ numerator <= 256 ] |
2580 | | 42039/12655 < L < 28738/8651 [ numerator <= 65536 ] |
2581 | | */ |
2582 | | |
2583 | | // Test if we have a zero number allowing for strings with no null terminators |
2584 | | // and zero decimals with non-zero exponents. |
2585 | | // |
2586 | | // We computed firstSigDigit by ignoring all zeros and dots. Thus if |
2587 | | // D->firstSigDigit equals str.end(), every digit must be a zero and there can |
2588 | | // be at most one dot. On the other hand, if we have a zero with a non-zero |
2589 | | // exponent, then we know that D.firstSigDigit will be non-numeric. |
2590 | 811k | if (D.firstSigDigit == str.end() || decDigitValue(*D.firstSigDigit) >= 10U) { |
2591 | 54.9k | category = fcZero; |
2592 | 54.9k | fs = opOK; |
2593 | | |
2594 | | /* Check whether the normalized exponent is high enough to overflow |
2595 | | max during the log-rebasing in the max-exponent check below. */ |
2596 | 756k | } else if (D.normalizedExponent - 1 > INT_MAX / 42039) { |
2597 | 4.22k | fs = handleOverflow(rounding_mode); |
2598 | | |
2599 | | /* If it wasn't, then it also wasn't high enough to overflow max |
2600 | | during the log-rebasing in the min-exponent check. Check that it |
2601 | | won't overflow min in either check, then perform the min-exponent |
2602 | | check. */ |
2603 | 752k | } else if (D.normalizedExponent - 1 < INT_MIN / 42039 || |
2604 | 750k | (D.normalizedExponent + 1) * 28738 <= |
2605 | 750k | 8651 * (semantics->minExponent - (int) semantics->precision)) { |
2606 | | /* Underflow to zero and round. */ |
2607 | 11.2k | category = fcNormal; |
2608 | 11.2k | zeroSignificand(); |
2609 | 11.2k | fs = normalize(rounding_mode, lfLessThanHalf); |
2610 | | |
2611 | | /* We can finally safely perform the max-exponent check. */ |
2612 | 741k | } else if ((D.normalizedExponent - 1) * 42039 |
2613 | 741k | >= 12655 * semantics->maxExponent) { |
2614 | | /* Overflow and round. */ |
2615 | 10.5k | fs = handleOverflow(rounding_mode); |
2616 | 730k | } else { |
2617 | 730k | integerPart *decSignificand; |
2618 | 730k | unsigned int partCount; |
2619 | | |
2620 | | /* A tight upper bound on number of bits required to hold an |
2621 | | N-digit decimal integer is N * 196 / 59. Allocate enough space |
2622 | | to hold the full significand, and an extra part required by |
2623 | | tcMultiplyPart. */ |
2624 | 730k | partCount = static_cast<unsigned int>(D.lastSigDigit - D.firstSigDigit) + 1; |
2625 | 730k | partCount = partCountForBits(1 + 196 * partCount / 59); |
2626 | 730k | decSignificand = new integerPart[partCount + 1]; |
2627 | 730k | partCount = 0; |
2628 | | |
2629 | | /* Convert to binary efficiently - we do almost all multiplication |
2630 | | in an integerPart. When this would overflow do we do a single |
2631 | | bignum multiplication, and then revert again to multiplication |
2632 | | in an integerPart. */ |
2633 | 908k | do { |
2634 | 908k | integerPart decValue, val, multiplier; |
2635 | | |
2636 | 908k | val = 0; |
2637 | 908k | multiplier = 1; |
2638 | | |
2639 | 5.69M | do { |
2640 | 5.69M | if (*p == '.') { |
2641 | 690k | p++; |
2642 | 690k | if (p == str.end()) { |
2643 | 0 | break; |
2644 | 0 | } |
2645 | 690k | } |
2646 | 5.69M | decValue = decDigitValue(*p++); |
2647 | 5.69M | assert(decValue < 10U && "Invalid character in significand"); |
2648 | 5.69M | multiplier *= 10; |
2649 | 5.69M | val = val * 10 + decValue; |
2650 | | /* The maximum number that can be multiplied by ten with any |
2651 | | digit added without overflowing an integerPart. */ |
2652 | 5.69M | } while (p <= D.lastSigDigit && multiplier <= (~ (integerPart) 0 - 9) / 10); |
2653 | | |
2654 | | /* Multiply out the current part. */ |
2655 | 908k | APInt::tcMultiplyPart(decSignificand, decSignificand, multiplier, val, |
2656 | 908k | partCount, partCount + 1, false); |
2657 | | |
2658 | | /* If we used another part (likely but not guaranteed), increase |
2659 | | the count. */ |
2660 | 908k | if (decSignificand[partCount]) |
2661 | 871k | partCount++; |
2662 | 908k | } while (p <= D.lastSigDigit); |
2663 | | |
2664 | 730k | category = fcNormal; |
2665 | 730k | fs = roundSignificandWithExponent(decSignificand, partCount, |
2666 | 730k | D.exponent, rounding_mode); |
2667 | | |
2668 | 730k | delete [] decSignificand; |
2669 | 730k | } |
2670 | | |
2671 | 811k | return fs; |
2672 | 811k | } |
2673 | | |
2674 | | bool |
2675 | 960k | APFloat::convertFromStringSpecials(StringRef str) { |
2676 | 960k | if (str.equals("inf") || str.equals("INFINITY")) { |
2677 | 0 | makeInf(false); |
2678 | 0 | return true; |
2679 | 0 | } |
2680 | | |
2681 | 960k | if (str.equals("-inf") || str.equals("-INFINITY")) { |
2682 | 0 | makeInf(true); |
2683 | 0 | return true; |
2684 | 0 | } |
2685 | | |
2686 | 960k | if (str.equals("nan") || str.equals("NaN")) { |
2687 | 0 | makeNaN(false, false); |
2688 | 0 | return true; |
2689 | 0 | } |
2690 | | |
2691 | 960k | if (str.equals("-nan") || str.equals("-NaN")) { |
2692 | 0 | makeNaN(false, true); |
2693 | 0 | return true; |
2694 | 0 | } |
2695 | | |
2696 | 960k | return false; |
2697 | 960k | } |
2698 | | |
2699 | | APFloat::opStatus |
2700 | | APFloat::convertFromString(StringRef str, roundingMode rounding_mode) |
2701 | 960k | { |
2702 | 960k | assert(!str.empty() && "Invalid string length"); |
2703 | | |
2704 | | // Handle special cases. |
2705 | 960k | if (convertFromStringSpecials(str)) |
2706 | 0 | return opOK; |
2707 | | |
2708 | | /* Handle a leading minus sign. */ |
2709 | 960k | StringRef::iterator p = str.begin(); |
2710 | 960k | size_t slen = str.size(); |
2711 | 960k | sign = *p == '-' ? 1 : 0; |
2712 | 960k | if (*p == '-' || *p == '+') { |
2713 | 0 | p++; |
2714 | 0 | slen--; |
2715 | 0 | assert(slen && "String has no digits"); |
2716 | 0 | } |
2717 | | |
2718 | 960k | if (slen >= 2 && p[0] == '0' && (p[1] == 'x' || p[1] == 'X')) { |
2719 | 131k | assert(slen - 2 && "Invalid string"); |
2720 | 131k | return convertFromHexadecimalString(StringRef(p + 2, slen - 2), |
2721 | 131k | rounding_mode); |
2722 | 131k | } |
2723 | | |
2724 | 828k | return convertFromDecimalString(StringRef(p, slen), rounding_mode); |
2725 | 960k | } |
2726 | | |
2727 | | /* Write out a hexadecimal representation of the floating point value |
2728 | | to DST, which must be of sufficient size, in the C99 form |
2729 | | [-]0xh.hhhhp[+-]d. Return the number of characters written, |
2730 | | excluding the terminating NUL. |
2731 | | |
2732 | | If UPPERCASE, the output is in upper case, otherwise in lower case. |
2733 | | |
2734 | | HEXDIGITS digits appear altogether, rounding the value if |
2735 | | necessary. If HEXDIGITS is 0, the minimal precision to display the |
2736 | | number precisely is used instead. If nothing would appear after |
2737 | | the decimal point it is suppressed. |
2738 | | |
2739 | | The decimal exponent is always printed and has at least one digit. |
2740 | | Zero values display an exponent of zero. Infinities and NaNs |
2741 | | appear as "infinity" or "nan" respectively. |
2742 | | |
2743 | | The above rules are as specified by C99. There is ambiguity about |
