1 | //===-- APInt.cpp - Implement APInt class ---------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements a class to represent arbitrary precision integer |
10 | // constant values and provide a variety of arithmetic operations on them. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/ADT/APInt.h" |
15 | #include "llvm/ADT/ArrayRef.h" |
16 | #include "llvm/ADT/FoldingSet.h" |
17 | #include "llvm/ADT/Hashing.h" |
18 | #include "llvm/ADT/SmallString.h" |
19 | #include "llvm/ADT/StringRef.h" |
20 | #include "llvm/ADT/bit.h" |
21 | #include "llvm/Config/llvm-config.h" |
22 | #include "llvm/Support/Alignment.h" |
23 | #include "llvm/Support/Debug.h" |
24 | #include "llvm/Support/ErrorHandling.h" |
25 | #include "llvm/Support/MathExtras.h" |
26 | #include "llvm/Support/raw_ostream.h" |
27 | #include <cmath> |
28 | #include <optional> |
29 | |
30 | using namespace llvm; |
31 | |
32 | #define DEBUG_TYPE "apint" |
33 | |
34 | /// A utility function for allocating memory, checking for allocation failures, |
35 | /// and ensuring the contents are zeroed. |
36 | inline static uint64_t* getClearedMemory(unsigned numWords) { |
37 | return new uint64_t[numWords](); |
38 | } |
39 | |
40 | /// A utility function for allocating memory and checking for allocation |
41 | /// failure. The content is not zeroed. |
42 | inline static uint64_t* getMemory(unsigned numWords) { |
43 | return new uint64_t[numWords]; |
44 | } |
45 | |
46 | /// A utility function that converts a character to a digit. |
47 | inline static unsigned getDigit(char cdigit, uint8_t radix) { |
48 | unsigned r; |
49 | |
50 | if (radix == 16 || radix == 36) { |
51 | r = cdigit - '0'; |
52 | if (r <= 9) |
53 | return r; |
54 | |
55 | r = cdigit - 'A'; |
56 | if (r <= radix - 11U) |
57 | return r + 10; |
58 | |
59 | r = cdigit - 'a'; |
60 | if (r <= radix - 11U) |
61 | return r + 10; |
62 | |
63 | radix = 10; |
64 | } |
65 | |
66 | r = cdigit - '0'; |
67 | if (r < radix) |
68 | return r; |
69 | |
70 | return UINT_MAX; |
71 | } |
72 | |
73 | |
74 | void APInt::initSlowCase(uint64_t val, bool isSigned) { |
75 | if (isSigned && int64_t(val) < 0) { |
76 | U.pVal = getMemory(numWords: getNumWords()); |
77 | U.pVal[0] = val; |
78 | memset(s: &U.pVal[1], c: 0xFF, n: APINT_WORD_SIZE * (getNumWords() - 1)); |
79 | clearUnusedBits(); |
80 | } else { |
81 | U.pVal = getClearedMemory(numWords: getNumWords()); |
82 | U.pVal[0] = val; |
83 | } |
84 | } |
85 | |
86 | void APInt::initSlowCase(const APInt& that) { |
87 | U.pVal = getMemory(numWords: getNumWords()); |
88 | memcpy(dest: U.pVal, src: that.U.pVal, n: getNumWords() * APINT_WORD_SIZE); |
89 | } |
90 | |
91 | void APInt::initFromArray(ArrayRef<uint64_t> bigVal) { |
92 | assert(bigVal.data() && "Null pointer detected!" ); |
93 | if (isSingleWord()) |
94 | U.VAL = bigVal[0]; |
95 | else { |
96 | // Get memory, cleared to 0 |
97 | U.pVal = getClearedMemory(numWords: getNumWords()); |
98 | // Calculate the number of words to copy |
99 | unsigned words = std::min<unsigned>(a: bigVal.size(), b: getNumWords()); |
100 | // Copy the words from bigVal to pVal |
101 | memcpy(dest: U.pVal, src: bigVal.data(), n: words * APINT_WORD_SIZE); |
102 | } |
103 | // Make sure unused high bits are cleared |
104 | clearUnusedBits(); |
105 | } |
106 | |
107 | APInt::APInt(unsigned numBits, ArrayRef<uint64_t> bigVal) : BitWidth(numBits) { |
108 | initFromArray(bigVal); |
109 | } |
110 | |
111 | APInt::APInt(unsigned numBits, unsigned numWords, const uint64_t bigVal[]) |
112 | : BitWidth(numBits) { |
113 | initFromArray(bigVal: ArrayRef(bigVal, numWords)); |
114 | } |
115 | |
116 | APInt::APInt(unsigned numbits, StringRef Str, uint8_t radix) |
117 | : BitWidth(numbits) { |
118 | fromString(numBits: numbits, str: Str, radix); |
119 | } |
120 | |
121 | void APInt::reallocate(unsigned NewBitWidth) { |
122 | // If the number of words is the same we can just change the width and stop. |
123 | if (getNumWords() == getNumWords(BitWidth: NewBitWidth)) { |
124 | BitWidth = NewBitWidth; |
125 | return; |
126 | } |
127 | |
128 | // If we have an allocation, delete it. |
129 | if (!isSingleWord()) |
130 | delete [] U.pVal; |
131 | |
132 | // Update BitWidth. |
133 | BitWidth = NewBitWidth; |
134 | |
135 | // If we are supposed to have an allocation, create it. |
136 | if (!isSingleWord()) |
137 | U.pVal = getMemory(numWords: getNumWords()); |
138 | } |
139 | |
140 | void APInt::assignSlowCase(const APInt &RHS) { |
141 | // Don't do anything for X = X |
142 | if (this == &RHS) |
143 | return; |
144 | |
145 | // Adjust the bit width and handle allocations as necessary. |
146 | reallocate(NewBitWidth: RHS.getBitWidth()); |
147 | |
148 | // Copy the data. |
149 | if (isSingleWord()) |
150 | U.VAL = RHS.U.VAL; |
151 | else |
152 | memcpy(dest: U.pVal, src: RHS.U.pVal, n: getNumWords() * APINT_WORD_SIZE); |
153 | } |
154 | |
155 | /// This method 'profiles' an APInt for use with FoldingSet. |
156 | void APInt::Profile(FoldingSetNodeID& ID) const { |
157 | ID.AddInteger(I: BitWidth); |
158 | |
159 | if (isSingleWord()) { |
160 | ID.AddInteger(I: U.VAL); |
161 | return; |
162 | } |
163 | |
164 | unsigned NumWords = getNumWords(); |
165 | for (unsigned i = 0; i < NumWords; ++i) |
166 | ID.AddInteger(I: U.pVal[i]); |
167 | } |
168 | |
169 | bool APInt::isAligned(Align A) const { |
170 | if (isZero()) |
171 | return true; |
172 | const unsigned TrailingZeroes = countr_zero(); |
173 | const unsigned MinimumTrailingZeroes = Log2(A); |
174 | return TrailingZeroes >= MinimumTrailingZeroes; |
175 | } |
176 | |
177 | /// Prefix increment operator. Increments the APInt by one. |
178 | APInt& APInt::operator++() { |
179 | if (isSingleWord()) |
180 | ++U.VAL; |
181 | else |
182 | tcIncrement(dst: U.pVal, parts: getNumWords()); |
183 | return clearUnusedBits(); |
184 | } |
185 | |
186 | /// Prefix decrement operator. Decrements the APInt by one. |
187 | APInt& APInt::operator--() { |
188 | if (isSingleWord()) |
189 | --U.VAL; |
190 | else |
191 | tcDecrement(dst: U.pVal, parts: getNumWords()); |
192 | return clearUnusedBits(); |
193 | } |
194 | |
195 | /// Adds the RHS APInt to this APInt. |
196 | /// @returns this, after addition of RHS. |
197 | /// Addition assignment operator. |
198 | APInt& APInt::operator+=(const APInt& RHS) { |
199 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
200 | if (isSingleWord()) |
201 | U.VAL += RHS.U.VAL; |
202 | else |
203 | tcAdd(U.pVal, RHS.U.pVal, carry: 0, getNumWords()); |
204 | return clearUnusedBits(); |
205 | } |
206 | |
207 | APInt& APInt::operator+=(uint64_t RHS) { |
208 | if (isSingleWord()) |
209 | U.VAL += RHS; |
210 | else |
211 | tcAddPart(U.pVal, RHS, getNumWords()); |
212 | return clearUnusedBits(); |
213 | } |
214 | |
215 | /// Subtracts the RHS APInt from this APInt |
216 | /// @returns this, after subtraction |
217 | /// Subtraction assignment operator. |
218 | APInt& APInt::operator-=(const APInt& RHS) { |
219 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
220 | if (isSingleWord()) |
221 | U.VAL -= RHS.U.VAL; |
222 | else |
223 | tcSubtract(U.pVal, RHS.U.pVal, carry: 0, getNumWords()); |
224 | return clearUnusedBits(); |
225 | } |
226 | |
227 | APInt& APInt::operator-=(uint64_t RHS) { |
228 | if (isSingleWord()) |
229 | U.VAL -= RHS; |
230 | else |
231 | tcSubtractPart(U.pVal, RHS, getNumWords()); |
232 | return clearUnusedBits(); |
233 | } |
234 | |
235 | APInt APInt::operator*(const APInt& RHS) const { |
236 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
237 | if (isSingleWord()) |
238 | return APInt(BitWidth, U.VAL * RHS.U.VAL, /*isSigned=*/false, |
239 | /*implicitTrunc=*/true); |
240 | |
241 | APInt Result(getMemory(numWords: getNumWords()), getBitWidth()); |
242 | tcMultiply(Result.U.pVal, U.pVal, RHS.U.pVal, getNumWords()); |
243 | Result.clearUnusedBits(); |
244 | return Result; |
245 | } |
246 | |
247 | void APInt::andAssignSlowCase(const APInt &RHS) { |
248 | WordType *dst = U.pVal, *rhs = RHS.U.pVal; |
249 | for (size_t i = 0, e = getNumWords(); i != e; ++i) |
250 | dst[i] &= rhs[i]; |
251 | } |
252 | |
253 | void APInt::orAssignSlowCase(const APInt &RHS) { |
254 | WordType *dst = U.pVal, *rhs = RHS.U.pVal; |
255 | for (size_t i = 0, e = getNumWords(); i != e; ++i) |
256 | dst[i] |= rhs[i]; |
257 | } |
258 | |
259 | void APInt::xorAssignSlowCase(const APInt &RHS) { |
260 | WordType *dst = U.pVal, *rhs = RHS.U.pVal; |
261 | for (size_t i = 0, e = getNumWords(); i != e; ++i) |
262 | dst[i] ^= rhs[i]; |
263 | } |
264 | |
265 | APInt &APInt::operator*=(const APInt &RHS) { |
266 | *this = *this * RHS; |
267 | return *this; |
268 | } |
269 | |
270 | APInt& APInt::operator*=(uint64_t RHS) { |
271 | if (isSingleWord()) { |
272 | U.VAL *= RHS; |
273 | } else { |
274 | unsigned NumWords = getNumWords(); |
275 | tcMultiplyPart(dst: U.pVal, src: U.pVal, multiplier: RHS, carry: 0, srcParts: NumWords, dstParts: NumWords, add: false); |
276 | } |
277 | return clearUnusedBits(); |
278 | } |
279 | |
280 | bool APInt::equalSlowCase(const APInt &RHS) const { |
281 | return std::equal(first1: U.pVal, last1: U.pVal + getNumWords(), first2: RHS.U.pVal); |
282 | } |
283 | |
284 | int APInt::compare(const APInt& RHS) const { |
285 | assert(BitWidth == RHS.BitWidth && "Bit widths must be same for comparison" ); |
286 | if (isSingleWord()) |
287 | return U.VAL < RHS.U.VAL ? -1 : U.VAL > RHS.U.VAL; |
288 | |
289 | return tcCompare(U.pVal, RHS.U.pVal, getNumWords()); |
290 | } |
291 | |
292 | int APInt::compareSigned(const APInt& RHS) const { |
293 | assert(BitWidth == RHS.BitWidth && "Bit widths must be same for comparison" ); |
294 | if (isSingleWord()) { |
295 | int64_t lhsSext = SignExtend64(X: U.VAL, B: BitWidth); |
296 | int64_t rhsSext = SignExtend64(X: RHS.U.VAL, B: BitWidth); |
297 | return lhsSext < rhsSext ? -1 : lhsSext > rhsSext; |
298 | } |
299 | |
300 | bool lhsNeg = isNegative(); |
301 | bool rhsNeg = RHS.isNegative(); |
302 | |
303 | // If the sign bits don't match, then (LHS < RHS) if LHS is negative |
304 | if (lhsNeg != rhsNeg) |
305 | return lhsNeg ? -1 : 1; |
306 | |
307 | // Otherwise we can just use an unsigned comparison, because even negative |
308 | // numbers compare correctly this way if both have the same signed-ness. |
309 | return tcCompare(U.pVal, RHS.U.pVal, getNumWords()); |
310 | } |
311 | |
312 | void APInt::setBitsSlowCase(unsigned loBit, unsigned hiBit) { |
313 | unsigned loWord = whichWord(bitPosition: loBit); |
314 | unsigned hiWord = whichWord(bitPosition: hiBit); |
315 | |
316 | // Create an initial mask for the low word with zeros below loBit. |
317 | uint64_t loMask = WORDTYPE_MAX << whichBit(bitPosition: loBit); |
318 | |
319 | // If hiBit is not aligned, we need a high mask. |
320 | unsigned hiShiftAmt = whichBit(bitPosition: hiBit); |
321 | if (hiShiftAmt != 0) { |
322 | // Create a high mask with zeros above hiBit. |
323 | uint64_t hiMask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - hiShiftAmt); |
324 | // If loWord and hiWord are equal, then we combine the masks. Otherwise, |
325 | // set the bits in hiWord. |
326 | if (hiWord == loWord) |
327 | loMask &= hiMask; |
328 | else |
329 | U.pVal[hiWord] |= hiMask; |
330 | } |
331 | // Apply the mask to the low word. |
332 | U.pVal[loWord] |= loMask; |
333 | |
334 | // Fill any words between loWord and hiWord with all ones. |
335 | for (unsigned word = loWord + 1; word < hiWord; ++word) |
336 | U.pVal[word] = WORDTYPE_MAX; |
337 | } |
338 | |
339 | void APInt::clearBitsSlowCase(unsigned LoBit, unsigned HiBit) { |
340 | unsigned LoWord = whichWord(bitPosition: LoBit); |
341 | unsigned HiWord = whichWord(bitPosition: HiBit); |
342 | |
343 | // Create an initial mask for the low word with ones below loBit. |
344 | uint64_t LoMask = ~(WORDTYPE_MAX << whichBit(bitPosition: LoBit)); |
345 | |
346 | // If HiBit is not aligned, we need a high mask. |
347 | unsigned HiShiftAmt = whichBit(bitPosition: HiBit); |
348 | if (HiShiftAmt != 0) { |
349 | // Create a high mask with ones above HiBit. |
350 | uint64_t HiMask = ~(WORDTYPE_MAX >> (APINT_BITS_PER_WORD - HiShiftAmt)); |
351 | // If LoWord and HiWord are equal, then we combine the masks. Otherwise, |
352 | // clear the bits in HiWord. |
353 | if (HiWord == LoWord) |
354 | LoMask |= HiMask; |
355 | else |
356 | U.pVal[HiWord] &= HiMask; |
357 | } |
358 | // Apply the mask to the low word. |
359 | U.pVal[LoWord] &= LoMask; |
360 | |
361 | // Fill any words between LoWord and HiWord with all zeros. |
362 | for (unsigned Word = LoWord + 1; Word < HiWord; ++Word) |
363 | U.pVal[Word] = 0; |
364 | } |
365 | |
366 | // Complement a bignum in-place. |
367 | static void tcComplement(APInt::WordType *dst, unsigned parts) { |
368 | for (unsigned i = 0; i < parts; i++) |
369 | dst[i] = ~dst[i]; |
370 | } |
371 | |
372 | /// Toggle every bit to its opposite value. |
373 | void APInt::flipAllBitsSlowCase() { |
374 | tcComplement(dst: U.pVal, parts: getNumWords()); |
375 | clearUnusedBits(); |
376 | } |
377 | |
378 | /// Concatenate the bits from "NewLSB" onto the bottom of *this. This is |
379 | /// equivalent to: |
380 | /// (this->zext(NewWidth) << NewLSB.getBitWidth()) | NewLSB.zext(NewWidth) |
381 | /// In the slow case, we know the result is large. |
382 | APInt APInt::concatSlowCase(const APInt &NewLSB) const { |
383 | unsigned NewWidth = getBitWidth() + NewLSB.getBitWidth(); |
384 | APInt Result = NewLSB.zext(width: NewWidth); |
385 | Result.insertBits(SubBits: *this, bitPosition: NewLSB.getBitWidth()); |
386 | return Result; |
387 | } |
388 | |
389 | /// Toggle a given bit to its opposite value whose position is given |
390 | /// as "bitPosition". |
391 | /// Toggles a given bit to its opposite value. |
392 | void APInt::flipBit(unsigned bitPosition) { |
393 | assert(bitPosition < BitWidth && "Out of the bit-width range!" ); |
394 | setBitVal(BitPosition: bitPosition, BitValue: !(*this)[bitPosition]); |
395 | } |
396 | |
397 | void APInt::insertBits(const APInt &subBits, unsigned bitPosition) { |
398 | unsigned subBitWidth = subBits.getBitWidth(); |
399 | assert((subBitWidth + bitPosition) <= BitWidth && "Illegal bit insertion" ); |
400 | |
401 | // inserting no bits is a noop. |
402 | if (subBitWidth == 0) |
403 | return; |
404 | |
405 | // Insertion is a direct copy. |
406 | if (subBitWidth == BitWidth) { |
407 | *this = subBits; |
408 | return; |
409 | } |
410 | |
411 | // Single word result can be done as a direct bitmask. |