2744 | | what the leading hexadecimal digit should be. This implementation |
2745 | | uses whatever is necessary so that the exponent is displayed as |
2746 | | stored. This implies the exponent will fall within the IEEE format |
2747 | | range, and the leading hexadecimal digit will be 0 (for denormals), |
2748 | | 1 (normal numbers) or 2 (normal numbers rounded-away-from-zero with |
2749 | | any other digits zero). |
2750 | | */ |
2751 | | unsigned int |
2752 | | APFloat::convertToHexString(char *dst, unsigned int hexDigits, |
2753 | | bool upperCase, roundingMode rounding_mode) const |
2754 | 0 | { |
2755 | 0 | char *p; |
2756 | |
|
2757 | 0 | p = dst; |
2758 | 0 | if (sign) |
2759 | 0 | *dst++ = '-'; |
2760 | |
|
2761 | 0 | switch (category) { |
2762 | 0 | case fcInfinity: |
2763 | 0 | memcpy (dst, upperCase ? infinityU: infinityL, sizeof infinityU - 1); |
2764 | 0 | dst += sizeof infinityL - 1; |
2765 | 0 | break; |
2766 | | |
2767 | 0 | case fcNaN: |
2768 | 0 | memcpy (dst, upperCase ? NaNU: NaNL, sizeof NaNU - 1); |
2769 | 0 | dst += sizeof NaNU - 1; |
2770 | 0 | break; |
2771 | | |
2772 | 0 | case fcZero: |
2773 | 0 | *dst++ = '0'; |
2774 | 0 | *dst++ = upperCase ? 'X': 'x'; |
2775 | 0 | *dst++ = '0'; |
2776 | 0 | if (hexDigits > 1) { |
2777 | 0 | *dst++ = '.'; |
2778 | 0 | memset (dst, '0', hexDigits - 1); |
2779 | 0 | dst += hexDigits - 1; |
2780 | 0 | } |
2781 | 0 | *dst++ = upperCase ? 'P': 'p'; |
2782 | 0 | *dst++ = '0'; |
2783 | 0 | break; |
2784 | | |
2785 | 0 | case fcNormal: |
2786 | 0 | dst = convertNormalToHexString (dst, hexDigits, upperCase, rounding_mode); |
2787 | 0 | break; |
2788 | 0 | } |
2789 | | |
2790 | 0 | *dst = 0; |
2791 | |
|
2792 | 0 | return static_cast<unsigned int>(dst - p); |
2793 | 0 | } |
2794 | | |
2795 | | /* Does the hard work of outputting the correctly rounded hexadecimal |
2796 | | form of a normal floating point number with the specified number of |
2797 | | hexadecimal digits. If HEXDIGITS is zero the minimum number of |
2798 | | digits necessary to print the value precisely is output. */ |
2799 | | char * |
2800 | | APFloat::convertNormalToHexString(char *dst, unsigned int hexDigits, |
2801 | | bool upperCase, |
2802 | | roundingMode rounding_mode) const |
2803 | 0 | { |
2804 | 0 | unsigned int count, valueBits, shift, partsCount, outputDigits; |
2805 | 0 | const char *hexDigitChars; |
2806 | 0 | const integerPart *significand; |
2807 | 0 | char *p; |
2808 | 0 | bool roundUp; |
2809 | |
|
2810 | 0 | *dst++ = '0'; |
2811 | 0 | *dst++ = upperCase ? 'X': 'x'; |
2812 | |
|
2813 | 0 | roundUp = false; |
2814 | 0 | hexDigitChars = upperCase ? hexDigitsUpper: hexDigitsLower; |
2815 | |
|
2816 | 0 | significand = significandParts(); |
2817 | 0 | partsCount = partCount(); |
2818 | | |
2819 | | /* +3 because the first digit only uses the single integer bit, so |
2820 | | we have 3 virtual zero most-significant-bits. */ |
2821 | 0 | valueBits = semantics->precision + 3; |
2822 | 0 | shift = integerPartWidth - valueBits % integerPartWidth; |
2823 | | |
2824 | | /* The natural number of digits required ignoring trailing |
2825 | | insignificant zeroes. */ |
2826 | 0 | outputDigits = (valueBits - significandLSB () + 3) / 4; |
2827 | | |
2828 | | /* hexDigits of zero means use the required number for the |
2829 | | precision. Otherwise, see if we are truncating. If we are, |
2830 | | find out if we need to round away from zero. */ |
2831 | 0 | if (hexDigits) { |
2832 | 0 | if (hexDigits < outputDigits) { |
2833 | | /* We are dropping non-zero bits, so need to check how to round. |
2834 | | "bits" is the number of dropped bits. */ |
2835 | 0 | unsigned int bits; |
2836 | 0 | lostFraction fraction; |
2837 | |
|
2838 | 0 | bits = valueBits - hexDigits * 4; |
2839 | 0 | fraction = lostFractionThroughTruncation (significand, partsCount, bits); |
2840 | 0 | roundUp = roundAwayFromZero(rounding_mode, fraction, bits); |
2841 | 0 | } |
2842 | 0 | outputDigits = hexDigits; |
2843 | 0 | } |
2844 | | |
2845 | | /* Write the digits consecutively, and start writing in the location |
2846 | | of the hexadecimal point. We move the most significant digit |
2847 | | left and add the hexadecimal point later. */ |
2848 | 0 | p = ++dst; |
2849 | |
|
2850 | 0 | count = (valueBits + integerPartWidth - 1) / integerPartWidth; |
2851 | |
|
2852 | 0 | while (outputDigits && count) { |
2853 | 0 | integerPart part; |
2854 | | |
2855 | | /* Put the most significant integerPartWidth bits in "part". */ |
2856 | 0 | if (--count == partsCount) |
2857 | 0 | part = 0; /* An imaginary higher zero part. */ |
2858 | 0 | else |
2859 | 0 | part = significand[count] << shift; |
2860 | |
|
2861 | 0 | if (count && shift) |
2862 | 0 | part |= significand[count - 1] >> (integerPartWidth - shift); |
2863 | | |
2864 | | /* Convert as much of "part" to hexdigits as we can. */ |
2865 | 0 | unsigned int curDigits = integerPartWidth / 4; |
2866 | |
|
2867 | 0 | if (curDigits > outputDigits) |
2868 | 0 | curDigits = outputDigits; |
2869 | 0 | dst += partAsHex (dst, part, curDigits, hexDigitChars); |
2870 | 0 | outputDigits -= curDigits; |
2871 | 0 | } |
2872 | |
|
2873 | 0 | if (roundUp) { |
2874 | 0 | char *q = dst; |
2875 | | |
2876 | | /* Note that hexDigitChars has a trailing '0'. */ |
2877 | 0 | do { |
2878 | 0 | q--; |
2879 | 0 | *q = hexDigitChars[hexDigitValue (*q) + 1]; |
2880 | 0 | } while (*q == '0'); |
2881 | 0 | assert(q >= p); |
2882 | 0 | } else { |
2883 | | /* Add trailing zeroes. */ |
2884 | 0 | memset (dst, '0', outputDigits); |
2885 | 0 | dst += outputDigits; |
2886 | 0 | } |
2887 | | |
2888 | | /* Move the most significant digit to before the point, and if there |
2889 | | is something after the decimal point add it. This must come |
2890 | | after rounding above. */ |
2891 | 0 | p[-1] = p[0]; |
2892 | 0 | if (dst -1 == p) |
2893 | 0 | dst--; |
2894 | 0 | else |
2895 | 0 | p[0] = '.'; |
2896 | | |
2897 | | /* Finally output the exponent. */ |
2898 | 0 | *dst++ = upperCase ? 'P': 'p'; |
2899 | |
|
2900 | 0 | return writeSignedDecimal (dst, exponent); |
2901 | 0 | } |
2902 | | |
2903 | 0 | hash_code llvm_ks::hash_value(const APFloat &Arg) { |
2904 | 0 | if (!Arg.isFiniteNonZero()) |
2905 | 0 | return hash_combine((uint8_t)Arg.category, |
2906 | | // NaN has no sign, fix it at zero. |
2907 | 0 | Arg.isNaN() ? (uint8_t)0 : (uint8_t)Arg.sign, |
2908 | 0 | Arg.semantics->precision); |
2909 | | |
2910 | | // Normal floats need their exponent and significand hashed. |
2911 | 0 | return hash_combine((uint8_t)Arg.category, (uint8_t)Arg.sign, |
2912 | 0 | Arg.semantics->precision, Arg.exponent, |
2913 | 0 | hash_combine_range( |
2914 | 0 | Arg.significandParts(), |
2915 | 0 | Arg.significandParts() + Arg.partCount())); |
2916 | 0 | } |
2917 | | |
2918 | | // Conversion from APFloat to/from host float/double. It may eventually be |
2919 | | // possible to eliminate these and have everybody deal with APFloats, but that |
2920 | | // will take a while. This approach will not easily extend to long double. |
2921 | | // Current implementation requires integerPartWidth==64, which is correct at |
2922 | | // the moment but could be made more general. |
2923 | | |
2924 | | // Denormals have exponent minExponent in APFloat, but minExponent-1 in |
2925 | | // the actual IEEE respresentations. We compensate for that here. |
2926 | | |
2927 | | APInt |
2928 | | APFloat::convertF80LongDoubleAPFloatToAPInt() const |
2929 | 0 | { |
2930 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&x87DoubleExtended); |
2931 | 0 | assert(partCount()==2); |
2932 | | |
2933 | 0 | uint64_t myexponent, mysignificand; |
2934 | |
|
2935 | 0 | if (isFiniteNonZero()) { |
2936 | 0 | myexponent = exponent+16383; //bias |
2937 | 0 | mysignificand = significandParts()[0]; |
2938 | 0 | if (myexponent==1 && !(mysignificand & 0x8000000000000000ULL)) |
2939 | 0 | myexponent = 0; // denormal |
2940 | 0 | } else if (category==fcZero) { |
2941 | 0 | myexponent = 0; |
2942 | 0 | mysignificand = 0; |
2943 | 0 | } else if (category==fcInfinity) { |
2944 | 0 | myexponent = 0x7fff; |
2945 | 0 | mysignificand = 0x8000000000000000ULL; |
2946 | 0 | } else { |
2947 | 0 | assert(category == fcNaN && "Unknown category"); |
2948 | 0 | myexponent = 0x7fff; |
2949 | 0 | mysignificand = significandParts()[0]; |
2950 | 0 | } |
2951 | | |
2952 | 0 | uint64_t words[2]; |
2953 | 0 | words[0] = mysignificand; |
2954 | 0 | words[1] = ((uint64_t)(sign & 1) << 15) | |
2955 | 0 | (myexponent & 0x7fffLL); |
2956 | 0 | return APInt(80, words); |
2957 | 0 | } |
2958 | | |
2959 | | APInt |
2960 | | APFloat::convertPPCDoubleDoubleAPFloatToAPInt() const |
2961 | 0 | { |
2962 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&PPCDoubleDouble); |
2963 | 0 | assert(partCount()==2); |
2964 | | |
2965 | 0 | uint64_t words[2]; |
2966 | 0 | opStatus fs; |
2967 | 0 | bool losesInfo; |
2968 | | |
2969 | | // Convert number to double. To avoid spurious underflows, we re- |
2970 | | // normalize against the "double" minExponent first, and only *then* |
2971 | | // truncate the mantissa. The result of that second conversion |
2972 | | // may be inexact, but should never underflow. |
2973 | | // Declare fltSemantics before APFloat that uses it (and |
2974 | | // saves pointer to it) to ensure correct destruction order. |
2975 | 0 | fltSemantics extendedSemantics = *semantics; |
2976 | 0 | extendedSemantics.minExponent = IEEEdouble.minExponent; |
2977 | 0 | APFloat extended(*this); |
2978 | 0 | fs = extended.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo); |
2979 | 0 | assert(fs == opOK && !losesInfo); |
2980 | 0 | (void)fs; |
2981 | |
|
2982 | 0 | APFloat u(extended); |
2983 | 0 | fs = u.convert(IEEEdouble, rmNearestTiesToEven, &losesInfo); |
2984 | 0 | assert(fs == opOK || fs == opInexact); |
2985 | 0 | (void)fs; |
2986 | 0 | words[0] = *u.convertDoubleAPFloatToAPInt().getRawData(); |
2987 | | |
2988 | | // If conversion was exact or resulted in a special case, we're done; |
2989 | | // just set the second double to zero. Otherwise, re-convert back to |
2990 | | // the extended format and compute the difference. This now should |
2991 | | // convert exactly to double. |
2992 | 0 | if (u.isFiniteNonZero() && losesInfo) { |
2993 | 0 | fs = u.convert(extendedSemantics, rmNearestTiesToEven, &losesInfo); |
2994 | 0 | assert(fs == opOK && !losesInfo); |
2995 | 0 | (void)fs; |
2996 | |
|
2997 | 0 | APFloat v(extended); |
2998 | 0 | v.subtract(u, rmNearestTiesToEven); |
2999 | 0 | fs = v.convert(IEEEdouble, rmNearestTiesToEven, &losesInfo); |
3000 | 0 | assert(fs == opOK && !losesInfo); |
3001 | 0 | (void)fs; |
3002 | 0 | words[1] = *v.convertDoubleAPFloatToAPInt().getRawData(); |
3003 | 0 | } else { |
3004 | 0 | words[1] = 0; |
3005 | 0 | } |
3006 | | |
3007 | 0 | return APInt(128, words); |
3008 | 0 | } |
3009 | | |
3010 | | APInt |
3011 | | APFloat::convertQuadrupleAPFloatToAPInt() const |
3012 | 0 | { |
3013 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEquad); |
3014 | 0 | assert(partCount()==2); |
3015 | | |
3016 | 0 | uint64_t myexponent, mysignificand, mysignificand2; |
3017 | |
|
3018 | 0 | if (isFiniteNonZero()) { |
3019 | 0 | myexponent = exponent+16383; //bias |
3020 | 0 | mysignificand = significandParts()[0]; |
3021 | 0 | mysignificand2 = significandParts()[1]; |
3022 | 0 | if (myexponent==1 && !(mysignificand2 & 0x1000000000000LL)) |
3023 | 0 | myexponent = 0; // denormal |
3024 | 0 | } else if (category==fcZero) { |
3025 | 0 | myexponent = 0; |
3026 | 0 | mysignificand = mysignificand2 = 0; |
3027 | 0 | } else if (category==fcInfinity) { |
3028 | 0 | myexponent = 0x7fff; |
3029 | 0 | mysignificand = mysignificand2 = 0; |
3030 | 0 | } else { |
3031 | 0 | assert(category == fcNaN && "Unknown category!"); |
3032 | 0 | myexponent = 0x7fff; |
3033 | 0 | mysignificand = significandParts()[0]; |
3034 | 0 | mysignificand2 = significandParts()[1]; |
3035 | 0 | } |
3036 | | |
3037 | 0 | uint64_t words[2]; |
3038 | 0 | words[0] = mysignificand; |
3039 | 0 | words[1] = ((uint64_t)(sign & 1) << 63) | |
3040 | 0 | ((myexponent & 0x7fff) << 48) | |
3041 | 0 | (mysignificand2 & 0xffffffffffffLL); |
3042 | |
|
3043 | 0 | return APInt(128, words); |
3044 | 0 | } |
3045 | | |
3046 | | APInt |
3047 | | APFloat::convertDoubleAPFloatToAPInt() const |
3048 | 883k | { |
3049 | 883k | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEdouble); |
3050 | 883k | assert(partCount()==1); |
3051 | | |
3052 | 883k | uint64_t myexponent, mysignificand; |
3053 | | |
3054 | 883k | if (isFiniteNonZero()) { |
3055 | 774k | myexponent = exponent+1023; //bias |
3056 | 774k | mysignificand = *significandParts(); |
3057 | 774k | if (myexponent==1 && !(mysignificand & 0x10000000000000LL)) |
3058 | 4.80k | myexponent = 0; // denormal |
3059 | 774k | } else if (category==fcZero) { |
3060 | 77.8k | myexponent = 0; |
3061 | 77.8k | mysignificand = 0; |
3062 | 77.8k | } else if (category==fcInfinity) { |
3063 | 29.9k | myexponent = 0x7ff; |
3064 | 29.9k | mysignificand = 0; |
3065 | 29.9k | } else { |
3066 | 1.52k | assert(category == fcNaN && "Unknown category!"); |
3067 | 1.52k | myexponent = 0x7ff; |
3068 | 1.52k | mysignificand = *significandParts(); |
3069 | 1.52k | } |
3070 | | |
3071 | 883k | return APInt(64, ((((uint64_t)(sign & 1) << 63) | |
3072 | 883k | ((myexponent & 0x7ff) << 52) | |
3073 | 883k | (mysignificand & 0xfffffffffffffLL)))); |
3074 | 883k | } |
3075 | | |
3076 | | APInt |
3077 | | APFloat::convertFloatAPFloatToAPInt() const |
3078 | 93.5k | { |
3079 | 93.5k | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEsingle); |
3080 | 93.5k | assert(partCount()==1); |
3081 | | |
3082 | 93.5k | uint32_t myexponent, mysignificand; |
3083 | | |
3084 | 93.5k | if (isFiniteNonZero()) { |
3085 | 62.3k | myexponent = exponent+127; //bias |
3086 | 62.3k | mysignificand = (uint32_t)*significandParts(); |
3087 | 62.3k | if (myexponent == 1 && !(mysignificand & 0x800000)) |
3088 | 1.56k | myexponent = 0; // denormal |
3089 | 62.3k | } else if (category==fcZero) { |
3090 | 14.8k | myexponent = 0; |
3091 | 14.8k | mysignificand = 0; |
3092 | 16.4k | } else if (category==fcInfinity) { |
3093 | 11.5k | myexponent = 0xff; |
3094 | 11.5k | mysignificand = 0; |
3095 | 11.5k | } else { |