412 | if (isSingleWord()) { |
413 | uint64_t mask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - subBitWidth); |
414 | U.VAL &= ~(mask << bitPosition); |
415 | U.VAL |= (subBits.U.VAL << bitPosition); |
416 | return; |
417 | } |
418 | |
419 | unsigned loBit = whichBit(bitPosition); |
420 | unsigned loWord = whichWord(bitPosition); |
421 | unsigned hi1Word = whichWord(bitPosition: bitPosition + subBitWidth - 1); |
422 | |
423 | // Insertion within a single word can be done as a direct bitmask. |
424 | if (loWord == hi1Word) { |
425 | uint64_t mask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - subBitWidth); |
426 | U.pVal[loWord] &= ~(mask << loBit); |
427 | U.pVal[loWord] |= (subBits.U.VAL << loBit); |
428 | return; |
429 | } |
430 | |
431 | // Insert on word boundaries. |
432 | if (loBit == 0) { |
433 | // Direct copy whole words. |
434 | unsigned numWholeSubWords = subBitWidth / APINT_BITS_PER_WORD; |
435 | memcpy(dest: U.pVal + loWord, src: subBits.getRawData(), |
436 | n: numWholeSubWords * APINT_WORD_SIZE); |
437 | |
438 | // Mask+insert remaining bits. |
439 | unsigned remainingBits = subBitWidth % APINT_BITS_PER_WORD; |
440 | if (remainingBits != 0) { |
441 | uint64_t mask = WORDTYPE_MAX >> (APINT_BITS_PER_WORD - remainingBits); |
442 | U.pVal[hi1Word] &= ~mask; |
443 | U.pVal[hi1Word] |= subBits.getWord(bitPosition: subBitWidth - 1); |
444 | } |
445 | return; |
446 | } |
447 | |
448 | // General case - set/clear individual bits in dst based on src. |
449 | // TODO - there is scope for optimization here, but at the moment this code |
450 | // path is barely used so prefer readability over performance. |
451 | for (unsigned i = 0; i != subBitWidth; ++i) |
452 | setBitVal(BitPosition: bitPosition + i, BitValue: subBits[i]); |
453 | } |
454 | |
455 | void APInt::insertBits(uint64_t subBits, unsigned bitPosition, unsigned numBits) { |
456 | uint64_t maskBits = maskTrailingOnes<uint64_t>(N: numBits); |
457 | subBits &= maskBits; |
458 | if (isSingleWord()) { |
459 | U.VAL &= ~(maskBits << bitPosition); |
460 | U.VAL |= subBits << bitPosition; |
461 | return; |
462 | } |
463 | |
464 | unsigned loBit = whichBit(bitPosition); |
465 | unsigned loWord = whichWord(bitPosition); |
466 | unsigned hiWord = whichWord(bitPosition: bitPosition + numBits - 1); |
467 | if (loWord == hiWord) { |
468 | U.pVal[loWord] &= ~(maskBits << loBit); |
469 | U.pVal[loWord] |= subBits << loBit; |
470 | return; |
471 | } |
472 | |
473 | static_assert(8 * sizeof(WordType) <= 64, "This code assumes only two words affected" ); |
474 | unsigned wordBits = 8 * sizeof(WordType); |
475 | U.pVal[loWord] &= ~(maskBits << loBit); |
476 | U.pVal[loWord] |= subBits << loBit; |
477 | |
478 | U.pVal[hiWord] &= ~(maskBits >> (wordBits - loBit)); |
479 | U.pVal[hiWord] |= subBits >> (wordBits - loBit); |
480 | } |
481 | |
482 | APInt APInt::(unsigned numBits, unsigned bitPosition) const { |
483 | assert(bitPosition < BitWidth && (numBits + bitPosition) <= BitWidth && |
484 | "Illegal bit extraction" ); |
485 | |
486 | if (isSingleWord()) |
487 | return APInt(numBits, U.VAL >> bitPosition, /*isSigned=*/false, |
488 | /*implicitTrunc=*/true); |
489 | |
490 | unsigned loBit = whichBit(bitPosition); |
491 | unsigned loWord = whichWord(bitPosition); |
492 | unsigned hiWord = whichWord(bitPosition: bitPosition + numBits - 1); |
493 | |
494 | // Single word result extracting bits from a single word source. |
495 | if (loWord == hiWord) |
496 | return APInt(numBits, U.pVal[loWord] >> loBit, /*isSigned=*/false, |
497 | /*implicitTrunc=*/true); |
498 | |
499 | // Extracting bits that start on a source word boundary can be done |
500 | // as a fast memory copy. |
501 | if (loBit == 0) |
502 | return APInt(numBits, ArrayRef(U.pVal + loWord, 1 + hiWord - loWord)); |
503 | |
504 | // General case - shift + copy source words directly into place. |
505 | APInt Result(numBits, 0); |
506 | unsigned NumSrcWords = getNumWords(); |
507 | unsigned NumDstWords = Result.getNumWords(); |
508 | |
509 | uint64_t *DestPtr = Result.isSingleWord() ? &Result.U.VAL : Result.U.pVal; |
510 | for (unsigned word = 0; word < NumDstWords; ++word) { |
511 | uint64_t w0 = U.pVal[loWord + word]; |
512 | uint64_t w1 = |
513 | (loWord + word + 1) < NumSrcWords ? U.pVal[loWord + word + 1] : 0; |
514 | DestPtr[word] = (w0 >> loBit) | (w1 << (APINT_BITS_PER_WORD - loBit)); |
515 | } |
516 | |
517 | return Result.clearUnusedBits(); |
518 | } |
519 | |
520 | uint64_t APInt::(unsigned numBits, |
521 | unsigned bitPosition) const { |
522 | assert(bitPosition < BitWidth && (numBits + bitPosition) <= BitWidth && |
523 | "Illegal bit extraction" ); |
524 | assert(numBits <= 64 && "Illegal bit extraction" ); |
525 | |
526 | uint64_t maskBits = maskTrailingOnes<uint64_t>(N: numBits); |
527 | if (isSingleWord()) |
528 | return (U.VAL >> bitPosition) & maskBits; |
529 | |
530 | static_assert(APINT_BITS_PER_WORD >= 64, |
531 | "This code assumes only two words affected" ); |
532 | unsigned loBit = whichBit(bitPosition); |
533 | unsigned loWord = whichWord(bitPosition); |
534 | unsigned hiWord = whichWord(bitPosition: bitPosition + numBits - 1); |
535 | if (loWord == hiWord) |
536 | return (U.pVal[loWord] >> loBit) & maskBits; |
537 | |
538 | uint64_t retBits = U.pVal[loWord] >> loBit; |
539 | retBits |= U.pVal[hiWord] << (APINT_BITS_PER_WORD - loBit); |
540 | retBits &= maskBits; |
541 | return retBits; |
542 | } |
543 | |
544 | unsigned APInt::getSufficientBitsNeeded(StringRef Str, uint8_t Radix) { |
545 | assert(!Str.empty() && "Invalid string length" ); |
546 | size_t StrLen = Str.size(); |
547 | |
548 | // Each computation below needs to know if it's negative. |
549 | unsigned IsNegative = false; |
550 | if (Str[0] == '-' || Str[0] == '+') { |
551 | IsNegative = Str[0] == '-'; |
552 | StrLen--; |
553 | assert(StrLen && "String is only a sign, needs a value." ); |
554 | } |
555 | |
556 | // For radixes of power-of-two values, the bits required is accurately and |
557 | // easily computed. |
558 | if (Radix == 2) |
559 | return StrLen + IsNegative; |
560 | if (Radix == 8) |
561 | return StrLen * 3 + IsNegative; |
562 | if (Radix == 16) |
563 | return StrLen * 4 + IsNegative; |
564 | |
565 | // Compute a sufficient number of bits that is always large enough but might |
566 | // be too large. This avoids the assertion in the constructor. This |
567 | // calculation doesn't work appropriately for the numbers 0-9, so just use 4 |
568 | // bits in that case. |
569 | if (Radix == 10) |
570 | return (StrLen == 1 ? 4 : StrLen * 64 / 18) + IsNegative; |
571 | |
572 | assert(Radix == 36); |
573 | return (StrLen == 1 ? 7 : StrLen * 16 / 3) + IsNegative; |
574 | } |
575 | |
576 | unsigned APInt::getBitsNeeded(StringRef str, uint8_t radix) { |
577 | // Compute a sufficient number of bits that is always large enough but might |
578 | // be too large. |
579 | unsigned sufficient = getSufficientBitsNeeded(Str: str, Radix: radix); |
580 | |
581 | // For bases 2, 8, and 16, the sufficient number of bits is exact and we can |
582 | // return the value directly. For bases 10 and 36, we need to do extra work. |
583 | if (radix == 2 || radix == 8 || radix == 16) |
584 | return sufficient; |
585 | |
586 | // This is grossly inefficient but accurate. We could probably do something |
587 | // with a computation of roughly slen*64/20 and then adjust by the value of |
588 | // the first few digits. But, I'm not sure how accurate that could be. |
589 | size_t slen = str.size(); |
590 | |
591 | // Each computation below needs to know if it's negative. |
592 | StringRef::iterator p = str.begin(); |
593 | unsigned isNegative = *p == '-'; |
594 | if (*p == '-' || *p == '+') { |
595 | p++; |
596 | slen--; |
597 | assert(slen && "String is only a sign, needs a value." ); |
598 | } |
599 | |
600 | |
601 | // Convert to the actual binary value. |
602 | APInt tmp(sufficient, StringRef(p, slen), radix); |
603 | |
604 | // Compute how many bits are required. If the log is infinite, assume we need |
605 | // just bit. If the log is exact and value is negative, then the value is |
606 | // MinSignedValue with (log + 1) bits. |
607 | unsigned log = tmp.logBase2(); |
608 | if (log == (unsigned)-1) { |
609 | return isNegative + 1; |
610 | } else if (isNegative && tmp.isPowerOf2()) { |
611 | return isNegative + log; |
612 | } else { |
613 | return isNegative + log + 1; |
614 | } |
615 | } |
616 | |
617 | hash_code llvm::hash_value(const APInt &Arg) { |
618 | if (Arg.isSingleWord()) |
619 | return hash_combine(args: Arg.BitWidth, args: Arg.U.VAL); |
620 | |
621 | return hash_combine( |
622 | args: Arg.BitWidth, |
623 | args: hash_combine_range(first: Arg.U.pVal, last: Arg.U.pVal + Arg.getNumWords())); |
624 | } |
625 | |
626 | unsigned DenseMapInfo<APInt, void>::getHashValue(const APInt &Key) { |
627 | return static_cast<unsigned>(hash_value(Arg: Key)); |
628 | } |
629 | |
630 | bool APInt::isSplat(unsigned SplatSizeInBits) const { |
631 | assert(getBitWidth() % SplatSizeInBits == 0 && |
632 | "SplatSizeInBits must divide width!" ); |
633 | // We can check that all parts of an integer are equal by making use of a |
634 | // little trick: rotate and check if it's still the same value. |
635 | return *this == rotl(rotateAmt: SplatSizeInBits); |
636 | } |
637 | |
638 | /// This function returns the high "numBits" bits of this APInt. |
639 | APInt APInt::getHiBits(unsigned numBits) const { |
640 | return this->lshr(shiftAmt: BitWidth - numBits); |
641 | } |
642 | |
643 | /// This function returns the low "numBits" bits of this APInt. |
644 | APInt APInt::getLoBits(unsigned numBits) const { |
645 | APInt Result(getLowBitsSet(numBits: BitWidth, loBitsSet: numBits)); |
646 | Result &= *this; |
647 | return Result; |
648 | } |
649 | |
650 | /// Return a value containing V broadcasted over NewLen bits. |
651 | APInt APInt::getSplat(unsigned NewLen, const APInt &V) { |
652 | assert(NewLen >= V.getBitWidth() && "Can't splat to smaller bit width!" ); |
653 | |
654 | APInt Val = V.zext(width: NewLen); |
655 | for (unsigned I = V.getBitWidth(); I < NewLen; I <<= 1) |
656 | Val |= Val << I; |
657 | |
658 | return Val; |
659 | } |
660 | |
661 | unsigned APInt::countLeadingZerosSlowCase() const { |
662 | unsigned Count = 0; |
663 | for (int i = getNumWords()-1; i >= 0; --i) { |
664 | uint64_t V = U.pVal[i]; |
665 | if (V == 0) |
666 | Count += APINT_BITS_PER_WORD; |
667 | else { |
668 | Count += llvm::countl_zero(Val: V); |
669 | break; |
670 | } |
671 | } |
672 | // Adjust for unused bits in the most significant word (they are zero). |
673 | unsigned Mod = BitWidth % APINT_BITS_PER_WORD; |
674 | Count -= Mod > 0 ? APINT_BITS_PER_WORD - Mod : 0; |
675 | return Count; |
676 | } |
677 | |
678 | unsigned APInt::countLeadingOnesSlowCase() const { |
679 | unsigned highWordBits = BitWidth % APINT_BITS_PER_WORD; |
680 | unsigned shift; |
681 | if (!highWordBits) { |
682 | highWordBits = APINT_BITS_PER_WORD; |
683 | shift = 0; |
684 | } else { |
685 | shift = APINT_BITS_PER_WORD - highWordBits; |
686 | } |
687 | int i = getNumWords() - 1; |
688 | unsigned Count = llvm::countl_one(Value: U.pVal[i] << shift); |
689 | if (Count == highWordBits) { |
690 | for (i--; i >= 0; --i) { |
691 | if (U.pVal[i] == WORDTYPE_MAX) |
692 | Count += APINT_BITS_PER_WORD; |
693 | else { |
694 | Count += llvm::countl_one(Value: U.pVal[i]); |
695 | break; |
696 | } |
697 | } |
698 | } |
699 | return Count; |
700 | } |
701 | |
702 | unsigned APInt::countTrailingZerosSlowCase() const { |
703 | unsigned Count = 0; |
704 | unsigned i = 0; |
705 | for (; i < getNumWords() && U.pVal[i] == 0; ++i) |
706 | Count += APINT_BITS_PER_WORD; |
707 | if (i < getNumWords()) |
708 | Count += llvm::countr_zero(Val: U.pVal[i]); |
709 | return std::min(a: Count, b: BitWidth); |
710 | } |
711 | |
712 | unsigned APInt::countTrailingOnesSlowCase() const { |
713 | unsigned Count = 0; |
714 | unsigned i = 0; |
715 | for (; i < getNumWords() && U.pVal[i] == WORDTYPE_MAX; ++i) |
716 | Count += APINT_BITS_PER_WORD; |
717 | if (i < getNumWords()) |
718 | Count += llvm::countr_one(Value: U.pVal[i]); |
719 | assert(Count <= BitWidth); |
720 | return Count; |
721 | } |
722 | |
723 | unsigned APInt::countPopulationSlowCase() const { |
724 | unsigned Count = 0; |
725 | for (unsigned i = 0; i < getNumWords(); ++i) |
726 | Count += llvm::popcount(Value: U.pVal[i]); |
727 | return Count; |
728 | } |
729 | |
730 | bool APInt::intersectsSlowCase(const APInt &RHS) const { |
731 | for (unsigned i = 0, e = getNumWords(); i != e; ++i) |
732 | if ((U.pVal[i] & RHS.U.pVal[i]) != 0) |
733 | return true; |
734 | |
735 | return false; |
736 | } |
737 | |
738 | bool APInt::isSubsetOfSlowCase(const APInt &RHS) const { |
739 | for (unsigned i = 0, e = getNumWords(); i != e; ++i) |
740 | if ((U.pVal[i] & ~RHS.U.pVal[i]) != 0) |
741 | return false; |
742 | |
743 | return true; |
744 | } |
745 | |
746 | APInt APInt::byteSwap() const { |
747 | assert(BitWidth >= 16 && BitWidth % 8 == 0 && "Cannot byteswap!" ); |
748 | if (BitWidth == 16) |
749 | return APInt(BitWidth, llvm::byteswap<uint16_t>(V: U.VAL)); |
750 | if (BitWidth == 32) |
751 | return APInt(BitWidth, llvm::byteswap<uint32_t>(V: U.VAL)); |
752 | if (BitWidth <= 64) { |
753 | uint64_t Tmp1 = llvm::byteswap<uint64_t>(V: U.VAL); |
754 | Tmp1 >>= (64 - BitWidth); |
755 | return APInt(BitWidth, Tmp1); |
756 | } |
757 | |
758 | APInt Result(getNumWords() * APINT_BITS_PER_WORD, 0); |
759 | for (unsigned I = 0, N = getNumWords(); I != N; ++I) |
760 | Result.U.pVal[I] = llvm::byteswap<uint64_t>(V: U.pVal[N - I - 1]); |
761 | if (Result.BitWidth != BitWidth) { |
762 | Result.lshrInPlace(ShiftAmt: Result.BitWidth - BitWidth); |
763 | Result.BitWidth = BitWidth; |
764 | } |
765 | return Result; |
766 | } |
767 | |
768 | APInt APInt::reverseBits() const { |
769 | switch (BitWidth) { |
770 | case 64: |
771 | return APInt(BitWidth, llvm::reverseBits<uint64_t>(Val: U.VAL)); |
772 | case 32: |
773 | return APInt(BitWidth, llvm::reverseBits<uint32_t>(Val: U.VAL)); |
774 | case 16: |
775 | return APInt(BitWidth, llvm::reverseBits<uint16_t>(Val: U.VAL)); |
776 | case 8: |
777 | return APInt(BitWidth, llvm::reverseBits<uint8_t>(Val: U.VAL)); |
778 | case 0: |
779 | return *this; |
780 | default: |
781 | break; |
782 | } |
783 | |
784 | APInt Val(*this); |
785 | APInt Reversed(BitWidth, 0); |
786 | unsigned S = BitWidth; |