3096 | 4.88k | assert(category == fcNaN && "Unknown category!"); |
3097 | 4.88k | myexponent = 0xff; |
3098 | 4.88k | mysignificand = (uint32_t)*significandParts(); |
3099 | 4.88k | } |
3100 | | |
3101 | 93.5k | return APInt(32, (((sign&1) << 31) | ((myexponent&0xff) << 23) | |
3102 | 93.5k | (mysignificand & 0x7fffff))); |
3103 | 93.5k | } |
3104 | | |
3105 | | APInt |
3106 | | APFloat::convertHalfAPFloatToAPInt() const |
3107 | 0 | { |
3108 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEhalf); |
3109 | 0 | assert(partCount()==1); |
3110 | | |
3111 | 0 | uint32_t myexponent, mysignificand; |
3112 | |
|
3113 | 0 | if (isFiniteNonZero()) { |
3114 | 0 | myexponent = exponent+15; //bias |
3115 | 0 | mysignificand = (uint32_t)*significandParts(); |
3116 | 0 | if (myexponent == 1 && !(mysignificand & 0x400)) |
3117 | 0 | myexponent = 0; // denormal |
3118 | 0 | } else if (category==fcZero) { |
3119 | 0 | myexponent = 0; |
3120 | 0 | mysignificand = 0; |
3121 | 0 | } else if (category==fcInfinity) { |
3122 | 0 | myexponent = 0x1f; |
3123 | 0 | mysignificand = 0; |
3124 | 0 | } else { |
3125 | 0 | assert(category == fcNaN && "Unknown category!"); |
3126 | 0 | myexponent = 0x1f; |
3127 | 0 | mysignificand = (uint32_t)*significandParts(); |
3128 | 0 | } |
3129 | | |
3130 | 0 | return APInt(16, (((sign&1) << 15) | ((myexponent&0x1f) << 10) | |
3131 | 0 | (mysignificand & 0x3ff))); |
3132 | 0 | } |
3133 | | |
3134 | | // This function creates an APInt that is just a bit map of the floating |
3135 | | // point constant as it would appear in memory. It is not a conversion, |
3136 | | // and treating the result as a normal integer is unlikely to be useful. |
3137 | | |
3138 | | APInt |
3139 | | APFloat::bitcastToAPInt() const |
3140 | 977k | { |
3141 | 977k | if (semantics == (const llvm_ks::fltSemantics*)&IEEEhalf) |
3142 | 0 | return convertHalfAPFloatToAPInt(); |
3143 | | |
3144 | 977k | if (semantics == (const llvm_ks::fltSemantics*)&IEEEsingle) |
3145 | 93.5k | return convertFloatAPFloatToAPInt(); |
3146 | | |
3147 | 883k | if (semantics == (const llvm_ks::fltSemantics*)&IEEEdouble) |
3148 | 883k | return convertDoubleAPFloatToAPInt(); |
3149 | | |
3150 | 0 | if (semantics == (const llvm_ks::fltSemantics*)&IEEEquad) |
3151 | 0 | return convertQuadrupleAPFloatToAPInt(); |
3152 | | |
3153 | 0 | if (semantics == (const llvm_ks::fltSemantics*)&PPCDoubleDouble) |
3154 | 0 | return convertPPCDoubleDoubleAPFloatToAPInt(); |
3155 | | |
3156 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&x87DoubleExtended && |
3157 | 0 | "unknown format!"); |
3158 | 0 | return convertF80LongDoubleAPFloatToAPInt(); |
3159 | 0 | } |
3160 | | |
3161 | | float |
3162 | | APFloat::convertToFloat() const |
3163 | 0 | { |
3164 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEsingle && |
3165 | 0 | "Float semantics are not IEEEsingle"); |
3166 | 0 | APInt api = bitcastToAPInt(); |
3167 | 0 | return api.bitsToFloat(); |
3168 | 0 | } |
3169 | | |
3170 | | double |
3171 | | APFloat::convertToDouble() const |
3172 | 0 | { |
3173 | 0 | assert(semantics == (const llvm_ks::fltSemantics*)&IEEEdouble && |
3174 | 0 | "Float semantics are not IEEEdouble"); |
3175 | 0 | APInt api = bitcastToAPInt(); |
3176 | 0 | return api.bitsToDouble(); |
3177 | 0 | } |
3178 | | |
3179 | | /// Integer bit is explicit in this format. Intel hardware (387 and later) |
3180 | | /// does not support these bit patterns: |
3181 | | /// exponent = all 1's, integer bit 0, significand 0 ("pseudoinfinity") |
3182 | | /// exponent = all 1's, integer bit 0, significand nonzero ("pseudoNaN") |
3183 | | /// exponent = 0, integer bit 1 ("pseudodenormal") |
3184 | | /// exponent!=0 nor all 1's, integer bit 0 ("unnormal") |
3185 | | /// At the moment, the first two are treated as NaNs, the second two as Normal. |
3186 | | void |
3187 | | APFloat::initFromF80LongDoubleAPInt(const APInt &api) |
3188 | 0 | { |
3189 | 0 | assert(api.getBitWidth()==80); |
3190 | 0 | uint64_t i1 = api.getRawData()[0]; |
3191 | 0 | uint64_t i2 = api.getRawData()[1]; |
3192 | 0 | uint64_t myexponent = (i2 & 0x7fff); |
3193 | 0 | uint64_t mysignificand = i1; |
3194 | |
|
3195 | 0 | initialize(&APFloat::x87DoubleExtended); |
3196 | 0 | assert(partCount()==2); |
3197 | | |
3198 | 0 | sign = static_cast<unsigned int>(i2>>15); |
3199 | 0 | if (myexponent==0 && mysignificand==0) { |
3200 | | // exponent, significand meaningless |
3201 | 0 | category = fcZero; |
3202 | 0 | } else if (myexponent==0x7fff && mysignificand==0x8000000000000000ULL) { |
3203 | | // exponent, significand meaningless |
3204 | 0 | category = fcInfinity; |
3205 | 0 | } else if (myexponent==0x7fff && mysignificand!=0x8000000000000000ULL) { |
3206 | | // exponent meaningless |
3207 | 0 | category = fcNaN; |
3208 | 0 | significandParts()[0] = mysignificand; |
3209 | 0 | significandParts()[1] = 0; |
3210 | 0 | } else { |
3211 | 0 | category = fcNormal; |
3212 | 0 | exponent = myexponent - 16383; |
3213 | 0 | significandParts()[0] = mysignificand; |
3214 | 0 | significandParts()[1] = 0; |
3215 | 0 | if (myexponent==0) // denormal |
3216 | 0 | exponent = -16382; |
3217 | 0 | } |
3218 | 0 | } |
3219 | | |
3220 | | void |
3221 | | APFloat::initFromPPCDoubleDoubleAPInt(const APInt &api) |
3222 | 0 | { |
3223 | 0 | assert(api.getBitWidth()==128); |
3224 | 0 | uint64_t i1 = api.getRawData()[0]; |
3225 | 0 | uint64_t i2 = api.getRawData()[1]; |
3226 | 0 | opStatus fs; |
3227 | 0 | bool losesInfo; |
3228 | | |
3229 | | // Get the first double and convert to our format. |
3230 | 0 | initFromDoubleAPInt(APInt(64, i1)); |
3231 | 0 | fs = convert(PPCDoubleDouble, rmNearestTiesToEven, &losesInfo); |
3232 | 0 | assert(fs == opOK && !losesInfo); |
3233 | 0 | (void)fs; |
3234 | | |
3235 | | // Unless we have a special case, add in second double. |
3236 | 0 | if (isFiniteNonZero()) { |
3237 | 0 | APFloat v(IEEEdouble, APInt(64, i2)); |
3238 | 0 | fs = v.convert(PPCDoubleDouble, rmNearestTiesToEven, &losesInfo); |
3239 | 0 | assert(fs == opOK && !losesInfo); |
3240 | 0 | (void)fs; |
3241 | |
|
3242 | 0 | add(v, rmNearestTiesToEven); |
3243 | 0 | } |
3244 | 0 | } |
3245 | | |
3246 | | void |
3247 | | APFloat::initFromQuadrupleAPInt(const APInt &api) |
3248 | 0 | { |
3249 | 0 | assert(api.getBitWidth()==128); |
3250 | 0 | uint64_t i1 = api.getRawData()[0]; |
3251 | 0 | uint64_t i2 = api.getRawData()[1]; |
3252 | 0 | uint64_t myexponent = (i2 >> 48) & 0x7fff; |
3253 | 0 | uint64_t mysignificand = i1; |
3254 | 0 | uint64_t mysignificand2 = i2 & 0xffffffffffffLL; |
3255 | |
|
3256 | 0 | initialize(&APFloat::IEEEquad); |
3257 | 0 | assert(partCount()==2); |
3258 | | |
3259 | 0 | sign = static_cast<unsigned int>(i2>>63); |
3260 | 0 | if (myexponent==0 && |
3261 | 0 | (mysignificand==0 && mysignificand2==0)) { |
3262 | | // exponent, significand meaningless |
3263 | 0 | category = fcZero; |
3264 | 0 | } else if (myexponent==0x7fff && |
3265 | 0 | (mysignificand==0 && mysignificand2==0)) { |
3266 | | // exponent, significand meaningless |
3267 | 0 | category = fcInfinity; |
3268 | 0 | } else if (myexponent==0x7fff && |
3269 | 0 | (mysignificand!=0 || mysignificand2 !=0)) { |
3270 | | // exponent meaningless |
3271 | 0 | category = fcNaN; |
3272 | 0 | significandParts()[0] = mysignificand; |
3273 | 0 | significandParts()[1] = mysignificand2; |
3274 | 0 | } else { |
3275 | 0 | category = fcNormal; |
3276 | 0 | exponent = myexponent - 16383; |
3277 | 0 | significandParts()[0] = mysignificand; |
3278 | 0 | significandParts()[1] = mysignificand2; |
3279 | 0 | if (myexponent==0) // denormal |
3280 | 0 | exponent = -16382; |
3281 | 0 | else |
3282 | 0 | significandParts()[1] |= 0x1000000000000LL; // integer bit |
3283 | 0 | } |
3284 | 0 | } |
3285 | | |
3286 | | void |
3287 | | APFloat::initFromDoubleAPInt(const APInt &api) |
3288 | 0 | { |
3289 | 0 | assert(api.getBitWidth()==64); |
3290 | 0 | uint64_t i = *api.getRawData(); |
3291 | 0 | uint64_t myexponent = (i >> 52) & 0x7ff; |
3292 | 0 | uint64_t mysignificand = i & 0xfffffffffffffLL; |
3293 | |
|
3294 | 0 | initialize(&APFloat::IEEEdouble); |
3295 | 0 | assert(partCount()==1); |
3296 | | |
3297 | 0 | sign = static_cast<unsigned int>(i>>63); |
3298 | 0 | if (myexponent==0 && mysignificand==0) { |
3299 | | // exponent, significand meaningless |
3300 | 0 | category = fcZero; |
3301 | 0 | } else if (myexponent==0x7ff && mysignificand==0) { |
3302 | | // exponent, significand meaningless |
3303 | 0 | category = fcInfinity; |
3304 | 0 | } else if (myexponent==0x7ff && mysignificand!=0) { |
3305 | | // exponent meaningless |
3306 | 0 | category = fcNaN; |
3307 | 0 | *significandParts() = mysignificand; |
3308 | 0 | } else { |
3309 | 0 | category = fcNormal; |
3310 | 0 | exponent = myexponent - 1023; |
3311 | 0 | *significandParts() = mysignificand; |
3312 | 0 | if (myexponent==0) // denormal |
3313 | 0 | exponent = -1022; |
3314 | 0 | else |
3315 | 0 | *significandParts() |= 0x10000000000000LL; // integer bit |
3316 | 0 | } |
3317 | 0 | } |
3318 | | |
3319 | | void |
3320 | | APFloat::initFromFloatAPInt(const APInt & api) |
3321 | 844 | { |
3322 | 844 | assert(api.getBitWidth()==32); |
3323 | 844 | uint32_t i = (uint32_t)*api.getRawData(); |
3324 | 844 | uint32_t myexponent = (i >> 23) & 0xff; |
3325 | 844 | uint32_t mysignificand = i & 0x7fffff; |
3326 | | |
3327 | 844 | initialize(&APFloat::IEEEsingle); |
3328 | 844 | assert(partCount()==1); |
3329 | | |
3330 | 844 | sign = i >> 31; |
3331 | 844 | if (myexponent==0 && mysignificand==0) { |
3332 | | // exponent, significand meaningless |
3333 | 0 | category = fcZero; |
3334 | 844 | } else if (myexponent==0xff && mysignificand==0) { |
3335 | | // exponent, significand meaningless |
3336 | 0 | category = fcInfinity; |
3337 | 844 | } else if (myexponent==0xff && mysignificand!=0) { |
3338 | | // sign, exponent, significand meaningless |
3339 | 0 | category = fcNaN; |
3340 | 0 | *significandParts() = mysignificand; |
3341 | 844 | } else { |
3342 | 844 | category = fcNormal; |
3343 | 844 | exponent = myexponent - 127; //bias |
3344 | 844 | *significandParts() = mysignificand; |
3345 | 844 | if (myexponent==0) // denormal |
3346 | 0 | exponent = -126; |
3347 | 844 | else |
3348 | 844 | *significandParts() |= 0x800000; // integer bit |
3349 | 844 | } |
3350 | 844 | } |
3351 | | |
3352 | | void |
3353 | | APFloat::initFromHalfAPInt(const APInt & api) |
3354 | 0 | { |
3355 | 0 | assert(api.getBitWidth()==16); |
3356 | 0 | uint32_t i = (uint32_t)*api.getRawData(); |
3357 | 0 | uint32_t myexponent = (i >> 10) & 0x1f; |
3358 | 0 | uint32_t mysignificand = i & 0x3ff; |
3359 | |
|
3360 | 0 | initialize(&APFloat::IEEEhalf); |
3361 | 0 | assert(partCount()==1); |
3362 | | |
3363 | 0 | sign = i >> 15; |
3364 | 0 | if (myexponent==0 && mysignificand==0) { |
3365 | | // exponent, significand meaningless |
3366 | 0 | category = fcZero; |
3367 | 0 | } else if (myexponent==0x1f && mysignificand==0) { |
3368 | | // exponent, significand meaningless |
3369 | 0 | category = fcInfinity; |
3370 | 0 | } else if (myexponent==0x1f && mysignificand!=0) { |
3371 | | // sign, exponent, significand meaningless |
3372 | 0 | category = fcNaN; |
3373 | 0 | *significandParts() = mysignificand; |
3374 | 0 | } else { |
3375 | 0 | category = fcNormal; |
3376 | 0 | exponent = myexponent - 15; //bias |
3377 | 0 | *significandParts() = mysignificand; |
3378 | 0 | if (myexponent==0) // denormal |
3379 | 0 | exponent = -14; |
3380 | 0 | else |
3381 | 0 | *significandParts() |= 0x400; // integer bit |
3382 | 0 | } |
3383 | 0 | } |
3384 | | |
3385 | | /// Treat api as containing the bits of a floating point number. Currently |
3386 | | /// we infer the floating point type from the size of the APInt. The |
3387 | | /// isIEEE argument distinguishes between PPC128 and IEEE128 (not meaningful |
3388 | | /// when the size is anything else). |
3389 | | void |
3390 | | APFloat::initFromAPInt(const fltSemantics* Sem, const APInt& api) |
3391 | 844 | { |
3392 | 844 | if (Sem == &IEEEhalf) |
3393 | 0 | return initFromHalfAPInt(api); |
3394 | 844 | if (Sem == &IEEEsingle) |
3395 | 844 | return initFromFloatAPInt(api); |
3396 | 0 | if (Sem == &IEEEdouble) |
3397 | 0 | return initFromDoubleAPInt(api); |
3398 | 0 | if (Sem == &x87DoubleExtended) |
3399 | 0 | return initFromF80LongDoubleAPInt(api); |
3400 | 0 | if (Sem == &IEEEquad) |
3401 | 0 | return initFromQuadrupleAPInt(api); |
3402 | 0 | if (Sem == &PPCDoubleDouble) |
3403 | 0 | return initFromPPCDoubleDoubleAPInt(api); |
3404 | | |
3405 | 0 | llvm_unreachable(nullptr); |
3406 | 0 | } |
3407 | | |
3408 | | APFloat |
3409 | | APFloat::getAllOnesValue(unsigned BitWidth, bool isIEEE) |
3410 | 0 | { |
3411 | 0 | switch (BitWidth) { |
3412 | 0 | case 16: |
3413 | 0 | return APFloat(IEEEhalf, APInt::getAllOnesValue(BitWidth)); |
3414 | 0 | case 32: |
3415 | 0 | return APFloat(IEEEsingle, APInt::getAllOnesValue(BitWidth)); |
3416 | 0 | case 64: |
3417 | 0 | return APFloat(IEEEdouble, APInt::getAllOnesValue(BitWidth)); |
3418 | 0 | case 80: |
3419 | 0 | return APFloat(x87DoubleExtended, APInt::getAllOnesValue(BitWidth)); |
3420 | 0 | case 128: |
3421 | 0 | if (isIEEE) |
3422 | 0 | return APFloat(IEEEquad, APInt::getAllOnesValue(BitWidth)); |
3423 | 0 | return APFloat(PPCDoubleDouble, APInt::getAllOnesValue(BitWidth)); |
3424 | 0 | default: |
3425 | 0 | llvm_unreachable("Unknown floating bit width"); |
3426 | 0 | } |
3427 | 0 | } |
3428 | | |
3429 | 0 | unsigned APFloat::getSizeInBits(const fltSemantics &Sem) { |
3430 | 0 | return Sem.sizeInBits; |
3431 | 0 | } |
3432 | | |
3433 | | /// Make this number the largest magnitude normal number in the given |
3434 | | /// semantics. |
3435 | 0 | void APFloat::makeLargest(bool Negative) { |
3436 | | // We want (in interchange format): |
3437 | | // sign = {Negative} |
3438 | | // exponent = 1..10 |
3439 | | // significand = 1..1 |
3440 | 0 | category = fcNormal; |
3441 | 0 | sign = Negative; |
3442 | 0 | exponent = semantics->maxExponent; |
3443 | | |
3444 | | // Use memset to set all but the highest integerPart to all ones. |
3445 | 0 | integerPart *significand = significandParts(); |
3446 | 0 | unsigned PartCount = partCount(); |
3447 | 0 | memset(significand, 0xFF, sizeof(integerPart)*(PartCount - 1)); |
3448 | | |