787 | |
788 | for (; Val != 0; Val.lshrInPlace(ShiftAmt: 1)) { |
789 | Reversed <<= 1; |
790 | Reversed |= Val[0]; |
791 | --S; |
792 | } |
793 | |
794 | Reversed <<= S; |
795 | return Reversed; |
796 | } |
797 | |
798 | APInt llvm::APIntOps::GreatestCommonDivisor(APInt A, APInt B) { |
799 | // Fast-path a common case. |
800 | if (A == B) return A; |
801 | |
802 | // Corner cases: if either operand is zero, the other is the gcd. |
803 | if (!A) return B; |
804 | if (!B) return A; |
805 | |
806 | // Count common powers of 2 and remove all other powers of 2. |
807 | unsigned Pow2; |
808 | { |
809 | unsigned Pow2_A = A.countr_zero(); |
810 | unsigned Pow2_B = B.countr_zero(); |
811 | if (Pow2_A > Pow2_B) { |
812 | A.lshrInPlace(ShiftAmt: Pow2_A - Pow2_B); |
813 | Pow2 = Pow2_B; |
814 | } else if (Pow2_B > Pow2_A) { |
815 | B.lshrInPlace(ShiftAmt: Pow2_B - Pow2_A); |
816 | Pow2 = Pow2_A; |
817 | } else { |
818 | Pow2 = Pow2_A; |
819 | } |
820 | } |
821 | |
822 | // Both operands are odd multiples of 2^Pow_2: |
823 | // |
824 | // gcd(a, b) = gcd(|a - b| / 2^i, min(a, b)) |
825 | // |
826 | // This is a modified version of Stein's algorithm, taking advantage of |
827 | // efficient countTrailingZeros(). |
828 | while (A != B) { |
829 | if (A.ugt(RHS: B)) { |
830 | A -= B; |
831 | A.lshrInPlace(ShiftAmt: A.countr_zero() - Pow2); |
832 | } else { |
833 | B -= A; |
834 | B.lshrInPlace(ShiftAmt: B.countr_zero() - Pow2); |
835 | } |
836 | } |
837 | |
838 | return A; |
839 | } |
840 | |
841 | APInt llvm::APIntOps::RoundDoubleToAPInt(double Double, unsigned width) { |
842 | uint64_t I = bit_cast<uint64_t>(from: Double); |
843 | |
844 | // Get the sign bit from the highest order bit |
845 | bool isNeg = I >> 63; |
846 | |
847 | // Get the 11-bit exponent and adjust for the 1023 bit bias |
848 | int64_t exp = ((I >> 52) & 0x7ff) - 1023; |
849 | |
850 | // If the exponent is negative, the value is < 0 so just return 0. |
851 | if (exp < 0) |
852 | return APInt(width, 0u); |
853 | |
854 | // Extract the mantissa by clearing the top 12 bits (sign + exponent). |
855 | uint64_t mantissa = (I & (~0ULL >> 12)) | 1ULL << 52; |
856 | |
857 | // If the exponent doesn't shift all bits out of the mantissa |
858 | if (exp < 52) |
859 | return isNeg ? -APInt(width, mantissa >> (52 - exp)) : |
860 | APInt(width, mantissa >> (52 - exp)); |
861 | |
862 | // If the client didn't provide enough bits for us to shift the mantissa into |
863 | // then the result is undefined, just return 0 |
864 | if (width <= exp - 52) |
865 | return APInt(width, 0); |
866 | |
867 | // Otherwise, we have to shift the mantissa bits up to the right location |
868 | APInt Tmp(width, mantissa); |
869 | Tmp <<= (unsigned)exp - 52; |
870 | return isNeg ? -Tmp : Tmp; |
871 | } |
872 | |
873 | /// This function converts this APInt to a double. |
874 | /// The layout for double is as following (IEEE Standard 754): |
875 | /// -------------------------------------- |
876 | /// | Sign Exponent Fraction Bias | |
877 | /// |-------------------------------------- | |
878 | /// | 1[63] 11[62-52] 52[51-00] 1023 | |
879 | /// -------------------------------------- |
880 | double APInt::roundToDouble(bool isSigned) const { |
881 | // Handle the simple case where the value is contained in one uint64_t. |
882 | // It is wrong to optimize getWord(0) to VAL; there might be more than one word. |
883 | if (isSingleWord() || getActiveBits() <= APINT_BITS_PER_WORD) { |
884 | if (isSigned) { |
885 | int64_t sext = SignExtend64(X: getWord(bitPosition: 0), B: BitWidth); |
886 | return double(sext); |
887 | } |
888 | return double(getWord(bitPosition: 0)); |
889 | } |
890 | |
891 | // Determine if the value is negative. |
892 | bool isNeg = isSigned ? (*this)[BitWidth-1] : false; |
893 | |
894 | // Construct the absolute value if we're negative. |
895 | APInt Tmp(isNeg ? -(*this) : (*this)); |
896 | |
897 | // Figure out how many bits we're using. |
898 | unsigned n = Tmp.getActiveBits(); |
899 | |
900 | // The exponent (without bias normalization) is just the number of bits |
901 | // we are using. Note that the sign bit is gone since we constructed the |
902 | // absolute value. |
903 | uint64_t exp = n; |
904 | |
905 | // Return infinity for exponent overflow |
906 | if (exp > 1023) { |
907 | if (!isSigned || !isNeg) |
908 | return std::numeric_limits<double>::infinity(); |
909 | else |
910 | return -std::numeric_limits<double>::infinity(); |
911 | } |
912 | exp += 1023; // Increment for 1023 bias |
913 | |
914 | // Number of bits in mantissa is 52. To obtain the mantissa value, we must |
915 | // extract the high 52 bits from the correct words in pVal. |
916 | uint64_t mantissa; |
917 | unsigned hiWord = whichWord(bitPosition: n-1); |
918 | if (hiWord == 0) { |
919 | mantissa = Tmp.U.pVal[0]; |
920 | if (n > 52) |
921 | mantissa >>= n - 52; // shift down, we want the top 52 bits. |
922 | } else { |
923 | assert(hiWord > 0 && "huh?" ); |
924 | uint64_t hibits = Tmp.U.pVal[hiWord] << (52 - n % APINT_BITS_PER_WORD); |
925 | uint64_t lobits = Tmp.U.pVal[hiWord-1] >> (11 + n % APINT_BITS_PER_WORD); |
926 | mantissa = hibits | lobits; |
927 | } |
928 | |
929 | // The leading bit of mantissa is implicit, so get rid of it. |
930 | uint64_t sign = isNeg ? (1ULL << (APINT_BITS_PER_WORD - 1)) : 0; |
931 | uint64_t I = sign | (exp << 52) | mantissa; |
932 | return bit_cast<double>(from: I); |
933 | } |
934 | |
935 | // Truncate to new width. |
936 | APInt APInt::trunc(unsigned width) const { |
937 | assert(width <= BitWidth && "Invalid APInt Truncate request" ); |
938 | |
939 | if (width <= APINT_BITS_PER_WORD) |
940 | return APInt(width, getRawData()[0], /*isSigned=*/false, |
941 | /*implicitTrunc=*/true); |
942 | |
943 | if (width == BitWidth) |
944 | return *this; |
945 | |
946 | APInt Result(getMemory(numWords: getNumWords(BitWidth: width)), width); |
947 | |
948 | // Copy full words. |
949 | unsigned i; |
950 | for (i = 0; i != width / APINT_BITS_PER_WORD; i++) |
951 | Result.U.pVal[i] = U.pVal[i]; |
952 | |
953 | // Truncate and copy any partial word. |
954 | unsigned bits = (0 - width) % APINT_BITS_PER_WORD; |
955 | if (bits != 0) |
956 | Result.U.pVal[i] = U.pVal[i] << bits >> bits; |
957 | |
958 | return Result; |
959 | } |
960 | |
961 | // Truncate to new width with unsigned saturation. |
962 | APInt APInt::truncUSat(unsigned width) const { |
963 | assert(width <= BitWidth && "Invalid APInt Truncate request" ); |
964 | |
965 | // Can we just losslessly truncate it? |
966 | if (isIntN(N: width)) |
967 | return trunc(width); |
968 | // If not, then just return the new limit. |
969 | return APInt::getMaxValue(numBits: width); |
970 | } |
971 | |
972 | // Truncate to new width with signed saturation. |
973 | APInt APInt::truncSSat(unsigned width) const { |
974 | assert(width <= BitWidth && "Invalid APInt Truncate request" ); |
975 | |
976 | // Can we just losslessly truncate it? |
977 | if (isSignedIntN(N: width)) |
978 | return trunc(width); |
979 | // If not, then just return the new limits. |
980 | return isNegative() ? APInt::getSignedMinValue(numBits: width) |
981 | : APInt::getSignedMaxValue(numBits: width); |
982 | } |
983 | |
984 | // Sign extend to a new width. |
985 | APInt APInt::sext(unsigned Width) const { |
986 | assert(Width >= BitWidth && "Invalid APInt SignExtend request" ); |
987 | |
988 | if (Width <= APINT_BITS_PER_WORD) |
989 | return APInt(Width, SignExtend64(X: U.VAL, B: BitWidth), /*isSigned=*/true); |
990 | |
991 | if (Width == BitWidth) |
992 | return *this; |
993 | |
994 | APInt Result(getMemory(numWords: getNumWords(BitWidth: Width)), Width); |
995 | |
996 | // Copy words. |
997 | std::memcpy(dest: Result.U.pVal, src: getRawData(), n: getNumWords() * APINT_WORD_SIZE); |
998 | |
999 | // Sign extend the last word since there may be unused bits in the input. |
1000 | Result.U.pVal[getNumWords() - 1] = |
1001 | SignExtend64(X: Result.U.pVal[getNumWords() - 1], |
1002 | B: ((BitWidth - 1) % APINT_BITS_PER_WORD) + 1); |
1003 | |
1004 | // Fill with sign bits. |
1005 | std::memset(s: Result.U.pVal + getNumWords(), c: isNegative() ? -1 : 0, |
1006 | n: (Result.getNumWords() - getNumWords()) * APINT_WORD_SIZE); |
1007 | Result.clearUnusedBits(); |
1008 | return Result; |
1009 | } |
1010 | |
1011 | // Zero extend to a new width. |
1012 | APInt APInt::zext(unsigned width) const { |
1013 | assert(width >= BitWidth && "Invalid APInt ZeroExtend request" ); |
1014 | |
1015 | if (width <= APINT_BITS_PER_WORD) |
1016 | return APInt(width, U.VAL); |
1017 | |
1018 | if (width == BitWidth) |
1019 | return *this; |
1020 | |
1021 | APInt Result(getMemory(numWords: getNumWords(BitWidth: width)), width); |
1022 | |
1023 | // Copy words. |
1024 | std::memcpy(dest: Result.U.pVal, src: getRawData(), n: getNumWords() * APINT_WORD_SIZE); |
1025 | |
1026 | // Zero remaining words. |
1027 | std::memset(s: Result.U.pVal + getNumWords(), c: 0, |
1028 | n: (Result.getNumWords() - getNumWords()) * APINT_WORD_SIZE); |
1029 | |
1030 | return Result; |
1031 | } |
1032 | |
1033 | APInt APInt::zextOrTrunc(unsigned width) const { |
1034 | if (BitWidth < width) |
1035 | return zext(width); |
1036 | if (BitWidth > width) |
1037 | return trunc(width); |
1038 | return *this; |
1039 | } |
1040 | |
1041 | APInt APInt::sextOrTrunc(unsigned width) const { |
1042 | if (BitWidth < width) |
1043 | return sext(Width: width); |
1044 | if (BitWidth > width) |
1045 | return trunc(width); |
1046 | return *this; |
1047 | } |
1048 | |
1049 | /// Arithmetic right-shift this APInt by shiftAmt. |
1050 | /// Arithmetic right-shift function. |
1051 | void APInt::ashrInPlace(const APInt &shiftAmt) { |
1052 | ashrInPlace(ShiftAmt: (unsigned)shiftAmt.getLimitedValue(Limit: BitWidth)); |
1053 | } |
1054 | |
1055 | /// Arithmetic right-shift this APInt by shiftAmt. |
1056 | /// Arithmetic right-shift function. |
1057 | void APInt::ashrSlowCase(unsigned ShiftAmt) { |
1058 | // Don't bother performing a no-op shift. |
1059 | if (!ShiftAmt) |
1060 | return; |
1061 | |
1062 | // Save the original sign bit for later. |
1063 | bool Negative = isNegative(); |
1064 | |
1065 | // WordShift is the inter-part shift; BitShift is intra-part shift. |
1066 | unsigned WordShift = ShiftAmt / APINT_BITS_PER_WORD; |
1067 | unsigned BitShift = ShiftAmt % APINT_BITS_PER_WORD; |
1068 | |
1069 | unsigned WordsToMove = getNumWords() - WordShift; |
1070 | if (WordsToMove != 0) { |
1071 | // Sign extend the last word to fill in the unused bits. |
1072 | U.pVal[getNumWords() - 1] = SignExtend64( |
1073 | X: U.pVal[getNumWords() - 1], B: ((BitWidth - 1) % APINT_BITS_PER_WORD) + 1); |
1074 | |
1075 | // Fastpath for moving by whole words. |
1076 | if (BitShift == 0) { |
1077 | std::memmove(dest: U.pVal, src: U.pVal + WordShift, n: WordsToMove * APINT_WORD_SIZE); |
1078 | } else { |
1079 | // Move the words containing significant bits. |
1080 | for (unsigned i = 0; i != WordsToMove - 1; ++i) |
1081 | U.pVal[i] = (U.pVal[i + WordShift] >> BitShift) | |
1082 | (U.pVal[i + WordShift + 1] << (APINT_BITS_PER_WORD - BitShift)); |
1083 | |
1084 | // Handle the last word which has no high bits to copy. Use an arithmetic |
1085 | // shift to preserve the sign bit. |
1086 | U.pVal[WordsToMove - 1] = |
1087 | (int64_t)U.pVal[WordShift + WordsToMove - 1] >> BitShift; |
1088 | } |
1089 | } |
1090 | |
1091 | // Fill in the remainder based on the original sign. |
1092 | std::memset(s: U.pVal + WordsToMove, c: Negative ? -1 : 0, |
1093 | n: WordShift * APINT_WORD_SIZE); |
1094 | clearUnusedBits(); |
1095 | } |
1096 | |
1097 | /// Logical right-shift this APInt by shiftAmt. |
1098 | /// Logical right-shift function. |
1099 | void APInt::lshrInPlace(const APInt &shiftAmt) { |
1100 | lshrInPlace(ShiftAmt: (unsigned)shiftAmt.getLimitedValue(Limit: BitWidth)); |
1101 | } |
1102 | |
1103 | /// Logical right-shift this APInt by shiftAmt. |
1104 | /// Logical right-shift function. |
1105 | void APInt::lshrSlowCase(unsigned ShiftAmt) { |
1106 | tcShiftRight(U.pVal, Words: getNumWords(), Count: ShiftAmt); |
1107 | } |
1108 | |
1109 | /// Left-shift this APInt by shiftAmt. |
1110 | /// Left-shift function. |
1111 | APInt &APInt::operator<<=(const APInt &shiftAmt) { |
1112 | // It's undefined behavior in C to shift by BitWidth or greater. |
1113 | *this <<= (unsigned)shiftAmt.getLimitedValue(Limit: BitWidth); |
1114 | return *this; |
1115 | } |
1116 | |
1117 | void APInt::shlSlowCase(unsigned ShiftAmt) { |
1118 | tcShiftLeft(U.pVal, Words: getNumWords(), Count: ShiftAmt); |
1119 | clearUnusedBits(); |
1120 | } |
1121 | |
1122 | // Calculate the rotate amount modulo the bit width. |
1123 | static unsigned rotateModulo(unsigned BitWidth, const APInt &rotateAmt) { |
1124 | if (LLVM_UNLIKELY(BitWidth == 0)) |
1125 | return 0; |
1126 | unsigned rotBitWidth = rotateAmt.getBitWidth(); |
1127 | APInt rot = rotateAmt; |
1128 | if (rotBitWidth < BitWidth) { |
1129 | // Extend the rotate APInt, so that the urem doesn't divide by 0. |
1130 | // e.g. APInt(1, 32) would give APInt(1, 0). |
1131 | rot = rotateAmt.zext(width: BitWidth); |
1132 | } |
1133 | rot = rot.urem(RHS: APInt(rot.getBitWidth(), BitWidth)); |
1134 | return rot.getLimitedValue(Limit: BitWidth); |
1135 | } |
1136 | |
1137 | APInt APInt::rotl(const APInt &rotateAmt) const { |
1138 | return rotl(rotateAmt: rotateModulo(BitWidth, rotateAmt)); |
1139 | } |
1140 | |
1141 | APInt APInt::rotl(unsigned rotateAmt) const { |
1142 | if (LLVM_UNLIKELY(BitWidth == 0)) |
1143 | return *this; |
1144 | rotateAmt %= BitWidth; |
1145 | if (rotateAmt == 0) |
1146 | return *this; |
1147 | return shl(shiftAmt: rotateAmt) | lshr(shiftAmt: BitWidth - rotateAmt); |
1148 | } |
1149 | |
1150 | APInt APInt::rotr(const APInt &rotateAmt) const { |
1151 | return rotr(rotateAmt: rotateModulo(BitWidth, rotateAmt)); |
1152 | } |
1153 | |
1154 | APInt APInt::rotr(unsigned rotateAmt) const { |
1155 | if (BitWidth == 0) |
1156 | return *this; |
1157 | rotateAmt %= BitWidth; |
1158 | if (rotateAmt == 0) |
1159 | return *this; |
1160 | return lshr(shiftAmt: rotateAmt) | shl(shiftAmt: BitWidth - rotateAmt); |
1161 | } |
1162 | |
1163 | /// \returns the nearest log base 2 of this APInt. Ties round up. |
1164 | /// |
1165 | /// NOTE: When we have a BitWidth of 1, we define: |
1166 | /// |
1167 | /// log2(0) = UINT32_MAX |
1168 | /// log2(1) = 0 |
1169 | /// |
1170 | /// to get around any mathematical concerns resulting from |