3449 | | // Set the high integerPart especially setting all unused top bits for |
3450 | | // internal consistency. |
3451 | 0 | const unsigned NumUnusedHighBits = |
3452 | 0 | PartCount*integerPartWidth - semantics->precision; |
3453 | 0 | significand[PartCount - 1] = (NumUnusedHighBits < integerPartWidth) |
3454 | 0 | ? (~integerPart(0) >> NumUnusedHighBits) |
3455 | 0 | : 0; |
3456 | 0 | } |
3457 | | |
3458 | | /// Make this number the smallest magnitude denormal number in the given |
3459 | | /// semantics. |
3460 | 0 | void APFloat::makeSmallest(bool Negative) { |
3461 | | // We want (in interchange format): |
3462 | | // sign = {Negative} |
3463 | | // exponent = 0..0 |
3464 | | // significand = 0..01 |
3465 | 0 | category = fcNormal; |
3466 | 0 | sign = Negative; |
3467 | 0 | exponent = semantics->minExponent; |
3468 | 0 | APInt::tcSet(significandParts(), 1, partCount()); |
3469 | 0 | } |
3470 | | |
3471 | | |
3472 | 0 | APFloat APFloat::getLargest(const fltSemantics &Sem, bool Negative) { |
3473 | | // We want (in interchange format): |
3474 | | // sign = {Negative} |
3475 | | // exponent = 1..10 |
3476 | | // significand = 1..1 |
3477 | 0 | APFloat Val(Sem, uninitialized); |
3478 | 0 | Val.makeLargest(Negative); |
3479 | 0 | return Val; |
3480 | 0 | } |
3481 | | |
3482 | 0 | APFloat APFloat::getSmallest(const fltSemantics &Sem, bool Negative) { |
3483 | | // We want (in interchange format): |
3484 | | // sign = {Negative} |
3485 | | // exponent = 0..0 |
3486 | | // significand = 0..01 |
3487 | 0 | APFloat Val(Sem, uninitialized); |
3488 | 0 | Val.makeSmallest(Negative); |
3489 | 0 | return Val; |
3490 | 0 | } |
3491 | | |
3492 | 0 | APFloat APFloat::getSmallestNormalized(const fltSemantics &Sem, bool Negative) { |
3493 | 0 | APFloat Val(Sem, uninitialized); |
3494 | | |
3495 | | // We want (in interchange format): |
3496 | | // sign = {Negative} |
3497 | | // exponent = 0..0 |
3498 | | // significand = 10..0 |
3499 | |
|
3500 | 0 | Val.category = fcNormal; |
3501 | 0 | Val.zeroSignificand(); |
3502 | 0 | Val.sign = Negative; |
3503 | 0 | Val.exponent = Sem.minExponent; |
3504 | 0 | Val.significandParts()[partCountForBits(Sem.precision)-1] |= |
3505 | 0 | (((integerPart) 1) << ((Sem.precision - 1) % integerPartWidth)); |
3506 | |
|
3507 | 0 | return Val; |
3508 | 0 | } |
3509 | | |
3510 | 0 | APFloat::APFloat(const fltSemantics &Sem, const APInt &API) { |
3511 | 0 | initFromAPInt(&Sem, API); |
3512 | 0 | } |
3513 | | |
3514 | 844 | APFloat::APFloat(float f) { |
3515 | 844 | initFromAPInt(&IEEEsingle, APInt::floatToBits(f)); |
3516 | 844 | } |
3517 | | |
3518 | 0 | APFloat::APFloat(double d) { |
3519 | 0 | initFromAPInt(&IEEEdouble, APInt::doubleToBits(d)); |
3520 | 0 | } |
3521 | | |
3522 | | namespace { |
3523 | 0 | void append(SmallVectorImpl<char> &Buffer, StringRef Str) { |
3524 | 0 | Buffer.append(Str.begin(), Str.end()); |
3525 | 0 | } |
3526 | | |
3527 | | /// Removes data from the given significand until it is no more |
3528 | | /// precise than is required for the desired precision. |
3529 | | void AdjustToPrecision(APInt &significand, |
3530 | 0 | int &exp, unsigned FormatPrecision) { |
3531 | 0 | unsigned bits = significand.getActiveBits(); |
3532 | | |
3533 | | // 196/59 is a very slight overestimate of lg_2(10). |
3534 | 0 | unsigned bitsRequired = (FormatPrecision * 196 + 58) / 59; |
3535 | |
|
3536 | 0 | if (bits <= bitsRequired) return; |
3537 | | |
3538 | 0 | unsigned tensRemovable = (bits - bitsRequired) * 59 / 196; |
3539 | 0 | if (!tensRemovable) return; |
3540 | | |
3541 | 0 | exp += tensRemovable; |
3542 | |
|
3543 | 0 | APInt divisor(significand.getBitWidth(), 1); |
3544 | 0 | APInt powten(significand.getBitWidth(), 10); |
3545 | 0 | while (true) { |
3546 | 0 | if (tensRemovable & 1) |
3547 | 0 | divisor *= powten; |
3548 | 0 | tensRemovable >>= 1; |
3549 | 0 | if (!tensRemovable) break; |
3550 | 0 | powten *= powten; |
3551 | 0 | } |
3552 | |
|
3553 | 0 | significand = significand.udiv(divisor); |
3554 | | |
3555 | | // Truncate the significand down to its active bit count. |
3556 | 0 | significand = significand.trunc(significand.getActiveBits()); |
3557 | 0 | } |
3558 | | |
3559 | | |
3560 | | void AdjustToPrecision(SmallVectorImpl<char> &buffer, |
3561 | 0 | int &exp, unsigned FormatPrecision) { |
3562 | 0 | unsigned N = buffer.size(); |
3563 | 0 | if (N <= FormatPrecision) return; |
3564 | | |
3565 | | // The most significant figures are the last ones in the buffer. |
3566 | 0 | unsigned FirstSignificant = N - FormatPrecision; |
3567 | | |
3568 | | // Round. |
3569 | | // FIXME: this probably shouldn't use 'round half up'. |
3570 | | |
3571 | | // Rounding down is just a truncation, except we also want to drop |
3572 | | // trailing zeros from the new result. |
3573 | 0 | if (buffer[FirstSignificant - 1] < '5') { |
3574 | 0 | while (FirstSignificant < N && buffer[FirstSignificant] == '0') |
3575 | 0 | FirstSignificant++; |
3576 | |
|
3577 | 0 | exp += FirstSignificant; |
3578 | 0 | buffer.erase(&buffer[0], &buffer[FirstSignificant]); |
3579 | 0 | return; |
3580 | 0 | } |
3581 | | |
3582 | | // Rounding up requires a decimal add-with-carry. If we continue |
3583 | | // the carry, the newly-introduced zeros will just be truncated. |
3584 | 0 | for (unsigned I = FirstSignificant; I != N; ++I) { |
3585 | 0 | if (buffer[I] == '9') { |
3586 | 0 | FirstSignificant++; |
3587 | 0 | } else { |
3588 | 0 | buffer[I]++; |
3589 | 0 | break; |
3590 | 0 | } |
3591 | 0 | } |
3592 | | |
3593 | | // If we carried through, we have exactly one digit of precision. |
3594 | 0 | if (FirstSignificant == N) { |
3595 | 0 | exp += FirstSignificant; |
3596 | 0 | buffer.clear(); |
3597 | 0 | buffer.push_back('1'); |
3598 | 0 | return; |
3599 | 0 | } |
3600 | | |
3601 | 0 | exp += FirstSignificant; |
3602 | 0 | buffer.erase(&buffer[0], &buffer[FirstSignificant]); |
3603 | 0 | } |
3604 | | } |
3605 | | |
3606 | | void APFloat::toString(SmallVectorImpl<char> &Str, |
3607 | | unsigned FormatPrecision, |
3608 | 0 | unsigned FormatMaxPadding) const { |
3609 | 0 | switch (category) { |
3610 | 0 | case fcInfinity: |
3611 | 0 | if (isNegative()) |
3612 | 0 | return append(Str, "-Inf"); |
3613 | 0 | else |
3614 | 0 | return append(Str, "+Inf"); |
3615 | | |
3616 | 0 | case fcNaN: return append(Str, "NaN"); |
3617 | | |
3618 | 0 | case fcZero: |
3619 | 0 | if (isNegative()) |
3620 | 0 | Str.push_back('-'); |
3621 | |
|
3622 | 0 | if (!FormatMaxPadding) |
3623 | 0 | append(Str, "0.0E+0"); |
3624 | 0 | else |
3625 | 0 | Str.push_back('0'); |
3626 | 0 | return; |
3627 | | |
3628 | 0 | case fcNormal: |
3629 | 0 | break; |
3630 | 0 | } |
3631 | | |
3632 | 0 | if (isNegative()) |
3633 | 0 | Str.push_back('-'); |
3634 | | |
3635 | | // Decompose the number into an APInt and an exponent. |
3636 | 0 | int exp = exponent - ((int) semantics->precision - 1); |
3637 | 0 | APInt significand(semantics->precision, |
3638 | 0 | makeArrayRef(significandParts(), |
3639 | 0 | partCountForBits(semantics->precision))); |
3640 | | |
3641 | | // Set FormatPrecision if zero. We want to do this before we |
3642 | | // truncate trailing zeros, as those are part of the precision. |