1171 | /// referencing 2 in a space where 2 does no exist. |
1172 | unsigned APInt::nearestLogBase2() const { |
1173 | // Special case when we have a bitwidth of 1. If VAL is 1, then we |
1174 | // get 0. If VAL is 0, we get WORDTYPE_MAX which gets truncated to |
1175 | // UINT32_MAX. |
1176 | if (BitWidth == 1) |
1177 | return U.VAL - 1; |
1178 | |
1179 | // Handle the zero case. |
1180 | if (isZero()) |
1181 | return UINT32_MAX; |
1182 | |
1183 | // The non-zero case is handled by computing: |
1184 | // |
1185 | // nearestLogBase2(x) = logBase2(x) + x[logBase2(x)-1]. |
1186 | // |
1187 | // where x[i] is referring to the value of the ith bit of x. |
1188 | unsigned lg = logBase2(); |
1189 | return lg + unsigned((*this)[lg - 1]); |
1190 | } |
1191 | |
1192 | // Square Root - this method computes and returns the square root of "this". |
1193 | // Three mechanisms are used for computation. For small values (<= 5 bits), |
1194 | // a table lookup is done. This gets some performance for common cases. For |
1195 | // values using less than 52 bits, the value is converted to double and then |
1196 | // the libc sqrt function is called. The result is rounded and then converted |
1197 | // back to a uint64_t which is then used to construct the result. Finally, |
1198 | // the Babylonian method for computing square roots is used. |
1199 | APInt APInt::sqrt() const { |
1200 | |
1201 | // Determine the magnitude of the value. |
1202 | unsigned magnitude = getActiveBits(); |
1203 | |
1204 | // Use a fast table for some small values. This also gets rid of some |
1205 | // rounding errors in libc sqrt for small values. |
1206 | if (magnitude <= 5) { |
1207 | static const uint8_t results[32] = { |
1208 | /* 0 */ 0, |
1209 | /* 1- 2 */ 1, 1, |
1210 | /* 3- 6 */ 2, 2, 2, 2, |
1211 | /* 7-12 */ 3, 3, 3, 3, 3, 3, |
1212 | /* 13-20 */ 4, 4, 4, 4, 4, 4, 4, 4, |
1213 | /* 21-30 */ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, |
1214 | /* 31 */ 6 |
1215 | }; |
1216 | return APInt(BitWidth, results[ (isSingleWord() ? U.VAL : U.pVal[0]) ]); |
1217 | } |
1218 | |
1219 | // If the magnitude of the value fits in less than 52 bits (the precision of |
1220 | // an IEEE double precision floating point value), then we can use the |
1221 | // libc sqrt function which will probably use a hardware sqrt computation. |
1222 | // This should be faster than the algorithm below. |
1223 | if (magnitude < 52) { |
1224 | return APInt(BitWidth, |
1225 | uint64_t(::round(x: ::sqrt(x: double(isSingleWord() ? U.VAL |
1226 | : U.pVal[0]))))); |
1227 | } |
1228 | |
1229 | // Okay, all the short cuts are exhausted. We must compute it. The following |
1230 | // is a classical Babylonian method for computing the square root. This code |
1231 | // was adapted to APInt from a wikipedia article on such computations. |
1232 | // See http://www.wikipedia.org/ and go to the page named |
1233 | // Calculate_an_integer_square_root. |
1234 | unsigned nbits = BitWidth, i = 4; |
1235 | APInt testy(BitWidth, 16); |
1236 | APInt x_old(BitWidth, 1); |
1237 | APInt x_new(BitWidth, 0); |
1238 | APInt two(BitWidth, 2); |
1239 | |
1240 | // Select a good starting value using binary logarithms. |
1241 | for (;; i += 2, testy = testy.shl(shiftAmt: 2)) |
1242 | if (i >= nbits || this->ule(RHS: testy)) { |
1243 | x_old = x_old.shl(shiftAmt: i / 2); |
1244 | break; |
1245 | } |
1246 | |
1247 | // Use the Babylonian method to arrive at the integer square root: |
1248 | for (;;) { |
1249 | x_new = (this->udiv(RHS: x_old) + x_old).udiv(RHS: two); |
1250 | if (x_old.ule(RHS: x_new)) |
1251 | break; |
1252 | x_old = x_new; |
1253 | } |
1254 | |
1255 | // Make sure we return the closest approximation |
1256 | // NOTE: The rounding calculation below is correct. It will produce an |
1257 | // off-by-one discrepancy with results from pari/gp. That discrepancy has been |
1258 | // determined to be a rounding issue with pari/gp as it begins to use a |
1259 | // floating point representation after 192 bits. There are no discrepancies |
1260 | // between this algorithm and pari/gp for bit widths < 192 bits. |
1261 | APInt square(x_old * x_old); |
1262 | APInt nextSquare((x_old + 1) * (x_old +1)); |
1263 | if (this->ult(RHS: square)) |
1264 | return x_old; |
1265 | assert(this->ule(nextSquare) && "Error in APInt::sqrt computation" ); |
1266 | APInt midpoint((nextSquare - square).udiv(RHS: two)); |
1267 | APInt offset(*this - square); |
1268 | if (offset.ult(RHS: midpoint)) |
1269 | return x_old; |
1270 | return x_old + 1; |
1271 | } |
1272 | |
1273 | /// \returns the multiplicative inverse of an odd APInt modulo 2^BitWidth. |
1274 | APInt APInt::multiplicativeInverse() const { |
1275 | assert((*this)[0] && |
1276 | "multiplicative inverse is only defined for odd numbers!" ); |
1277 | |
1278 | // Use Newton's method. |
1279 | APInt Factor = *this; |
1280 | APInt T; |
1281 | while (!(T = *this * Factor).isOne()) |
1282 | Factor *= 2 - std::move(T); |
1283 | return Factor; |
1284 | } |
1285 | |
1286 | /// Implementation of Knuth's Algorithm D (Division of nonnegative integers) |
1287 | /// from "Art of Computer Programming, Volume 2", section 4.3.1, p. 272. The |
1288 | /// variables here have the same names as in the algorithm. Comments explain |
1289 | /// the algorithm and any deviation from it. |
1290 | static void KnuthDiv(uint32_t *u, uint32_t *v, uint32_t *q, uint32_t* r, |
1291 | unsigned m, unsigned n) { |
1292 | assert(u && "Must provide dividend" ); |
1293 | assert(v && "Must provide divisor" ); |
1294 | assert(q && "Must provide quotient" ); |
1295 | assert(u != v && u != q && v != q && "Must use different memory" ); |
1296 | assert(n>1 && "n must be > 1" ); |
1297 | |
1298 | // b denotes the base of the number system. In our case b is 2^32. |
1299 | const uint64_t b = uint64_t(1) << 32; |
1300 | |
1301 | // The DEBUG macros here tend to be spam in the debug output if you're not |
1302 | // debugging this code. Disable them unless KNUTH_DEBUG is defined. |
1303 | #ifdef KNUTH_DEBUG |
1304 | #define DEBUG_KNUTH(X) LLVM_DEBUG(X) |
1305 | #else |
1306 | #define DEBUG_KNUTH(X) do {} while(false) |
1307 | #endif |
1308 | |
1309 | DEBUG_KNUTH(dbgs() << "KnuthDiv: m=" << m << " n=" << n << '\n'); |
1310 | DEBUG_KNUTH(dbgs() << "KnuthDiv: original:" ); |
1311 | DEBUG_KNUTH(for (int i = m + n; i >= 0; i--) dbgs() << " " << u[i]); |
1312 | DEBUG_KNUTH(dbgs() << " by" ); |
1313 | DEBUG_KNUTH(for (int i = n; i > 0; i--) dbgs() << " " << v[i - 1]); |
1314 | DEBUG_KNUTH(dbgs() << '\n'); |
1315 | // D1. [Normalize.] Set d = b / (v[n-1] + 1) and multiply all the digits of |
1316 | // u and v by d. Note that we have taken Knuth's advice here to use a power |
1317 | // of 2 value for d such that d * v[n-1] >= b/2 (b is the base). A power of |
1318 | // 2 allows us to shift instead of multiply and it is easy to determine the |
1319 | // shift amount from the leading zeros. We are basically normalizing the u |
1320 | // and v so that its high bits are shifted to the top of v's range without |
1321 | // overflow. Note that this can require an extra word in u so that u must |
1322 | // be of length m+n+1. |
1323 | unsigned shift = llvm::countl_zero(Val: v[n - 1]); |
1324 | uint32_t v_carry = 0; |
1325 | uint32_t u_carry = 0; |
1326 | if (shift) { |
1327 | for (unsigned i = 0; i < m+n; ++i) { |
1328 | uint32_t u_tmp = u[i] >> (32 - shift); |
1329 | u[i] = (u[i] << shift) | u_carry; |
1330 | u_carry = u_tmp; |
1331 | } |
1332 | for (unsigned i = 0; i < n; ++i) { |
1333 | uint32_t v_tmp = v[i] >> (32 - shift); |
1334 | v[i] = (v[i] << shift) | v_carry; |
1335 | v_carry = v_tmp; |
1336 | } |
1337 | } |
1338 | u[m+n] = u_carry; |
1339 | |
1340 | DEBUG_KNUTH(dbgs() << "KnuthDiv: normal:" ); |
1341 | DEBUG_KNUTH(for (int i = m + n; i >= 0; i--) dbgs() << " " << u[i]); |
1342 | DEBUG_KNUTH(dbgs() << " by" ); |
1343 | DEBUG_KNUTH(for (int i = n; i > 0; i--) dbgs() << " " << v[i - 1]); |
1344 | DEBUG_KNUTH(dbgs() << '\n'); |
1345 | |
1346 | // D2. [Initialize j.] Set j to m. This is the loop counter over the places. |
1347 | int j = m; |
1348 | do { |
1349 | DEBUG_KNUTH(dbgs() << "KnuthDiv: quotient digit #" << j << '\n'); |
1350 | // D3. [Calculate q'.]. |
1351 | // Set qp = (u[j+n]*b + u[j+n-1]) / v[n-1]. (qp=qprime=q') |
1352 | // Set rp = (u[j+n]*b + u[j+n-1]) % v[n-1]. (rp=rprime=r') |
1353 | // Now test if qp == b or qp*v[n-2] > b*rp + u[j+n-2]; if so, decrease |
1354 | // qp by 1, increase rp by v[n-1], and repeat this test if rp < b. The test |
1355 | // on v[n-2] determines at high speed most of the cases in which the trial |
1356 | // value qp is one too large, and it eliminates all cases where qp is two |
1357 | // too large. |
1358 | uint64_t dividend = Make_64(High: u[j+n], Low: u[j+n-1]); |
1359 | DEBUG_KNUTH(dbgs() << "KnuthDiv: dividend == " << dividend << '\n'); |
1360 | uint64_t qp = dividend / v[n-1]; |
1361 | uint64_t rp = dividend % v[n-1]; |
1362 | if (qp == b || qp*v[n-2] > b*rp + u[j+n-2]) { |
1363 | qp--; |
1364 | rp += v[n-1]; |
1365 | if (rp < b && (qp == b || qp*v[n-2] > b*rp + u[j+n-2])) |
1366 | qp--; |
1367 | } |
1368 | DEBUG_KNUTH(dbgs() << "KnuthDiv: qp == " << qp << ", rp == " << rp << '\n'); |
1369 | |
1370 | // D4. [Multiply and subtract.] Replace (u[j+n]u[j+n-1]...u[j]) with |
1371 | // (u[j+n]u[j+n-1]..u[j]) - qp * (v[n-1]...v[1]v[0]). This computation |
1372 | // consists of a simple multiplication by a one-place number, combined with |
1373 | // a subtraction. |
1374 | // The digits (u[j+n]...u[j]) should be kept positive; if the result of |
1375 | // this step is actually negative, (u[j+n]...u[j]) should be left as the |
1376 | // true value plus b**(n+1), namely as the b's complement of |
1377 | // the true value, and a "borrow" to the left should be remembered. |
1378 | int64_t borrow = 0; |
1379 | for (unsigned i = 0; i < n; ++i) { |
1380 | uint64_t p = uint64_t(qp) * uint64_t(v[i]); |
1381 | int64_t subres = int64_t(u[j+i]) - borrow - Lo_32(Value: p); |
1382 | u[j+i] = Lo_32(Value: subres); |
1383 | borrow = Hi_32(Value: p) - Hi_32(Value: subres); |
1384 | DEBUG_KNUTH(dbgs() << "KnuthDiv: u[j+i] = " << u[j + i] |
1385 | << ", borrow = " << borrow << '\n'); |
1386 | } |
1387 | bool isNeg = u[j+n] < borrow; |
1388 | u[j+n] -= Lo_32(Value: borrow); |
1389 | |
1390 | DEBUG_KNUTH(dbgs() << "KnuthDiv: after subtraction:" ); |
1391 | DEBUG_KNUTH(for (int i = m + n; i >= 0; i--) dbgs() << " " << u[i]); |
1392 | DEBUG_KNUTH(dbgs() << '\n'); |
1393 | |
1394 | // D5. [Test remainder.] Set q[j] = qp. If the result of step D4 was |
1395 | // negative, go to step D6; otherwise go on to step D7. |
1396 | q[j] = Lo_32(Value: qp); |
1397 | if (isNeg) { |
1398 | // D6. [Add back]. The probability that this step is necessary is very |
1399 | // small, on the order of only 2/b. Make sure that test data accounts for |
1400 | // this possibility. Decrease q[j] by 1 |
1401 | q[j]--; |
1402 | // and add (0v[n-1]...v[1]v[0]) to (u[j+n]u[j+n-1]...u[j+1]u[j]). |
1403 | // A carry will occur to the left of u[j+n], and it should be ignored |
1404 | // since it cancels with the borrow that occurred in D4. |
1405 | bool carry = false; |
1406 | for (unsigned i = 0; i < n; i++) { |
1407 | uint32_t limit = std::min(a: u[j+i],b: v[i]); |
1408 | u[j+i] += v[i] + carry; |
1409 | carry = u[j+i] < limit || (carry && u[j+i] == limit); |
1410 | } |
1411 | u[j+n] += carry; |
1412 | } |
1413 | DEBUG_KNUTH(dbgs() << "KnuthDiv: after correction:" ); |
1414 | DEBUG_KNUTH(for (int i = m + n; i >= 0; i--) dbgs() << " " << u[i]); |
1415 | DEBUG_KNUTH(dbgs() << "\nKnuthDiv: digit result = " << q[j] << '\n'); |
1416 | |
1417 | // D7. [Loop on j.] Decrease j by one. Now if j >= 0, go back to D3. |
1418 | } while (--j >= 0); |
1419 | |
1420 | DEBUG_KNUTH(dbgs() << "KnuthDiv: quotient:" ); |
1421 | DEBUG_KNUTH(for (int i = m; i >= 0; i--) dbgs() << " " << q[i]); |
1422 | DEBUG_KNUTH(dbgs() << '\n'); |
1423 | |
1424 | // D8. [Unnormalize]. Now q[...] is the desired quotient, and the desired |
1425 | // remainder may be obtained by dividing u[...] by d. If r is non-null we |
1426 | // compute the remainder (urem uses this). |
1427 | if (r) { |
1428 | // The value d is expressed by the "shift" value above since we avoided |
1429 | // multiplication by d by using a shift left. So, all we have to do is |
1430 | // shift right here. |
1431 | if (shift) { |
1432 | uint32_t carry = 0; |
1433 | DEBUG_KNUTH(dbgs() << "KnuthDiv: remainder:" ); |
1434 | for (int i = n-1; i >= 0; i--) { |
1435 | r[i] = (u[i] >> shift) | carry; |
1436 | carry = u[i] << (32 - shift); |
1437 | DEBUG_KNUTH(dbgs() << " " << r[i]); |
1438 | } |
1439 | } else { |
1440 | for (int i = n-1; i >= 0; i--) { |
1441 | r[i] = u[i]; |
1442 | DEBUG_KNUTH(dbgs() << " " << r[i]); |
1443 | } |
1444 | } |
1445 | DEBUG_KNUTH(dbgs() << '\n'); |
1446 | } |
1447 | DEBUG_KNUTH(dbgs() << '\n'); |
1448 | } |
1449 | |
1450 | void APInt::divide(const WordType *LHS, unsigned lhsWords, const WordType *RHS, |
1451 | unsigned rhsWords, WordType *Quotient, WordType *Remainder) { |
1452 | assert(lhsWords >= rhsWords && "Fractional result" ); |
1453 | |
1454 | // First, compose the values into an array of 32-bit words instead of |
1455 | // 64-bit words. This is a necessity of both the "short division" algorithm |
1456 | // and the Knuth "classical algorithm" which requires there to be native |
1457 | // operations for +, -, and * on an m bit value with an m*2 bit result. We |
1458 | // can't use 64-bit operands here because we don't have native results of |
1459 | // 128-bits. Furthermore, casting the 64-bit values to 32-bit values won't |
1460 | // work on large-endian machines. |
1461 | unsigned n = rhsWords * 2; |
1462 | unsigned m = (lhsWords * 2) - n; |
1463 | |
1464 | // Allocate space for the temporary values we need either on the stack, if |
1465 | // it will fit, or on the heap if it won't. |
1466 | uint32_t SPACE[128]; |
1467 | uint32_t *U = nullptr; |
1468 | uint32_t *V = nullptr; |
1469 | uint32_t *Q = nullptr; |
1470 | uint32_t *R = nullptr; |
1471 | if ((Remainder?4:3)*n+2*m+1 <= 128) { |
1472 | U = &SPACE[0]; |
1473 | V = &SPACE[m+n+1]; |
1474 | Q = &SPACE[(m+n+1) + n]; |
1475 | if (Remainder) |
1476 | R = &SPACE[(m+n+1) + n + (m+n)]; |
1477 | } else { |
1478 | U = new uint32_t[m + n + 1]; |
1479 | V = new uint32_t[n]; |
1480 | Q = new uint32_t[m+n]; |
1481 | if (Remainder) |
1482 | R = new uint32_t[n]; |
1483 | } |
1484 | |
1485 | // Initialize the dividend |
1486 | memset(s: U, c: 0, n: (m+n+1)*sizeof(uint32_t)); |
1487 | for (unsigned i = 0; i < lhsWords; ++i) { |
1488 | uint64_t tmp = LHS[i]; |
1489 | U[i * 2] = Lo_32(Value: tmp); |
1490 | U[i * 2 + 1] = Hi_32(Value: tmp); |
1491 | } |
1492 | U[m+n] = 0; // this extra word is for "spill" in the Knuth algorithm. |
1493 | |
1494 | // Initialize the divisor |
1495 | memset(s: V, c: 0, n: (n)*sizeof(uint32_t)); |
1496 | for (unsigned i = 0; i < rhsWords; ++i) { |
1497 | uint64_t tmp = RHS[i]; |
1498 | V[i * 2] = Lo_32(Value: tmp); |
1499 | V[i * 2 + 1] = Hi_32(Value: tmp); |
1500 | } |
1501 | |
1502 | // initialize the quotient and remainder |
1503 | memset(s: Q, c: 0, n: (m+n) * sizeof(uint32_t)); |
1504 | if (Remainder) |
1505 | memset(s: R, c: 0, n: n * sizeof(uint32_t)); |
1506 | |
1507 | // Now, adjust m and n for the Knuth division. n is the number of words in |
1508 | // the divisor. m is the number of words by which the dividend exceeds the |
1509 | // divisor (i.e. m+n is the length of the dividend). These sizes must not |
1510 | // contain any zero words or the Knuth algorithm fails. |
1511 | for (unsigned i = n; i > 0 && V[i-1] == 0; i--) { |
1512 | n--; |
1513 | m++; |
1514 | } |
1515 | for (unsigned i = m+n; i > 0 && U[i-1] == 0; i--) |
1516 | m--; |
1517 | |
1518 | // If we're left with only a single word for the divisor, Knuth doesn't work |
1519 | // so we implement the short division algorithm here. This is much simpler |
1520 | // and faster because we are certain that we can divide a 64-bit quantity |
1521 | // by a 32-bit quantity at hardware speed and short division is simply a |
1522 | // series of such operations. This is just like doing short division but we |
1523 | // are using base 2^32 instead of base 10. |
1524 | assert(n != 0 && "Divide by zero?" ); |
1525 | if (n == 1) { |
1526 | uint32_t divisor = V[0]; |
1527 | uint32_t remainder = 0; |
1528 | for (int i = m; i >= 0; i--) { |
1529 | uint64_t partial_dividend = Make_64(High: remainder, Low: U[i]); |
1530 | if (partial_dividend == 0) { |
1531 | Q[i] = 0; |
1532 | remainder = 0; |
1533 | } else if (partial_dividend < divisor) { |
1534 | Q[i] = 0; |
1535 | remainder = Lo_32(Value: partial_dividend); |
1536 | } else if (partial_dividend == divisor) { |
1537 | Q[i] = 1; |
1538 | remainder = 0; |
1539 | } else { |
1540 | Q[i] = Lo_32(Value: partial_dividend / divisor); |
1541 | remainder = Lo_32(Value: partial_dividend - (Q[i] * divisor)); |
1542 | } |
1543 | } |
1544 | if (R) |
1545 | R[0] = remainder; |
1546 | } else { |
1547 | // Now we're ready to invoke the Knuth classical divide algorithm. In this |
1548 | // case n > 1. |
1549 | KnuthDiv(u: U, v: V, q: Q, r: R, m, n); |
1550 | } |
1551 | |
1552 | // If the caller wants the quotient |
1553 | if (Quotient) { |
1554 | for (unsigned i = 0; i < lhsWords; ++i) |
1555 | Quotient[i] = Make_64(High: Q[i*2+1], Low: Q[i*2]); |
1556 | } |
1557 | |
1558 | // If the caller wants the remainder |
1559 | if (Remainder) { |
1560 | for (unsigned i = 0; i < rhsWords; ++i) |
1561 | Remainder[i] = Make_64(High: R[i*2+1], Low: R[i*2]); |
1562 | } |
1563 | |
1564 | // Clean up the memory we allocated. |
1565 | if (U != &SPACE[0]) { |
1566 | delete [] U; |
1567 | delete [] V; |
1568 | delete [] Q; |
1569 | delete [] R; |
1570 | } |
1571 | } |
1572 | |
1573 | APInt APInt::udiv(const APInt &RHS) const { |
1574 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
1575 | |
1576 | // First, deal with the easy case |
1577 | if (isSingleWord()) { |
1578 | assert(RHS.U.VAL != 0 && "Divide by zero?" ); |
1579 | return APInt(BitWidth, U.VAL / RHS.U.VAL); |
1580 | } |
1581 | |
1582 | // Get some facts about the LHS and RHS number of bits and words |
1583 | unsigned lhsWords = getNumWords(BitWidth: getActiveBits()); |
1584 | unsigned rhsBits = RHS.getActiveBits(); |
1585 | unsigned rhsWords = getNumWords(BitWidth: rhsBits); |
1586 | assert(rhsWords && "Divided by zero???" ); |
1587 | |
1588 | // Deal with some degenerate cases |
1589 | if (!lhsWords) |
1590 | // 0 / X ===> 0 |
1591 | return APInt(BitWidth, 0); |
1592 | if (rhsBits == 1) |
1593 | // X / 1 ===> X |
1594 | return *this; |
1595 | if (lhsWords < rhsWords || this->ult(RHS)) |
1596 | // X / Y ===> 0, iff X < Y |
1597 | return APInt(BitWidth, 0); |
1598 | if (*this == RHS) |
1599 | // X / X ===> 1 |
1600 | return APInt(BitWidth, 1); |
1601 | if (lhsWords == 1) // rhsWords is 1 if lhsWords is 1. |
1602 | // All high words are zero, just use native divide |
1603 | return APInt(BitWidth, this->U.pVal[0] / RHS.U.pVal[0]); |
1604 | |
1605 | // We have to compute it the hard way. Invoke the Knuth divide algorithm. |
1606 | APInt Quotient(BitWidth, 0); // to hold result. |
1607 | divide(LHS: U.pVal, lhsWords, RHS: RHS.U.pVal, rhsWords, Quotient: Quotient.U.pVal, Remainder: nullptr); |
1608 | return Quotient; |
1609 | } |
1610 | |
1611 | APInt APInt::udiv(uint64_t RHS) const { |
1612 | assert(RHS != 0 && "Divide by zero?" ); |
1613 | |
1614 | // First, deal with the easy case |
1615 | if (isSingleWord()) |
1616 | return APInt(BitWidth, U.VAL / RHS); |
1617 | |
1618 | // Get some facts about the LHS words. |
1619 | unsigned lhsWords = getNumWords(BitWidth: getActiveBits()); |
1620 | |
1621 | // Deal with some degenerate cases |
1622 | if (!lhsWords) |
1623 | // 0 / X ===> 0 |
1624 | return APInt(BitWidth, 0); |
1625 | if (RHS == 1) |
1626 | // X / 1 ===> X |
1627 | return *this; |
1628 | if (this->ult(RHS)) |
1629 | // X / Y ===> 0, iff X < Y |
1630 | return APInt(BitWidth, 0); |
1631 | if (*this == RHS) |
1632 | // X / X ===> 1 |
1633 | return APInt(BitWidth, 1); |
1634 | if (lhsWords == 1) // rhsWords is 1 if lhsWords is 1. |
1635 | // All high words are zero, just use native divide |
1636 | return APInt(BitWidth, this->U.pVal[0] / RHS); |
1637 | |
1638 | // We have to compute it the hard way. Invoke the Knuth divide algorithm. |
1639 | APInt Quotient(BitWidth, 0); // to hold result. |
1640 | divide(LHS: U.pVal, lhsWords, RHS: &RHS, rhsWords: 1, Quotient: Quotient.U.pVal, Remainder: nullptr); |
1641 | return Quotient; |
1642 | } |
1643 | |
1644 | APInt APInt::sdiv(const APInt &RHS) const { |
1645 | if (isNegative()) { |
1646 | if (RHS.isNegative()) |
1647 | return (-(*this)).udiv(RHS: -RHS); |
1648 | return -((-(*this)).udiv(RHS)); |
1649 | } |
1650 | if (RHS.isNegative()) |
1651 | return -(this->udiv(RHS: -RHS)); |
1652 | return this->udiv(RHS); |
1653 | } |
1654 | |
1655 | APInt APInt::sdiv(int64_t RHS) const { |
1656 | if (isNegative()) { |
1657 | if (RHS < 0) |
1658 | return (-(*this)).udiv(RHS: -RHS); |
1659 | return -((-(*this)).udiv(RHS)); |
1660 | } |
1661 | if (RHS < 0) |
1662 | return -(this->udiv(RHS: -RHS)); |
1663 | return this->udiv(RHS); |
1664 | } |
1665 | |
1666 | APInt APInt::urem(const APInt &RHS) const { |
1667 | assert(BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
1668 | if (isSingleWord()) { |
1669 | assert(RHS.U.VAL != 0 && "Remainder by zero?" ); |
1670 | return APInt(BitWidth, U.VAL % RHS.U.VAL); |
1671 | } |
1672 | |
1673 | // Get some facts about the LHS |
1674 | unsigned lhsWords = getNumWords(BitWidth: getActiveBits()); |
1675 | |
1676 | // Get some facts about the RHS |
1677 | unsigned rhsBits = RHS.getActiveBits(); |
1678 | unsigned rhsWords = getNumWords(BitWidth: rhsBits); |
1679 | assert(rhsWords && "Performing remainder operation by zero ???" ); |
1680 | |
1681 | // Check the degenerate cases |
1682 | if (lhsWords == 0) |
1683 | // 0 % Y ===> 0 |
1684 | return APInt(BitWidth, 0); |
1685 | if (rhsBits == 1) |
1686 | // X % 1 ===> 0 |
1687 | return APInt(BitWidth, 0); |
1688 | if (lhsWords < rhsWords || this->ult(RHS)) |
1689 | // X % Y ===> X, iff X < Y |
1690 | return *this; |
1691 | if (*this == RHS) |
1692 | // X % X == 0; |
1693 | return APInt(BitWidth, 0); |
1694 | if (lhsWords == 1) |
1695 | // All high words are zero, just use native remainder |
1696 | return APInt(BitWidth, U.pVal[0] % RHS.U.pVal[0]); |
1697 | |
1698 | // We have to compute it the hard way. Invoke the Knuth divide algorithm. |
1699 | APInt Remainder(BitWidth, 0); |
1700 | divide(LHS: U.pVal, lhsWords, RHS: RHS.U.pVal, rhsWords, Quotient: nullptr, Remainder: Remainder.U.pVal); |
1701 | return Remainder; |
1702 | } |
1703 | |
1704 | uint64_t APInt::urem(uint64_t RHS) const { |
1705 | assert(RHS != 0 && "Remainder by zero?" ); |
1706 | |
1707 | if (isSingleWord()) |
1708 | return U.VAL % RHS; |
1709 | |
1710 | // Get some facts about the LHS |
1711 | unsigned lhsWords = getNumWords(BitWidth: getActiveBits()); |
1712 | |
1713 | // Check the degenerate cases |
1714 | if (lhsWords == 0) |
1715 | // 0 % Y ===> 0 |
1716 | return 0; |
1717 | if (RHS == 1) |
1718 | // X % 1 ===> 0 |
1719 | return 0; |
1720 | if (this->ult(RHS)) |
1721 | // X % Y ===> X, iff X < Y |
1722 | return getZExtValue(); |
1723 | if (*this == RHS) |
1724 | // X % X == 0; |
1725 | return 0; |
1726 | if (lhsWords == 1) |
1727 | // All high words are zero, just use native remainder |
1728 | return U.pVal[0] % RHS; |
1729 | |
1730 | // We have to compute it the hard way. Invoke the Knuth divide algorithm. |
1731 | uint64_t Remainder; |
1732 | divide(LHS: U.pVal, lhsWords, RHS: &RHS, rhsWords: 1, Quotient: nullptr, Remainder: &Remainder); |
1733 | return Remainder; |
1734 | } |
1735 | |
1736 | APInt APInt::srem(const APInt &RHS) const { |
1737 | if (isNegative()) { |
1738 | if (RHS.isNegative()) |
1739 | return -((-(*this)).urem(RHS: -RHS)); |
1740 | return -((-(*this)).urem(RHS)); |
1741 | } |
1742 | if (RHS.isNegative()) |
1743 | return this->urem(RHS: -RHS); |
1744 | return this->urem(RHS); |
1745 | } |
1746 | |
1747 | int64_t APInt::srem(int64_t RHS) const { |
1748 | if (isNegative()) { |
1749 | if (RHS < 0) |
1750 | return -((-(*this)).urem(RHS: -RHS)); |
1751 | return -((-(*this)).urem(RHS)); |
1752 | } |
1753 | if (RHS < 0) |
1754 | return this->urem(RHS: -RHS); |
1755 | return this->urem(RHS); |
1756 | } |
1757 | |
1758 | void APInt::udivrem(const APInt &LHS, const APInt &RHS, |
1759 | APInt &Quotient, APInt &Remainder) { |
1760 | assert(LHS.BitWidth == RHS.BitWidth && "Bit widths must be the same" ); |
1761 | unsigned BitWidth = LHS.BitWidth; |
1762 | |
1763 | // First, deal with the easy case |
1764 | if (LHS.isSingleWord()) { |
1765 | assert(RHS.U.VAL != 0 && "Divide by zero?" ); |
1766 | uint64_t QuotVal = LHS.U.VAL / RHS.U.VAL; |
1767 | uint64_t RemVal = LHS.U.VAL % RHS.U.VAL; |
1768 | Quotient = APInt(BitWidth, QuotVal); |
1769 | Remainder = APInt(BitWidth, RemVal); |
1770 | return; |
1771 | } |
1772 | |
1773 | // Get some size facts about the dividend and divisor |
1774 | unsigned lhsWords = getNumWords(BitWidth: LHS.getActiveBits()); |
1775 | unsigned rhsBits = RHS.getActiveBits(); |
1776 | unsigned rhsWords = getNumWords(BitWidth: rhsBits); |
1777 | assert(rhsWords && "Performing divrem operation by zero ???" ); |
1778 | |
1779 | // Check the degenerate cases |
1780 | if (lhsWords == 0) { |
1781 | Quotient = APInt(BitWidth, 0); // 0 / Y ===> 0 |
1782 | Remainder = APInt(BitWidth, 0); // 0 % Y ===> 0 |
1783 | return; |
1784 | } |
1785 | |
1786 | if (rhsBits == 1) { |
1787 | Quotient = LHS; // X / 1 ===> X |
1788 | Remainder = APInt(BitWidth, 0); // X % 1 ===> 0 |
1789 | } |
1790 | |
1791 | if (lhsWords < rhsWords || LHS.ult(RHS)) { |
1792 | Remainder = LHS; // X % Y ===> X, iff X < Y |
1793 | Quotient = APInt(BitWidth, 0); // X / Y ===> 0, iff X < Y |
1794 | return; |
1795 | } |
1796 | |
1797 | if (LHS == RHS) { |
1798 | Quotient = APInt(BitWidth, 1); // X / X ===> 1 |
1799 | Remainder = APInt(BitWidth, 0); // X % X ===> 0; |
1800 | return; |
1801 | } |
1802 | |
1803 | // Make sure there is enough space to hold the results. |
1804 | // NOTE: This assumes that reallocate won't affect any bits if it doesn't |
1805 | // change the size. This is necessary if Quotient or Remainder is aliased |
1806 | // with LHS or RHS. |
1807 | Quotient.reallocate(NewBitWidth: BitWidth); |
1808 | Remainder.reallocate(NewBitWidth: BitWidth); |
1809 | |
1810 | if (lhsWords == 1) { // rhsWords is 1 if lhsWords is 1. |
1811 | // There is only one word to consider so use the native versions. |
1812 | uint64_t lhsValue = LHS.U.pVal[0]; |
1813 | uint64_t rhsValue = RHS.U.pVal[0]; |
1814 | Quotient = lhsValue / rhsValue; |
1815 | Remainder = lhsValue % rhsValue; |
1816 | return; |
1817 | } |
1818 | |
1819 | // Okay, lets do it the long way |
1820 | divide(LHS: LHS.U.pVal, lhsWords, RHS: RHS.U.pVal, rhsWords, Quotient: Quotient.U.pVal, |
1821 | Remainder: Remainder.U.pVal); |
1822 | // Clear the rest of the Quotient and Remainder. |
1823 | std::memset(s: Quotient.U.pVal + lhsWords, c: 0, |
1824 | n: (getNumWords(BitWidth) - lhsWords) * APINT_WORD_SIZE); |
1825 | std::memset(s: Remainder.U.pVal + rhsWords, c: 0, |
1826 | n: (getNumWords(BitWidth) - rhsWords) * APINT_WORD_SIZE); |
1827 | } |
1828 | |
1829 | void APInt::udivrem(const APInt &LHS, uint64_t RHS, APInt &Quotient, |
1830 | uint64_t &Remainder) { |
1831 | assert(RHS != 0 && "Divide by zero?" ); |
1832 | unsigned BitWidth = LHS.BitWidth; |
1833 | |
1834 | // First, deal with the easy case |
1835 | if (LHS.isSingleWord()) { |
1836 | uint64_t QuotVal = LHS.U.VAL / RHS; |
1837 | Remainder = LHS.U.VAL % RHS; |
1838 | Quotient = APInt(BitWidth, QuotVal); |
1839 | return; |
1840 | } |
1841 | |
1842 | // Get some size facts about the dividend and divisor |
1843 | unsigned lhsWords = getNumWords(BitWidth: LHS.getActiveBits()); |
1844 | |
1845 | // Check the degenerate cases |
1846 | if (lhsWords == 0) { |
1847 | Quotient = APInt(BitWidth, 0); // 0 / Y ===> 0 |
1848 | Remainder = 0; // 0 % Y ===> 0 |
1849 | return; |
1850 | } |
1851 | |
1852 | if (RHS == 1) { |
1853 | Quotient = LHS; // X / 1 ===> X |
1854 | Remainder = 0; // X % 1 ===> 0 |
1855 | return; |
1856 | } |
1857 | |
1858 | if (LHS.ult(RHS)) { |
1859 | Remainder = LHS.getZExtValue(); // X % Y ===> X, iff X < Y |
1860 | Quotient = APInt(BitWidth, 0); // X / Y ===> 0, iff X < Y |
1861 | return; |
1862 | } |
1863 | |
1864 | if (LHS == RHS) { |
1865 | Quotient = APInt(BitWidth, 1); // X / X ===> 1 |
1866 | Remainder = 0; // X % X ===> 0; |
1867 | return; |
1868 | } |
1869 | |
1870 | // Make sure there is enough space to hold the results. |
1871 | // NOTE: This assumes that reallocate won't affect any bits if it doesn't |
1872 | // change the size. This is necessary if Quotient is aliased with LHS. |