3643 | 0 | if (!FormatPrecision) { |
3644 | | // We use enough digits so the number can be round-tripped back to an |
3645 | | // APFloat. The formula comes from "How to Print Floating-Point Numbers |
3646 | | // Accurately" by Steele and White. |
3647 | | // FIXME: Using a formula based purely on the precision is conservative; |
3648 | | // we can print fewer digits depending on the actual value being printed. |
3649 | | |
3650 | | // FormatPrecision = 2 + floor(significandBits / lg_2(10)) |
3651 | 0 | FormatPrecision = 2 + semantics->precision * 59 / 196; |
3652 | 0 | } |
3653 | | |
3654 | | // Ignore trailing binary zeros. |
3655 | 0 | int trailingZeros = significand.countTrailingZeros(); |
3656 | 0 | exp += trailingZeros; |
3657 | 0 | significand = significand.lshr(trailingZeros); |
3658 | | |
3659 | | // Change the exponent from 2^e to 10^e. |
3660 | 0 | if (exp == 0) { |
3661 | | // Nothing to do. |
3662 | 0 | } else if (exp > 0) { |
3663 | | // Just shift left. |
3664 | 0 | significand = significand.zext(semantics->precision + exp); |
3665 | 0 | significand <<= exp; |
3666 | 0 | exp = 0; |
3667 | 0 | } else { /* exp < 0 */ |
3668 | 0 | int texp = -exp; |
3669 | | |
3670 | | // We transform this using the identity: |
3671 | | // (N)(2^-e) == (N)(5^e)(10^-e) |
3672 | | // This means we have to multiply N (the significand) by 5^e. |
3673 | | // To avoid overflow, we have to operate on numbers large |
3674 | | // enough to store N * 5^e: |
3675 | | // log2(N * 5^e) == log2(N) + e * log2(5) |
3676 | | // <= semantics->precision + e * 137 / 59 |
3677 | | // (log_2(5) ~ 2.321928 < 2.322034 ~ 137/59) |
3678 | |
|
3679 | 0 | unsigned precision = semantics->precision + (137 * texp + 136) / 59; |
3680 | | |
3681 | | // Multiply significand by 5^e. |
3682 | | // N * 5^0101 == N * 5^(1*1) * 5^(0*2) * 5^(1*4) * 5^(0*8) |
3683 | 0 | significand = significand.zext(precision); |
3684 | 0 | APInt five_to_the_i(precision, 5); |
3685 | 0 | while (true) { |
3686 | 0 | if (texp & 1) significand *= five_to_the_i; |
3687 | |
|
3688 | 0 | texp >>= 1; |
3689 | 0 | if (!texp) break; |
3690 | 0 | five_to_the_i *= five_to_the_i; |
3691 | 0 | } |
3692 | 0 | } |
3693 | |
|
3694 | 0 | AdjustToPrecision(significand, exp, FormatPrecision); |
3695 | |
|
3696 | 0 | SmallVector<char, 256> buffer; |
3697 | | |
3698 | | // Fill the buffer. |
3699 | 0 | unsigned precision = significand.getBitWidth(); |
3700 | 0 | APInt ten(precision, 10); |
3701 | 0 | APInt digit(precision, 0); |
3702 | |
|
3703 | 0 | bool inTrail = true; |
3704 | 0 | while (significand != 0) { |
3705 | | // digit <- significand % 10 |
3706 | | // significand <- significand / 10 |
3707 | 0 | APInt::udivrem(significand, ten, significand, digit); |
3708 | |
|
3709 | 0 | unsigned d = digit.getZExtValue(); |
3710 | | |
3711 | | // Drop trailing zeros. |
3712 | 0 | if (inTrail && !d) exp++; |
3713 | 0 | else { |
3714 | 0 | buffer.push_back((char) ('0' + d)); |
3715 | 0 | inTrail = false; |
3716 | 0 | } |
3717 | 0 | } |
3718 | |
|
3719 | 0 | assert(!buffer.empty() && "no characters in buffer!"); |
3720 | | |
3721 | | // Drop down to FormatPrecision. |
3722 | | // TODO: don't do more precise calculations above than are required. |
3723 | 0 | AdjustToPrecision(buffer, exp, FormatPrecision); |
3724 | |
|
3725 | 0 | unsigned NDigits = buffer.size(); |
3726 | | |
3727 | | // Check whether we should use scientific notation. |
3728 | 0 | bool FormatScientific; |
3729 | 0 | if (!FormatMaxPadding) |
3730 | 0 | FormatScientific = true; |
3731 | 0 | else { |
3732 | 0 | if (exp >= 0) { |
3733 | | // 765e3 --> 765000 |
3734 | | // ^^^ |
3735 | | // But we shouldn't make the number look more precise than it is. |
3736 | 0 | FormatScientific = ((unsigned) exp > FormatMaxPadding || |
3737 | 0 | NDigits + (unsigned) exp > FormatPrecision); |
3738 | 0 | } else { |
3739 | | // Power of the most significant digit. |
3740 | 0 | int MSD = exp + (int) (NDigits - 1); |
3741 | 0 | if (MSD >= 0) { |
3742 | | // 765e-2 == 7.65 |
3743 | 0 | FormatScientific = false; |
3744 | 0 | } else { |
3745 | | // 765e-5 == 0.00765 |
3746 | | // ^ ^^ |
3747 | 0 | FormatScientific = ((unsigned) -MSD) > FormatMaxPadding; |
3748 | 0 | } |
3749 | 0 | } |
3750 | 0 | } |
3751 | | |
3752 | | // Scientific formatting is pretty straightforward. |
3753 | 0 | if (FormatScientific) { |
3754 | 0 | exp += (NDigits - 1); |
3755 | |
|
3756 | 0 | Str.push_back(buffer[NDigits-1]); |
3757 | 0 | Str.push_back('.'); |
3758 | 0 | if (NDigits == 1) |
3759 | 0 | Str.push_back('0'); |
3760 | 0 | else |
3761 | 0 | for (unsigned I = 1; I != NDigits; ++I) |
3762 | 0 | Str.push_back(buffer[NDigits-1-I]); |
3763 | 0 | Str.push_back('E'); |
3764 | |
|
3765 | 0 | Str.push_back(exp >= 0 ? '+' : '-'); |
3766 | 0 | if (exp < 0) exp = -exp; |
3767 | 0 | SmallVector<char, 6> expbuf; |
3768 | 0 | do { |
3769 | 0 | expbuf.push_back((char) ('0' + (exp % 10))); |
3770 | 0 | exp /= 10; |
3771 | 0 | } while (exp); |
3772 | 0 | for (unsigned I = 0, E = expbuf.size(); I != E; ++I) |
3773 | 0 | Str.push_back(expbuf[E-1-I]); |
3774 | 0 | return; |
3775 | 0 | } |
3776 | | |
3777 | | // Non-scientific, positive exponents. |
3778 | 0 | if (exp >= 0) { |
3779 | 0 | for (unsigned I = 0; I != NDigits; ++I) |
3780 | 0 | Str.push_back(buffer[NDigits-1-I]); |
3781 | 0 | for (unsigned I = 0; I != (unsigned) exp; ++I) |
3782 | 0 | Str.push_back('0'); |
3783 | 0 | return; |
3784 | 0 | } |
3785 | | |
3786 | | // Non-scientific, negative exponents. |
3787 | | |
3788 | | // The number of digits to the left of the decimal point. |
3789 | 0 | int NWholeDigits = exp + (int) NDigits; |
3790 | |
|
3791 | 0 | unsigned I = 0; |
3792 | 0 | if (NWholeDigits > 0) { |
3793 | 0 | for (; I != (unsigned) NWholeDigits; ++I) |
3794 | 0 | Str.push_back(buffer[NDigits-I-1]); |
3795 | 0 | Str.push_back('.'); |
3796 | 0 | } else { |
3797 | 0 | unsigned NZeros = 1 + (unsigned) -NWholeDigits; |
3798 | |
|
3799 | 0 | Str.push_back('0'); |
3800 | 0 | Str.push_back('.'); |
3801 | 0 | for (unsigned Z = 1; Z != NZeros; ++Z) |
3802 | 0 | Str.push_back('0'); |
3803 | 0 | } |
3804 | |
|
3805 | 0 | for (; I != NDigits; ++I) |
3806 | 0 | Str.push_back(buffer[NDigits-I-1]); |
3807 | 0 | } |
3808 | | |
3809 | 0 | bool APFloat::getExactInverse(APFloat *inv) const { |
3810 | | // Special floats and denormals have no exact inverse. |
3811 | 0 | if (!isFiniteNonZero()) |
3812 | 0 | return false; |
3813 | | |
3814 | | // Check that the number is a power of two by making sure that only the |
3815 | | // integer bit is set in the significand. |
3816 | 0 | if (significandLSB() != semantics->precision - 1) |
3817 | 0 | return false; |
3818 | | |
3819 | | // Get the inverse. |
3820 | 0 | APFloat reciprocal(*semantics, 1ULL); |
3821 | 0 | if (reciprocal.divide(*this, rmNearestTiesToEven) != opOK) |
3822 | 0 | return false; |
3823 | | |
3824 | | // Avoid multiplication with a denormal, it is not safe on all platforms and |
3825 | | // may be slower than a normal division. |
3826 | 0 | if (reciprocal.isDenormal()) |
3827 | 0 | return false; |
3828 | | |
3829 | 0 | assert(reciprocal.isFiniteNonZero() && |