1873 | Quotient.reallocate(NewBitWidth: BitWidth); |
1874 | |
1875 | if (lhsWords == 1) { // rhsWords is 1 if lhsWords is 1. |
1876 | // There is only one word to consider so use the native versions. |
1877 | uint64_t lhsValue = LHS.U.pVal[0]; |
1878 | Quotient = lhsValue / RHS; |
1879 | Remainder = lhsValue % RHS; |
1880 | return; |
1881 | } |
1882 | |
1883 | // Okay, lets do it the long way |
1884 | divide(LHS: LHS.U.pVal, lhsWords, RHS: &RHS, rhsWords: 1, Quotient: Quotient.U.pVal, Remainder: &Remainder); |
1885 | // Clear the rest of the Quotient. |
1886 | std::memset(s: Quotient.U.pVal + lhsWords, c: 0, |
1887 | n: (getNumWords(BitWidth) - lhsWords) * APINT_WORD_SIZE); |
1888 | } |
1889 | |
1890 | void APInt::sdivrem(const APInt &LHS, const APInt &RHS, |
1891 | APInt &Quotient, APInt &Remainder) { |
1892 | if (LHS.isNegative()) { |
1893 | if (RHS.isNegative()) |
1894 | APInt::udivrem(LHS: -LHS, RHS: -RHS, Quotient, Remainder); |
1895 | else { |
1896 | APInt::udivrem(LHS: -LHS, RHS, Quotient, Remainder); |
1897 | Quotient.negate(); |
1898 | } |
1899 | Remainder.negate(); |
1900 | } else if (RHS.isNegative()) { |
1901 | APInt::udivrem(LHS, RHS: -RHS, Quotient, Remainder); |
1902 | Quotient.negate(); |
1903 | } else { |
1904 | APInt::udivrem(LHS, RHS, Quotient, Remainder); |
1905 | } |
1906 | } |
1907 | |
1908 | void APInt::sdivrem(const APInt &LHS, int64_t RHS, |
1909 | APInt &Quotient, int64_t &Remainder) { |
1910 | uint64_t R = Remainder; |
1911 | if (LHS.isNegative()) { |
1912 | if (RHS < 0) |
1913 | APInt::udivrem(LHS: -LHS, RHS: -RHS, Quotient, Remainder&: R); |
1914 | else { |
1915 | APInt::udivrem(LHS: -LHS, RHS, Quotient, Remainder&: R); |
1916 | Quotient.negate(); |
1917 | } |
1918 | R = -R; |
1919 | } else if (RHS < 0) { |
1920 | APInt::udivrem(LHS, RHS: -RHS, Quotient, Remainder&: R); |
1921 | Quotient.negate(); |
1922 | } else { |
1923 | APInt::udivrem(LHS, RHS, Quotient, Remainder&: R); |
1924 | } |
1925 | Remainder = R; |
1926 | } |
1927 | |
1928 | APInt APInt::sadd_ov(const APInt &RHS, bool &Overflow) const { |
1929 | APInt Res = *this+RHS; |
1930 | Overflow = isNonNegative() == RHS.isNonNegative() && |
1931 | Res.isNonNegative() != isNonNegative(); |
1932 | return Res; |
1933 | } |
1934 | |
1935 | APInt APInt::uadd_ov(const APInt &RHS, bool &Overflow) const { |
1936 | APInt Res = *this+RHS; |
1937 | Overflow = Res.ult(RHS); |
1938 | return Res; |
1939 | } |
1940 | |
1941 | APInt APInt::ssub_ov(const APInt &RHS, bool &Overflow) const { |
1942 | APInt Res = *this - RHS; |
1943 | Overflow = isNonNegative() != RHS.isNonNegative() && |
1944 | Res.isNonNegative() != isNonNegative(); |
1945 | return Res; |
1946 | } |
1947 | |
1948 | APInt APInt::usub_ov(const APInt &RHS, bool &Overflow) const { |
1949 | APInt Res = *this-RHS; |
1950 | Overflow = Res.ugt(RHS: *this); |
1951 | return Res; |
1952 | } |
1953 | |
1954 | APInt APInt::sdiv_ov(const APInt &RHS, bool &Overflow) const { |
1955 | // MININT/-1 --> overflow. |
1956 | Overflow = isMinSignedValue() && RHS.isAllOnes(); |
1957 | return sdiv(RHS); |
1958 | } |
1959 | |
1960 | APInt APInt::smul_ov(const APInt &RHS, bool &Overflow) const { |
1961 | APInt Res = *this * RHS; |
1962 | |
1963 | if (RHS != 0) |
1964 | Overflow = Res.sdiv(RHS) != *this || |
1965 | (isMinSignedValue() && RHS.isAllOnes()); |
1966 | else |
1967 | Overflow = false; |
1968 | return Res; |
1969 | } |
1970 | |
1971 | APInt APInt::umul_ov(const APInt &RHS, bool &Overflow) const { |
1972 | if (countl_zero() + RHS.countl_zero() + 2 <= BitWidth) { |
1973 | Overflow = true; |
1974 | return *this * RHS; |
1975 | } |
1976 | |
1977 | APInt Res = lshr(shiftAmt: 1) * RHS; |
1978 | Overflow = Res.isNegative(); |
1979 | Res <<= 1; |
1980 | if ((*this)[0]) { |
1981 | Res += RHS; |
1982 | if (Res.ult(RHS)) |
1983 | Overflow = true; |
1984 | } |
1985 | return Res; |
1986 | } |
1987 | |
1988 | APInt APInt::sshl_ov(const APInt &ShAmt, bool &Overflow) const { |
1989 | return sshl_ov(Amt: ShAmt.getLimitedValue(Limit: getBitWidth()), Overflow); |
1990 | } |
1991 | |
1992 | APInt APInt::sshl_ov(unsigned ShAmt, bool &Overflow) const { |
1993 | Overflow = ShAmt >= getBitWidth(); |
1994 | if (Overflow) |
1995 | return APInt(BitWidth, 0); |
1996 | |
1997 | if (isNonNegative()) // Don't allow sign change. |
1998 | Overflow = ShAmt >= countl_zero(); |
1999 | else |
2000 | Overflow = ShAmt >= countl_one(); |
2001 | |
2002 | return *this << ShAmt; |
2003 | } |
2004 | |
2005 | APInt APInt::ushl_ov(const APInt &ShAmt, bool &Overflow) const { |
2006 | return ushl_ov(Amt: ShAmt.getLimitedValue(Limit: getBitWidth()), Overflow); |
2007 | } |
2008 | |
2009 | APInt APInt::ushl_ov(unsigned ShAmt, bool &Overflow) const { |
2010 | Overflow = ShAmt >= getBitWidth(); |
2011 | if (Overflow) |
2012 | return APInt(BitWidth, 0); |
2013 | |
2014 | Overflow = ShAmt > countl_zero(); |
2015 | |
2016 | return *this << ShAmt; |
2017 | } |
2018 | |
2019 | APInt APInt::sfloordiv_ov(const APInt &RHS, bool &Overflow) const { |
2020 | APInt quotient = sdiv_ov(RHS, Overflow); |
2021 | if ((quotient * RHS != *this) && (isNegative() != RHS.isNegative())) |
2022 | return quotient - 1; |
2023 | return quotient; |
2024 | } |
2025 | |
2026 | APInt APInt::sadd_sat(const APInt &RHS) const { |
2027 | bool Overflow; |
2028 | APInt Res = sadd_ov(RHS, Overflow); |
2029 | if (!Overflow) |
2030 | return Res; |
2031 | |
2032 | return isNegative() ? APInt::getSignedMinValue(numBits: BitWidth) |
2033 | : APInt::getSignedMaxValue(numBits: BitWidth); |
2034 | } |
2035 | |
2036 | APInt APInt::uadd_sat(const APInt &RHS) const { |
2037 | bool Overflow; |
2038 | APInt Res = uadd_ov(RHS, Overflow); |
2039 | if (!Overflow) |
2040 | return Res; |
2041 | |
2042 | return APInt::getMaxValue(numBits: BitWidth); |
2043 | } |
2044 | |
2045 | APInt APInt::ssub_sat(const APInt &RHS) const { |
2046 | bool Overflow; |
2047 | APInt Res = ssub_ov(RHS, Overflow); |
2048 | if (!Overflow) |
2049 | return Res; |
2050 | |
2051 | return isNegative() ? APInt::getSignedMinValue(numBits: BitWidth) |
2052 | : APInt::getSignedMaxValue(numBits: BitWidth); |
2053 | } |
2054 | |
2055 | APInt APInt::usub_sat(const APInt &RHS) const { |
2056 | bool Overflow; |
2057 | APInt Res = usub_ov(RHS, Overflow); |
2058 | if (!Overflow) |
2059 | return Res; |
2060 | |
2061 | return APInt(BitWidth, 0); |
2062 | } |
2063 | |
2064 | APInt APInt::smul_sat(const APInt &RHS) const { |
2065 | bool Overflow; |
2066 | APInt Res = smul_ov(RHS, Overflow); |
2067 | if (!Overflow) |
2068 | return Res; |
2069 | |
2070 | // The result is negative if one and only one of inputs is negative. |
2071 | bool ResIsNegative = isNegative() ^ RHS.isNegative(); |
2072 | |
2073 | return ResIsNegative ? APInt::getSignedMinValue(numBits: BitWidth) |
2074 | : APInt::getSignedMaxValue(numBits: BitWidth); |
2075 | } |
2076 | |
2077 | APInt APInt::umul_sat(const APInt &RHS) const { |
2078 | bool Overflow; |
2079 | APInt Res = umul_ov(RHS, Overflow); |
2080 | if (!Overflow) |
2081 | return Res; |
2082 | |
2083 | return APInt::getMaxValue(numBits: BitWidth); |
2084 | } |
2085 | |
2086 | APInt APInt::sshl_sat(const APInt &RHS) const { |
2087 | return sshl_sat(RHS: RHS.getLimitedValue(Limit: getBitWidth())); |
2088 | } |
2089 | |
2090 | APInt APInt::sshl_sat(unsigned RHS) const { |
2091 | bool Overflow; |
2092 | APInt Res = sshl_ov(ShAmt: RHS, Overflow); |
2093 | if (!Overflow) |
2094 | return Res; |
2095 | |
2096 | return isNegative() ? APInt::getSignedMinValue(numBits: BitWidth) |
2097 | : APInt::getSignedMaxValue(numBits: BitWidth); |
2098 | } |
2099 | |
2100 | APInt APInt::ushl_sat(const APInt &RHS) const { |
2101 | return ushl_sat(RHS: RHS.getLimitedValue(Limit: getBitWidth())); |
2102 | } |
2103 | |
2104 | APInt APInt::ushl_sat(unsigned RHS) const { |
2105 | bool Overflow; |
2106 | APInt Res = ushl_ov(ShAmt: RHS, Overflow); |
2107 | if (!Overflow) |
2108 | return Res; |
2109 | |
2110 | return APInt::getMaxValue(numBits: BitWidth); |
2111 | } |
2112 | |
2113 | void APInt::fromString(unsigned numbits, StringRef str, uint8_t radix) { |
2114 | // Check our assumptions here |
2115 | assert(!str.empty() && "Invalid string length" ); |
2116 | assert((radix == 10 || radix == 8 || radix == 16 || radix == 2 || |
2117 | radix == 36) && |
2118 | "Radix should be 2, 8, 10, 16, or 36!" ); |
2119 | |
2120 | StringRef::iterator p = str.begin(); |
2121 | size_t slen = str.size(); |
2122 | bool isNeg = *p == '-'; |
2123 | if (*p == '-' || *p == '+') { |
2124 | p++; |
2125 | slen--; |
2126 | assert(slen && "String is only a sign, needs a value." ); |
2127 | } |
2128 | assert((slen <= numbits || radix != 2) && "Insufficient bit width" ); |
2129 | assert(((slen-1)*3 <= numbits || radix != 8) && "Insufficient bit width" ); |
2130 | assert(((slen-1)*4 <= numbits || radix != 16) && "Insufficient bit width" ); |
2131 | assert((((slen-1)*64)/22 <= numbits || radix != 10) && |
2132 | "Insufficient bit width" ); |
2133 | |
2134 | // Allocate memory if needed |
2135 | if (isSingleWord()) |
2136 | U.VAL = 0; |
2137 | else |
2138 | U.pVal = getClearedMemory(numWords: getNumWords()); |
2139 | |
2140 | // Figure out if we can shift instead of multiply |
2141 | unsigned shift = (radix == 16 ? 4 : radix == 8 ? 3 : radix == 2 ? 1 : 0); |
2142 | |
2143 | // Enter digit traversal loop |
2144 | for (StringRef::iterator e = str.end(); p != e; ++p) { |
2145 | unsigned digit = getDigit(cdigit: *p, radix); |
2146 | assert(digit < radix && "Invalid character in digit string" ); |
2147 | |
2148 | // Shift or multiply the value by the radix |
2149 | if (slen > 1) { |
2150 | if (shift) |
2151 | *this <<= shift; |
2152 | else |
2153 | *this *= radix; |
2154 | } |
2155 | |
2156 | // Add in the digit we just interpreted |
2157 | *this += digit; |
2158 | } |
2159 | // If its negative, put it in two's complement form |
2160 | if (isNeg) |
2161 | this->negate(); |
2162 | } |
2163 | |
2164 | void APInt::toString(SmallVectorImpl<char> &Str, unsigned Radix, bool Signed, |
2165 | bool formatAsCLiteral, bool UpperCase, |
2166 | bool InsertSeparators) const { |
2167 | assert((Radix == 10 || Radix == 8 || Radix == 16 || Radix == 2 || |
2168 | Radix == 36) && |
2169 | "Radix should be 2, 8, 10, 16, or 36!" ); |
2170 | |
2171 | const char *Prefix = "" ; |
2172 | if (formatAsCLiteral) { |
2173 | switch (Radix) { |
2174 | case 2: |
2175 | // Binary literals are a non-standard extension added in gcc 4.3: |
2176 | // http://gcc.gnu.org/onlinedocs/gcc-4.3.0/gcc/Binary-constants.html |
2177 | Prefix = "0b" ; |
2178 | break; |
2179 | case 8: |
2180 | Prefix = "0" ; |
2181 | break; |
2182 | case 10: |
2183 | break; // No prefix |
2184 | case 16: |
2185 | Prefix = "0x" ; |
2186 | break; |
2187 | default: |
2188 | llvm_unreachable("Invalid radix!" ); |
2189 | } |
2190 | } |
2191 | |
2192 | // Number of digits in a group between separators. |
2193 | unsigned Grouping = (Radix == 8 || Radix == 10) ? 3 : 4; |
2194 | |
2195 | // First, check for a zero value and just short circuit the logic below. |
2196 | if (isZero()) { |
2197 | while (*Prefix) { |
2198 | Str.push_back(Elt: *Prefix); |
2199 | ++Prefix; |
2200 | }; |
2201 | Str.push_back(Elt: '0'); |
2202 | return; |
2203 | } |
2204 | |
2205 | static const char BothDigits[] = "0123456789abcdefghijklmnopqrstuvwxyz" |
2206 | "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" ; |
2207 | const char *Digits = BothDigits + (UpperCase ? 36 : 0); |
2208 | |
2209 | if (isSingleWord()) { |
2210 | char Buffer[65]; |
2211 | char *BufPtr = std::end(arr&: Buffer); |
2212 | |
2213 | uint64_t N; |
2214 | if (!Signed) { |
2215 | N = getZExtValue(); |
2216 | } else { |
2217 | int64_t I = getSExtValue(); |
2218 | if (I >= 0) { |
2219 | N = I; |
2220 | } else { |
2221 | Str.push_back(Elt: '-'); |
2222 | N = -(uint64_t)I; |
2223 | } |
2224 | } |
2225 | |
2226 | while (*Prefix) { |
2227 | Str.push_back(Elt: *Prefix); |
2228 | ++Prefix; |
2229 | }; |
2230 | |
2231 | int Pos = 0; |
2232 | while (N) { |
2233 | if (InsertSeparators && Pos % Grouping == 0 && Pos > 0) |
2234 | *--BufPtr = '\''; |
2235 | *--BufPtr = Digits[N % Radix]; |
2236 | N /= Radix; |
2237 | Pos++; |
2238 | } |
2239 | Str.append(in_start: BufPtr, in_end: std::end(arr&: Buffer)); |
2240 | return; |
2241 | } |
2242 | |
2243 | APInt Tmp(*this); |
2244 | |
2245 | if (Signed && isNegative()) { |
2246 | // They want to print the signed version and it is a negative value |
2247 | // Flip the bits and add one to turn it into the equivalent positive |
2248 | // value and put a '-' in the result. |
2249 | Tmp.negate(); |
2250 | Str.push_back(Elt: '-'); |
2251 | } |
2252 | |
2253 | while (*Prefix) { |
2254 | Str.push_back(Elt: *Prefix); |
2255 | ++Prefix; |
2256 | } |
2257 | |
2258 | // We insert the digits backward, then reverse them to get the right order. |
2259 | unsigned StartDig = Str.size(); |
2260 | |
2261 | // For the 2, 8 and 16 bit cases, we can just shift instead of divide |
2262 | // because the number of bits per digit (1, 3 and 4 respectively) divides |
2263 | // equally. We just shift until the value is zero. |
2264 | if (Radix == 2 || Radix == 8 || Radix == 16) { |
2265 | // Just shift tmp right for each digit width until it becomes zero |
2266 | unsigned ShiftAmt = (Radix == 16 ? 4 : (Radix == 8 ? 3 : 1)); |
2267 | unsigned MaskAmt = Radix - 1; |
2268 | |
2269 | int Pos = 0; |
2270 | while (Tmp.getBoolValue()) { |
2271 | unsigned Digit = unsigned(Tmp.getRawData()[0]) & MaskAmt; |
2272 | if (InsertSeparators && Pos % Grouping == 0 && Pos > 0) |
2273 | Str.push_back(Elt: '\''); |
2274 | |
2275 | Str.push_back(Elt: Digits[Digit]); |
2276 | Tmp.lshrInPlace(ShiftAmt); |
2277 | Pos++; |
2278 | } |
2279 | } else { |
2280 | int Pos = 0; |
2281 | while (Tmp.getBoolValue()) { |
2282 | uint64_t Digit; |
2283 | udivrem(LHS: Tmp, RHS: Radix, Quotient&: Tmp, Remainder&: Digit); |
2284 | assert(Digit < Radix && "divide failed" ); |
2285 | if (InsertSeparators && Pos % Grouping == 0 && Pos > 0) |
2286 | Str.push_back(Elt: '\''); |
2287 | |
2288 | Str.push_back(Elt: Digits[Digit]); |
2289 | Pos++; |
2290 | } |
2291 | } |
2292 | |
2293 | // Reverse the digits before returning. |
2294 | std::reverse(first: Str.begin()+StartDig, last: Str.end()); |
2295 | } |
2296 | |
2297 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
2298 | LLVM_DUMP_METHOD void APInt::dump() const { |
2299 | SmallString<40> S, U; |
2300 | this->toStringUnsigned(U); |
2301 | this->toStringSigned(S); |
2302 | dbgs() << "APInt(" << BitWidth << "b, " |
2303 | << U << "u " << S << "s)\n" ; |
2304 | } |
2305 | #endif |
2306 | |
2307 | void APInt::print(raw_ostream &OS, bool isSigned) const { |
2308 | SmallString<40> S; |
2309 | this->toString(Str&: S, Radix: 10, Signed: isSigned, /* formatAsCLiteral = */false); |
2310 | OS << S; |
2311 | } |
2312 | |
2313 | // This implements a variety of operations on a representation of |
2314 | // arbitrary precision, two's-complement, bignum integer values. |
2315 | |
2316 | // Assumed by lowHalf, highHalf, partMSB and partLSB. A fairly safe |
2317 | // and unrestricting assumption. |
2318 | static_assert(APInt::APINT_BITS_PER_WORD % 2 == 0, |
2319 | "Part width must be divisible by 2!" ); |
2320 | |
2321 | // Returns the integer part with the least significant BITS set. |
2322 | // BITS cannot be zero. |
2323 | static inline APInt::WordType lowBitMask(unsigned bits) { |
2324 | assert(bits != 0 && bits <= APInt::APINT_BITS_PER_WORD); |
2325 | return ~(APInt::WordType) 0 >> (APInt::APINT_BITS_PER_WORD - bits); |
2326 | } |
2327 | |
2328 | /// Returns the value of the lower half of PART. |
2329 | static inline APInt::WordType lowHalf(APInt::WordType part) { |
2330 | return part & lowBitMask(bits: APInt::APINT_BITS_PER_WORD / 2); |
2331 | } |
2332 | |
2333 | /// Returns the value of the upper half of PART. |
2334 | static inline APInt::WordType highHalf(APInt::WordType part) { |
2335 | return part >> (APInt::APINT_BITS_PER_WORD / 2); |
2336 | } |
2337 | |
2338 | /// Sets the least significant part of a bignum to the input value, and zeroes |
2339 | /// out higher parts. |
2340 | void APInt::tcSet(WordType *dst, WordType part, unsigned parts) { |
2341 | assert(parts > 0); |
2342 | dst[0] = part; |
2343 | for (unsigned i = 1; i < parts; i++) |
2344 | dst[i] = 0; |
2345 | } |
2346 | |
2347 | /// Assign one bignum to another. |
2348 | void APInt::tcAssign(WordType *dst, const WordType *src, unsigned parts) { |
2349 | for (unsigned i = 0; i < parts; i++) |
2350 | dst[i] = src[i]; |
2351 | } |
2352 | |
2353 | /// Returns true if a bignum is zero, false otherwise. |
2354 | bool APInt::tcIsZero(const WordType *src, unsigned parts) { |
2355 | for (unsigned i = 0; i < parts; i++) |
2356 | if (src[i]) |
2357 | return false; |
2358 | |
2359 | return true; |
2360 | } |
2361 | |
2362 | /// Extract the given bit of a bignum; returns 0 or 1. |
2363 | int APInt::(const WordType *parts, unsigned bit) { |
2364 | return (parts[whichWord(bitPosition: bit)] & maskBit(bitPosition: bit)) != 0; |
2365 | } |
2366 | |
2367 | /// Set the given bit of a bignum. |
2368 | void APInt::tcSetBit(WordType *parts, unsigned bit) { |
2369 | parts[whichWord(bitPosition: bit)] |= maskBit(bitPosition: bit); |
2370 | } |
2371 | |
2372 | /// Clears the given bit of a bignum. |
2373 | void APInt::tcClearBit(WordType *parts, unsigned bit) { |
2374 | parts[whichWord(bitPosition: bit)] &= ~maskBit(bitPosition: bit); |
2375 | } |
2376 | |
2377 | /// Returns the bit number of the least significant set bit of a number. If the |
2378 | /// input number has no bits set UINT_MAX is returned. |
2379 | unsigned APInt::tcLSB(const WordType *parts, unsigned n) { |
2380 | for (unsigned i = 0; i < n; i++) { |
2381 | if (parts[i] != 0) { |
2382 | unsigned lsb = llvm::countr_zero(Val: parts[i]); |
2383 | return lsb + i * APINT_BITS_PER_WORD; |
2384 | } |
2385 | } |
2386 | |
2387 | return UINT_MAX; |
2388 | } |
2389 | |
2390 | /// Returns the bit number of the most significant set bit of a number. |
2391 | /// If the input number has no bits set UINT_MAX is returned. |
2392 | unsigned APInt::tcMSB(const WordType *parts, unsigned n) { |
2393 | do { |
2394 | --n; |
2395 | |
2396 | if (parts[n] != 0) { |
2397 | static_assert(sizeof(parts[n]) <= sizeof(uint64_t)); |
2398 | unsigned msb = llvm::Log2_64(Value: parts[n]); |
2399 | |
2400 | return msb + n * APINT_BITS_PER_WORD; |
2401 | } |
2402 | } while (n); |
2403 | |
2404 | return UINT_MAX; |
2405 | } |
2406 | |
2407 | /// Copy the bit vector of width srcBITS from SRC, starting at bit srcLSB, to |
2408 | /// DST, of dstCOUNT parts, such that the bit srcLSB becomes the least |
2409 | /// significant bit of DST. All high bits above srcBITS in DST are zero-filled. |
2410 | /// */ |
2411 | void |
2412 | APInt::(WordType *dst, unsigned dstCount, const WordType *src, |
2413 | unsigned srcBits, unsigned srcLSB) { |
2414 | unsigned dstParts = (srcBits + APINT_BITS_PER_WORD - 1) / APINT_BITS_PER_WORD; |
2415 | assert(dstParts <= dstCount); |
2416 | |
2417 | unsigned firstSrcPart = srcLSB / APINT_BITS_PER_WORD; |
2418 | tcAssign(dst, src: src + firstSrcPart, parts: dstParts); |
2419 | |
2420 | unsigned shift = srcLSB % APINT_BITS_PER_WORD; |
2421 | tcShiftRight(dst, Words: dstParts, Count: shift); |
2422 | |
2423 | // We now have (dstParts * APINT_BITS_PER_WORD - shift) bits from SRC |
2424 | // in DST. If this is less that srcBits, append the rest, else |
2425 | // clear the high bits. |
2426 | unsigned n = dstParts * APINT_BITS_PER_WORD - shift; |
2427 | if (n < srcBits) { |
2428 | WordType mask = lowBitMask (bits: srcBits - n); |
2429 | dst[dstParts - 1] |= ((src[firstSrcPart + dstParts] & mask) |
2430 | << n % APINT_BITS_PER_WORD); |
2431 | } else if (n > srcBits) { |
2432 | if (srcBits % APINT_BITS_PER_WORD) |
2433 | dst[dstParts - 1] &= lowBitMask (bits: srcBits % APINT_BITS_PER_WORD); |
2434 | } |
2435 | |
2436 | // Clear high parts. |
2437 | while (dstParts < dstCount) |
2438 | dst[dstParts++] = 0; |
2439 | } |
2440 | |
2441 | //// DST += RHS + C where C is zero or one. Returns the carry flag. |
2442 | APInt::WordType APInt::tcAdd(WordType *dst, const WordType *rhs, |
2443 | WordType c, unsigned parts) { |
2444 | assert(c <= 1); |
2445 | |
2446 | for (unsigned i = 0; i < parts; i++) { |
2447 | WordType l = dst[i]; |
2448 | if (c) { |
2449 | dst[i] += rhs[i] + 1; |
2450 | c = (dst[i] <= l); |
2451 | } else { |
2452 | dst[i] += rhs[i]; |
2453 | c = (dst[i] < l); |
2454 | } |
2455 | } |
2456 | |
2457 | return c; |
2458 | } |
2459 | |
2460 | /// This function adds a single "word" integer, src, to the multiple |
2461 | /// "word" integer array, dst[]. dst[] is modified to reflect the addition and |
2462 | /// 1 is returned if there is a carry out, otherwise 0 is returned. |
2463 | /// @returns the carry of the addition. |
2464 | APInt::WordType APInt::tcAddPart(WordType *dst, WordType src, |
2465 | unsigned parts) { |
2466 | for (unsigned i = 0; i < parts; ++i) { |
2467 | dst[i] += src; |
2468 | if (dst[i] >= src) |
2469 | return 0; // No need to carry so exit early. |
2470 | src = 1; // Carry one to next digit. |
2471 | } |
2472 | |
2473 | return 1; |
2474 | } |
2475 | |
2476 | /// DST -= RHS + C where C is zero or one. Returns the carry flag. |
2477 | APInt::WordType APInt::tcSubtract(WordType *dst, const WordType *rhs, |
2478 | WordType c, unsigned parts) { |
2479 | assert(c <= 1); |
2480 | |
2481 | for (unsigned i = 0; i < parts; i++) { |
2482 | WordType l = dst[i]; |
2483 | if (c) { |
2484 | dst[i] -= rhs[i] + 1; |
2485 | c = (dst[i] >= l); |
2486 | } else { |
2487 | dst[i] -= rhs[i]; |
2488 | c = (dst[i] > l); |
2489 | } |
2490 | } |
2491 | |
2492 | return c; |
2493 | } |
2494 | |
2495 | /// This function subtracts a single "word" (64-bit word), src, from |
2496 | /// the multi-word integer array, dst[], propagating the borrowed 1 value until |
2497 | /// no further borrowing is needed or it runs out of "words" in dst. The result |
2498 | /// is 1 if "borrowing" exhausted the digits in dst, or 0 if dst was not |
2499 | /// exhausted. In other words, if src > dst then this function returns 1, |
2500 | /// otherwise 0. |
2501 | /// @returns the borrow out of the subtraction |
2502 | APInt::WordType APInt::tcSubtractPart(WordType *dst, WordType src, |
2503 | unsigned parts) { |
2504 | for (unsigned i = 0; i < parts; ++i) { |
2505 | WordType Dst = dst[i]; |
2506 | dst[i] -= src; |
2507 | if (src <= Dst) |
2508 | return 0; // No need to borrow so exit early. |
2509 | src = 1; // We have to "borrow 1" from next "word" |
2510 | } |
2511 | |
2512 | return 1; |
2513 | } |
2514 | |
2515 | /// Negate a bignum in-place. |
2516 | void APInt::tcNegate(WordType *dst, unsigned parts) { |
2517 | tcComplement(dst, parts); |
2518 | tcIncrement(dst, parts); |
2519 | } |
2520 | |
2521 | /// DST += SRC * MULTIPLIER + CARRY if add is true |
2522 | /// DST = SRC * MULTIPLIER + CARRY if add is false |
2523 | /// Requires 0 <= DSTPARTS <= SRCPARTS + 1. If DST overlaps SRC |
2524 | /// they must start at the same point, i.e. DST == SRC. |
2525 | /// If DSTPARTS == SRCPARTS + 1 no overflow occurs and zero is |
2526 | /// returned. Otherwise DST is filled with the least significant |
2527 | /// DSTPARTS parts of the result, and if all of the omitted higher |
2528 | /// parts were zero return zero, otherwise overflow occurred and |
2529 | /// return one. |
2530 | int APInt::tcMultiplyPart(WordType *dst, const WordType *src, |
2531 | WordType multiplier, WordType carry, |
2532 | unsigned srcParts, unsigned dstParts, |
2533 | bool add) { |
2534 | // Otherwise our writes of DST kill our later reads of SRC. |
2535 | assert(dst <= src || dst >= src + srcParts); |
2536 | assert(dstParts <= srcParts + 1); |
2537 | |
2538 | // N loops; minimum of dstParts and srcParts. |
2539 | unsigned n = std::min(a: dstParts, b: srcParts); |
2540 | |
2541 | for (unsigned i = 0; i < n; i++) { |
2542 | // [LOW, HIGH] = MULTIPLIER * SRC[i] + DST[i] + CARRY. |
2543 | // This cannot overflow, because: |
2544 | // (n - 1) * (n - 1) + 2 (n - 1) = (n - 1) * (n + 1) |
2545 | // which is less than n^2. |
2546 | WordType srcPart = src[i]; |
2547 | WordType low, mid, high; |
2548 | if (multiplier == 0 || srcPart == 0) { |
2549 | low = carry; |
2550 | high = 0; |
2551 | } else { |
2552 | low = lowHalf(part: srcPart) * lowHalf(part: multiplier); |
2553 | high = highHalf(part: srcPart) * highHalf(part: multiplier); |
2554 | |
2555 | mid = lowHalf(part: srcPart) * highHalf(part: multiplier); |
2556 | high += highHalf(part: mid); |
2557 | mid <<= APINT_BITS_PER_WORD / 2; |
2558 | if (low + mid < low) |
2559 | high++; |
2560 | low += mid; |
2561 | |
2562 | mid = highHalf(part: srcPart) * lowHalf(part: multiplier); |
2563 | high += highHalf(part: mid); |
2564 | mid <<= APINT_BITS_PER_WORD / 2; |
2565 | if (low + mid < low) |
2566 | high++; |
2567 | low += mid; |
2568 | |
2569 | // Now add carry. |
2570 | if (low + carry < low) |
2571 | high++; |
2572 | low += carry; |
2573 | } |
2574 | |
2575 | if (add) { |
2576 | // And now DST[i], and store the new low part there. |
2577 | if (low + dst[i] < low) |
2578 | high++; |
2579 | dst[i] += low; |
2580 | } else { |
2581 | dst[i] = low; |
2582 | } |
2583 | |
2584 | carry = high; |
2585 | } |
2586 | |
2587 | if (srcParts < dstParts) { |
2588 | // Full multiplication, there is no overflow. |
2589 | assert(srcParts + 1 == dstParts); |
2590 | dst[srcParts] = carry; |
2591 | return 0; |
2592 | } |
2593 | |
2594 | // We overflowed if there is carry. |
2595 | if (carry) |
2596 | return 1; |
2597 | |
2598 | // We would overflow if any significant unwritten parts would be |
2599 | // non-zero. This is true if any remaining src parts are non-zero |
2600 | // and the multiplier is non-zero. |
2601 | if (multiplier) |
2602 | for (unsigned i = dstParts; i < srcParts; i++) |
2603 | if (src[i]) |
2604 | return 1; |
2605 | |
2606 | // We fitted in the narrow destination. |
2607 | return 0; |
2608 | } |
2609 | |
2610 | /// DST = LHS * RHS, where DST has the same width as the operands and |
2611 | /// is filled with the least significant parts of the result. Returns |
2612 | /// one if overflow occurred, otherwise zero. DST must be disjoint |
2613 | /// from both operands. |
2614 | int APInt::tcMultiply(WordType *dst, const WordType *lhs, |
2615 | const WordType *rhs, unsigned parts) { |
2616 | assert(dst != lhs && dst != rhs); |
2617 | |
2618 | int overflow = 0; |
2619 | |
2620 | for (unsigned i = 0; i < parts; i++) { |
2621 | // Don't accumulate on the first iteration so we don't need to initalize |
2622 | // dst to 0. |
2623 | overflow |= |
2624 | tcMultiplyPart(dst: &dst[i], src: lhs, multiplier: rhs[i], carry: 0, srcParts: parts, dstParts: parts - i, add: i != 0); |
2625 | } |
2626 | |
2627 | return overflow; |
2628 | } |
2629 | |
2630 | /// DST = LHS * RHS, where DST has width the sum of the widths of the |
2631 | /// operands. No overflow occurs. DST must be disjoint from both operands. |
2632 | void APInt::tcFullMultiply(WordType *dst, const WordType *lhs, |
2633 | const WordType *rhs, unsigned lhsParts, |
2634 | unsigned rhsParts) { |
2635 | // Put the narrower number on the LHS for less loops below. |
2636 | if (lhsParts > rhsParts) |
2637 | return tcFullMultiply (dst, lhs: rhs, rhs: lhs, lhsParts: rhsParts, rhsParts: lhsParts); |
2638 | |
2639 | assert(dst != lhs && dst != rhs); |
2640 | |
2641 | for (unsigned i = 0; i < lhsParts; i++) { |
2642 | // Don't accumulate on the first iteration so we don't need to initalize |
2643 | // dst to 0. |
2644 | tcMultiplyPart(dst: &dst[i], src: rhs, multiplier: lhs[i], carry: 0, srcParts: rhsParts, dstParts: rhsParts + 1, add: i != 0); |
2645 | } |
2646 | } |
2647 | |
2648 | // If RHS is zero LHS and REMAINDER are left unchanged, return one. |
2649 | // Otherwise set LHS to LHS / RHS with the fractional part discarded, |
2650 | // set REMAINDER to the remainder, return zero. i.e. |
2651 | // |
2652 | // OLD_LHS = RHS * LHS + REMAINDER |
2653 | // |
2654 | // SCRATCH is a bignum of the same size as the operands and result for |
2655 | // use by the routine; its contents need not be initialized and are |
2656 | // destroyed. LHS, REMAINDER and SCRATCH must be distinct. |
2657 | int APInt::tcDivide(WordType *lhs, const WordType *rhs, |
2658 | WordType *remainder, WordType *srhs, |
2659 | unsigned parts) { |
2660 | assert(lhs != remainder && lhs != srhs && remainder != srhs); |
2661 | |
2662 | unsigned shiftCount = tcMSB(parts: rhs, n: parts) + 1; |
2663 | if (shiftCount == 0) |
2664 | return true; |
2665 | |
2666 | shiftCount = parts * APINT_BITS_PER_WORD - shiftCount; |
2667 | unsigned n = shiftCount / APINT_BITS_PER_WORD; |
2668 | WordType mask = (WordType) 1 << (shiftCount % APINT_BITS_PER_WORD); |
2669 | |
2670 | tcAssign(dst: srhs, src: rhs, parts); |