3830 | 0 | reciprocal.significandLSB() == reciprocal.semantics->precision - 1); |
3831 | | |
3832 | 0 | if (inv) |
3833 | 0 | *inv = reciprocal; |
3834 | |
|
3835 | 0 | return true; |
3836 | 0 | } |
3837 | | |
3838 | 0 | bool APFloat::isSignaling() const { |
3839 | 0 | if (!isNaN()) |
3840 | 0 | return false; |
3841 | | |
3842 | | // IEEE-754R 2008 6.2.1: A signaling NaN bit string should be encoded with the |
3843 | | // first bit of the trailing significand being 0. |
3844 | 0 | return !APInt::tcExtractBit(significandParts(), semantics->precision - 2); |
3845 | 0 | } |
3846 | | |
3847 | | /// IEEE-754R 2008 5.3.1: nextUp/nextDown. |
3848 | | /// |
3849 | | /// *NOTE* since nextDown(x) = -nextUp(-x), we only implement nextUp with |
3850 | | /// appropriate sign switching before/after the computation. |
3851 | 0 | APFloat::opStatus APFloat::next(bool nextDown) { |
3852 | | // If we are performing nextDown, swap sign so we have -x. |
3853 | 0 | if (nextDown) |
3854 | 0 | changeSign(); |
3855 | | |
3856 | | // Compute nextUp(x) |
3857 | 0 | opStatus result = opOK; |
3858 | | |
3859 | | // Handle each float category separately. |
3860 | 0 | switch (category) { |
3861 | 0 | case fcInfinity: |
3862 | | // nextUp(+inf) = +inf |
3863 | 0 | if (!isNegative()) |
3864 | 0 | break; |
3865 | | // nextUp(-inf) = -getLargest() |
3866 | 0 | makeLargest(true); |
3867 | 0 | break; |
3868 | 0 | case fcNaN: |
3869 | | // IEEE-754R 2008 6.2 Par 2: nextUp(sNaN) = qNaN. Set Invalid flag. |
3870 | | // IEEE-754R 2008 6.2: nextUp(qNaN) = qNaN. Must be identity so we do not |
3871 | | // change the payload. |
3872 | 0 | if (isSignaling()) { |
3873 | 0 | result = opInvalidOp; |
3874 | | // For consistency, propagate the sign of the sNaN to the qNaN. |
3875 | 0 | makeNaN(false, isNegative(), nullptr); |
3876 | 0 | } |
3877 | 0 | break; |
3878 | 0 | case fcZero: |
3879 | | // nextUp(pm 0) = +getSmallest() |
3880 | 0 | makeSmallest(false); |
3881 | 0 | break; |
3882 | 0 | case fcNormal: |
3883 | | // nextUp(-getSmallest()) = -0 |
3884 | 0 | if (isSmallest() && isNegative()) { |
3885 | 0 | APInt::tcSet(significandParts(), 0, partCount()); |
3886 | 0 | category = fcZero; |
3887 | 0 | exponent = 0; |
3888 | 0 | break; |
3889 | 0 | } |
3890 | | |
3891 | | // nextUp(getLargest()) == INFINITY |
3892 | 0 | if (isLargest() && !isNegative()) { |
3893 | 0 | APInt::tcSet(significandParts(), 0, partCount()); |
3894 | 0 | category = fcInfinity; |
3895 | 0 | exponent = semantics->maxExponent + 1; |
3896 | 0 | break; |
3897 | 0 | } |
3898 | | |
3899 | | // nextUp(normal) == normal + inc. |
3900 | 0 | if (isNegative()) { |
3901 | | // If we are negative, we need to decrement the significand. |
3902 | | |
3903 | | // We only cross a binade boundary that requires adjusting the exponent |
3904 | | // if: |
3905 | | // 1. exponent != semantics->minExponent. This implies we are not in the |
3906 | | // smallest binade or are dealing with denormals. |
3907 | | // 2. Our significand excluding the integral bit is all zeros. |
3908 | 0 | bool WillCrossBinadeBoundary = |
3909 | 0 | exponent != semantics->minExponent && isSignificandAllZeros(); |
3910 | | |
3911 | | // Decrement the significand. |
3912 | | // |
3913 | | // We always do this since: |
3914 | | // 1. If we are dealing with a non-binade decrement, by definition we |
3915 | | // just decrement the significand. |
3916 | | // 2. If we are dealing with a normal -> normal binade decrement, since |
3917 | | // we have an explicit integral bit the fact that all bits but the |
3918 | | // integral bit are zero implies that subtracting one will yield a |
3919 | | // significand with 0 integral bit and 1 in all other spots. Thus we |
3920 | | // must just adjust the exponent and set the integral bit to 1. |
3921 | | // 3. If we are dealing with a normal -> denormal binade decrement, |
3922 | | // since we set the integral bit to 0 when we represent denormals, we |
3923 | | // just decrement the significand. |
3924 | 0 | integerPart *Parts = significandParts(); |
3925 | 0 | APInt::tcDecrement(Parts, partCount()); |
3926 | |
|
3927 | 0 | if (WillCrossBinadeBoundary) { |
3928 | | // Our result is a normal number. Do the following: |
3929 | | // 1. Set the integral bit to 1. |
3930 | | // 2. Decrement the exponent. |
3931 | 0 | APInt::tcSetBit(Parts, semantics->precision - 1); |
3932 | 0 | exponent--; |
3933 | 0 | } |
3934 | 0 | } else { |
3935 | | // If we are positive, we need to increment the significand. |
3936 | | |
3937 | | // We only cross a binade boundary that requires adjusting the exponent if |
3938 | | // the input is not a denormal and all of said input's significand bits |
3939 | | // are set. If all of said conditions are true: clear the significand, set |
3940 | | // the integral bit to 1, and increment the exponent. If we have a |
3941 | | // denormal always increment since moving denormals and the numbers in the |
3942 | | // smallest normal binade have the same exponent in our representation. |
3943 | 0 | bool WillCrossBinadeBoundary = !isDenormal() && isSignificandAllOnes(); |
3944 | |
|
3945 | 0 | if (WillCrossBinadeBoundary) { |
3946 | 0 | integerPart *Parts = significandParts(); |
3947 | 0 | APInt::tcSet(Parts, 0, partCount()); |
3948 | 0 | APInt::tcSetBit(Parts, semantics->precision - 1); |
3949 | 0 | assert(exponent != semantics->maxExponent && |
3950 | 0 | "We can not increment an exponent beyond the maxExponent allowed" |
3951 | 0 | " by the given floating point semantics."); |
3952 | 0 | exponent++; |
3953 | 0 | } else { |
3954 | 0 | incrementSignificand(); |
3955 | 0 | } |
3956 | 0 | } |
3957 | 0 | break; |
3958 | 0 | } |
3959 | | |
3960 | | // If we are performing nextDown, swap sign so we have -nextUp(-x) |
3961 | 0 | if (nextDown) |
3962 | 0 | changeSign(); |
3963 | |
|
3964 | 0 | return result; |
3965 | 0 | } |
3966 | | |
3967 | | void |
3968 | 11.5k | APFloat::makeInf(bool Negative) { |
3969 | 11.5k | category = fcInfinity; |
3970 | 11.5k | sign = Negative; |
3971 | 11.5k | exponent = semantics->maxExponent + 1; |
3972 | 11.5k | APInt::tcSet(significandParts(), 0, partCount()); |
3973 | 11.5k | } |
3974 | | |
3975 | | void |
3976 | 767k | APFloat::makeZero(bool Negative) { |
3977 | 767k | category = fcZero; |
3978 | 767k | sign = Negative; |
3979 | 767k | exponent = semantics->minExponent-1; |
3980 | 767k | APInt::tcSet(significandParts(), 0, partCount()); |
3981 | 767k | } |
3982 | | |
3983 | 0 | APFloat llvm_ks::scalbn(APFloat X, int Exp) { |
3984 | 0 | if (X.isInfinity() || X.isZero() || X.isNaN()) |
3985 | 0 | return X; |
3986 | | |
3987 | 0 | auto MaxExp = X.getSemantics().maxExponent; |
3988 | 0 | auto MinExp = X.getSemantics().minExponent; |
3989 | 0 | if (Exp > (MaxExp - X.exponent)) |
3990 | | // Overflow saturates to infinity. |
3991 | 0 | return APFloat::getInf(X.getSemantics(), X.isNegative()); |
3992 | 0 | if (Exp < (MinExp - X.exponent)) |
3993 | | // Underflow saturates to zero. |
3994 | 0 | return APFloat::getZero(X.getSemantics(), X.isNegative()); |
3995 | | |
3996 | 0 | X.exponent += Exp; |
3997 | 0 | return X; |
3998 | 0 | } |