2671 | tcShiftLeft(srhs, Words: parts, Count: shiftCount); |
2672 | tcAssign(dst: remainder, src: lhs, parts); |
2673 | tcSet(dst: lhs, part: 0, parts); |
2674 | |
2675 | // Loop, subtracting SRHS if REMAINDER is greater and adding that to the |
2676 | // total. |
2677 | for (;;) { |
2678 | int compare = tcCompare(remainder, srhs, parts); |
2679 | if (compare >= 0) { |
2680 | tcSubtract(dst: remainder, rhs: srhs, c: 0, parts); |
2681 | lhs[n] |= mask; |
2682 | } |
2683 | |
2684 | if (shiftCount == 0) |
2685 | break; |
2686 | shiftCount--; |
2687 | tcShiftRight(srhs, Words: parts, Count: 1); |
2688 | if ((mask >>= 1) == 0) { |
2689 | mask = (WordType) 1 << (APINT_BITS_PER_WORD - 1); |
2690 | n--; |
2691 | } |
2692 | } |
2693 | |
2694 | return false; |
2695 | } |
2696 | |
2697 | /// Shift a bignum left Count bits in-place. Shifted in bits are zero. There are |
2698 | /// no restrictions on Count. |
2699 | void APInt::tcShiftLeft(WordType *Dst, unsigned Words, unsigned Count) { |
2700 | // Don't bother performing a no-op shift. |
2701 | if (!Count) |
2702 | return; |
2703 | |
2704 | // WordShift is the inter-part shift; BitShift is the intra-part shift. |
2705 | unsigned WordShift = std::min(a: Count / APINT_BITS_PER_WORD, b: Words); |
2706 | unsigned BitShift = Count % APINT_BITS_PER_WORD; |
2707 | |
2708 | // Fastpath for moving by whole words. |
2709 | if (BitShift == 0) { |
2710 | std::memmove(dest: Dst + WordShift, src: Dst, n: (Words - WordShift) * APINT_WORD_SIZE); |
2711 | } else { |
2712 | while (Words-- > WordShift) { |
2713 | Dst[Words] = Dst[Words - WordShift] << BitShift; |
2714 | if (Words > WordShift) |
2715 | Dst[Words] |= |
2716 | Dst[Words - WordShift - 1] >> (APINT_BITS_PER_WORD - BitShift); |
2717 | } |
2718 | } |
2719 | |
2720 | // Fill in the remainder with 0s. |
2721 | std::memset(s: Dst, c: 0, n: WordShift * APINT_WORD_SIZE); |
2722 | } |
2723 | |
2724 | /// Shift a bignum right Count bits in-place. Shifted in bits are zero. There |
2725 | /// are no restrictions on Count. |
2726 | void APInt::tcShiftRight(WordType *Dst, unsigned Words, unsigned Count) { |
2727 | // Don't bother performing a no-op shift. |
2728 | if (!Count) |
2729 | return; |
2730 | |
2731 | // WordShift is the inter-part shift; BitShift is the intra-part shift. |
2732 | unsigned WordShift = std::min(a: Count / APINT_BITS_PER_WORD, b: Words); |
2733 | unsigned BitShift = Count % APINT_BITS_PER_WORD; |
2734 | |
2735 | unsigned WordsToMove = Words - WordShift; |
2736 | // Fastpath for moving by whole words. |
2737 | if (BitShift == 0) { |
2738 | std::memmove(dest: Dst, src: Dst + WordShift, n: WordsToMove * APINT_WORD_SIZE); |
2739 | } else { |
2740 | for (unsigned i = 0; i != WordsToMove; ++i) { |
2741 | Dst[i] = Dst[i + WordShift] >> BitShift; |
2742 | if (i + 1 != WordsToMove) |
2743 | Dst[i] |= Dst[i + WordShift + 1] << (APINT_BITS_PER_WORD - BitShift); |
2744 | } |
2745 | } |
2746 | |
2747 | // Fill in the remainder with 0s. |
2748 | std::memset(s: Dst + WordsToMove, c: 0, n: WordShift * APINT_WORD_SIZE); |
2749 | } |
2750 | |
2751 | // Comparison (unsigned) of two bignums. |
2752 | int APInt::tcCompare(const WordType *lhs, const WordType *rhs, |
2753 | unsigned parts) { |
2754 | while (parts) { |
2755 | parts--; |
2756 | if (lhs[parts] != rhs[parts]) |
2757 | return (lhs[parts] > rhs[parts]) ? 1 : -1; |
2758 | } |
2759 | |
2760 | return 0; |
2761 | } |
2762 | |
2763 | APInt llvm::APIntOps::RoundingUDiv(const APInt &A, const APInt &B, |
2764 | APInt::Rounding RM) { |
2765 | // Currently udivrem always rounds down. |
2766 | switch (RM) { |
2767 | case APInt::Rounding::DOWN: |
2768 | case APInt::Rounding::TOWARD_ZERO: |
2769 | return A.udiv(RHS: B); |
2770 | case APInt::Rounding::UP: { |
2771 | APInt Quo, Rem; |
2772 | APInt::udivrem(LHS: A, RHS: B, Quotient&: Quo, Remainder&: Rem); |
2773 | if (Rem.isZero()) |
2774 | return Quo; |
2775 | return Quo + 1; |
2776 | } |
2777 | } |
2778 | llvm_unreachable("Unknown APInt::Rounding enum" ); |
2779 | } |
2780 | |
2781 | APInt llvm::APIntOps::RoundingSDiv(const APInt &A, const APInt &B, |
2782 | APInt::Rounding RM) { |
2783 | switch (RM) { |
2784 | case APInt::Rounding::DOWN: |
2785 | case APInt::Rounding::UP: { |
2786 | APInt Quo, Rem; |
2787 | APInt::sdivrem(LHS: A, RHS: B, Quotient&: Quo, Remainder&: Rem); |
2788 | if (Rem.isZero()) |
2789 | return Quo; |
2790 | // This algorithm deals with arbitrary rounding mode used by sdivrem. |
2791 | // We want to check whether the non-integer part of the mathematical value |
2792 | // is negative or not. If the non-integer part is negative, we need to round |
2793 | // down from Quo; otherwise, if it's positive or 0, we return Quo, as it's |
2794 | // already rounded down. |
2795 | if (RM == APInt::Rounding::DOWN) { |
2796 | if (Rem.isNegative() != B.isNegative()) |
2797 | return Quo - 1; |
2798 | return Quo; |
2799 | } |
2800 | if (Rem.isNegative() != B.isNegative()) |
2801 | return Quo; |
2802 | return Quo + 1; |
2803 | } |
2804 | // Currently sdiv rounds towards zero. |
2805 | case APInt::Rounding::TOWARD_ZERO: |
2806 | return A.sdiv(RHS: B); |
2807 | } |
2808 | llvm_unreachable("Unknown APInt::Rounding enum" ); |
2809 | } |
2810 | |
2811 | std::optional<APInt> |
2812 | llvm::APIntOps::SolveQuadraticEquationWrap(APInt A, APInt B, APInt C, |
2813 | unsigned RangeWidth) { |
2814 | unsigned CoeffWidth = A.getBitWidth(); |
2815 | assert(CoeffWidth == B.getBitWidth() && CoeffWidth == C.getBitWidth()); |
2816 | assert(RangeWidth <= CoeffWidth && |
2817 | "Value range width should be less than coefficient width" ); |
2818 | assert(RangeWidth > 1 && "Value range bit width should be > 1" ); |
2819 | |
2820 | LLVM_DEBUG(dbgs() << __func__ << ": solving " << A << "x^2 + " << B |
2821 | << "x + " << C << ", rw:" << RangeWidth << '\n'); |
2822 | |
2823 | // Identify 0 as a (non)solution immediately. |
2824 | if (C.sextOrTrunc(width: RangeWidth).isZero()) { |
2825 | LLVM_DEBUG(dbgs() << __func__ << ": zero solution\n" ); |
2826 | return APInt(CoeffWidth, 0); |
2827 | } |
2828 | |
2829 | // The result of APInt arithmetic has the same bit width as the operands, |
2830 | // so it can actually lose high bits. A product of two n-bit integers needs |
2831 | // 2n-1 bits to represent the full value. |
2832 | // The operation done below (on quadratic coefficients) that can produce |
2833 | // the largest value is the evaluation of the equation during bisection, |
2834 | // which needs 3 times the bitwidth of the coefficient, so the total number |
2835 | // of required bits is 3n. |
2836 | // |
2837 | // The purpose of this extension is to simulate the set Z of all integers, |
2838 | // where n+1 > n for all n in Z. In Z it makes sense to talk about positive |
2839 | // and negative numbers (not so much in a modulo arithmetic). The method |
2840 | // used to solve the equation is based on the standard formula for real |
2841 | // numbers, and uses the concepts of "positive" and "negative" with their |
2842 | // usual meanings. |
2843 | CoeffWidth *= 3; |
2844 | A = A.sext(Width: CoeffWidth); |
2845 | B = B.sext(Width: CoeffWidth); |
2846 | C = C.sext(Width: CoeffWidth); |
2847 | |
2848 | // Make A > 0 for simplicity. Negate cannot overflow at this point because |
2849 | // the bit width has increased. |
2850 | if (A.isNegative()) { |
2851 | A.negate(); |
2852 | B.negate(); |
2853 | C.negate(); |
2854 | } |
2855 | |
2856 | // Solving an equation q(x) = 0 with coefficients in modular arithmetic |
2857 | // is really solving a set of equations q(x) = kR for k = 0, 1, 2, ..., |
2858 | // and R = 2^BitWidth. |
2859 | // Since we're trying not only to find exact solutions, but also values |
2860 | // that "wrap around", such a set will always have a solution, i.e. an x |
2861 | // that satisfies at least one of the equations, or such that |q(x)| |
2862 | // exceeds kR, while |q(x-1)| for the same k does not. |
2863 | // |
2864 | // We need to find a value k, such that Ax^2 + Bx + C = kR will have a |
2865 | // positive solution n (in the above sense), and also such that the n |
2866 | // will be the least among all solutions corresponding to k = 0, 1, ... |
2867 | // (more precisely, the least element in the set |
2868 | // { n(k) | k is such that a solution n(k) exists }). |
2869 | // |
2870 | // Consider the parabola (over real numbers) that corresponds to the |
2871 | // quadratic equation. Since A > 0, the arms of the parabola will point |
2872 | // up. Picking different values of k will shift it up and down by R. |
2873 | // |
2874 | // We want to shift the parabola in such a way as to reduce the problem |
2875 | // of solving q(x) = kR to solving shifted_q(x) = 0. |
2876 | // (The interesting solutions are the ceilings of the real number |
2877 | // solutions.) |
2878 | APInt R = APInt::getOneBitSet(numBits: CoeffWidth, BitNo: RangeWidth); |
2879 | APInt TwoA = 2 * A; |
2880 | APInt SqrB = B * B; |
2881 | bool PickLow; |
2882 | |
2883 | auto RoundUp = [] (const APInt &V, const APInt &A) -> APInt { |
2884 | assert(A.isStrictlyPositive()); |
2885 | APInt T = V.abs().urem(RHS: A); |
2886 | if (T.isZero()) |
2887 | return V; |
2888 | return V.isNegative() ? V+T : V+(A-T); |
2889 | }; |
2890 | |
2891 | // The vertex of the parabola is at -B/2A, but since A > 0, it's negative |
2892 | // iff B is positive. |
2893 | if (B.isNonNegative()) { |
2894 | // If B >= 0, the vertex it at a negative location (or at 0), so in |
2895 | // order to have a non-negative solution we need to pick k that makes |
2896 | // C-kR negative. To satisfy all the requirements for the solution |
2897 | // that we are looking for, it needs to be closest to 0 of all k. |
2898 | C = C.srem(RHS: R); |
2899 | if (C.isStrictlyPositive()) |
2900 | C -= R; |
2901 | // Pick the greater solution. |
2902 | PickLow = false; |
2903 | } else { |
2904 | // If B < 0, the vertex is at a positive location. For any solution |
2905 | // to exist, the discriminant must be non-negative. This means that |
2906 | // C-kR <= B^2/4A is a necessary condition for k, i.e. there is a |
2907 | // lower bound on values of k: kR >= C - B^2/4A. |
2908 | APInt LowkR = C - SqrB.udiv(RHS: 2*TwoA); // udiv because all values > 0. |
2909 | // Round LowkR up (towards +inf) to the nearest kR. |
2910 | LowkR = RoundUp(LowkR, R); |
2911 | |
2912 | // If there exists k meeting the condition above, and such that |
2913 | // C-kR > 0, there will be two positive real number solutions of |
2914 | // q(x) = kR. Out of all such values of k, pick the one that makes |
2915 | // C-kR closest to 0, (i.e. pick maximum k such that C-kR > 0). |
2916 | // In other words, find maximum k such that LowkR <= kR < C. |
2917 | if (C.sgt(RHS: LowkR)) { |
2918 | // If LowkR < C, then such a k is guaranteed to exist because |
2919 | // LowkR itself is a multiple of R. |
2920 | C -= -RoundUp(-C, R); // C = C - RoundDown(C, R) |
2921 | // Pick the smaller solution. |
2922 | PickLow = true; |
2923 | } else { |
2924 | // If C-kR < 0 for all potential k's, it means that one solution |
2925 | // will be negative, while the other will be positive. The positive |
2926 | // solution will shift towards 0 if the parabola is moved up. |
2927 | // Pick the kR closest to the lower bound (i.e. make C-kR closest |
2928 | // to 0, or in other words, out of all parabolas that have solutions, |
2929 | // pick the one that is the farthest "up"). |
2930 | // Since LowkR is itself a multiple of R, simply take C-LowkR. |
2931 | C -= LowkR; |
2932 | // Pick the greater solution. |
2933 | PickLow = false; |
2934 | } |
2935 | } |
2936 | |
2937 | LLVM_DEBUG(dbgs() << __func__ << ": updated coefficients " << A << "x^2 + " |
2938 | << B << "x + " << C << ", rw:" << RangeWidth << '\n'); |
2939 | |
2940 | APInt D = SqrB - 4*A*C; |
2941 | assert(D.isNonNegative() && "Negative discriminant" ); |
2942 | APInt SQ = D.sqrt(); |
2943 | |
2944 | APInt Q = SQ * SQ; |
2945 | bool InexactSQ = Q != D; |
2946 | // The calculated SQ may actually be greater than the exact (non-integer) |
2947 | // value. If that's the case, decrement SQ to get a value that is lower. |
2948 | if (Q.sgt(RHS: D)) |
2949 | SQ -= 1; |
2950 | |
2951 | APInt X; |
2952 | APInt Rem; |
2953 | |
2954 | // SQ is rounded down (i.e SQ * SQ <= D), so the roots may be inexact. |
2955 | // When using the quadratic formula directly, the calculated low root |
2956 | // may be greater than the exact one, since we would be subtracting SQ. |
2957 | // To make sure that the calculated root is not greater than the exact |
2958 | // one, subtract SQ+1 when calculating the low root (for inexact value |
2959 | // of SQ). |
2960 | if (PickLow) |
2961 | APInt::sdivrem(LHS: -B - (SQ+InexactSQ), RHS: TwoA, Quotient&: X, Remainder&: Rem); |
2962 | else |
2963 | APInt::sdivrem(LHS: -B + SQ, RHS: TwoA, Quotient&: X, Remainder&: Rem); |
2964 | |
2965 | // The updated coefficients should be such that the (exact) solution is |
2966 | // positive. Since APInt division rounds towards 0, the calculated one |
2967 | // can be 0, but cannot be negative. |
2968 | assert(X.isNonNegative() && "Solution should be non-negative" ); |
2969 | |
2970 | if (!InexactSQ && Rem.isZero()) { |
2971 | LLVM_DEBUG(dbgs() << __func__ << ": solution (root): " << X << '\n'); |
2972 | return X; |
2973 | } |
2974 | |
2975 | assert((SQ*SQ).sle(D) && "SQ = |_sqrt(D)_|, so SQ*SQ <= D" ); |
2976 | // The exact value of the square root of D should be between SQ and SQ+1. |
2977 | // This implies that the solution should be between that corresponding to |
2978 | // SQ (i.e. X) and that corresponding to SQ+1. |
2979 | // |
2980 | // The calculated X cannot be greater than the exact (real) solution. |
2981 | // Actually it must be strictly less than the exact solution, while |
2982 | // X+1 will be greater than or equal to it. |
2983 | |
2984 | APInt VX = (A*X + B)*X + C; |
2985 | APInt VY = VX + TwoA*X + A + B; |
2986 | bool SignChange = |
2987 | VX.isNegative() != VY.isNegative() || VX.isZero() != VY.isZero(); |
2988 | // If the sign did not change between X and X+1, X is not a valid solution. |
2989 | // This could happen when the actual (exact) roots don't have an integer |
2990 | // between them, so they would both be contained between X and X+1. |
2991 | if (!SignChange) { |
2992 | LLVM_DEBUG(dbgs() << __func__ << ": no valid solution\n" ); |
2993 | return std::nullopt; |
2994 | } |
2995 | |
2996 | X += 1; |
2997 | LLVM_DEBUG(dbgs() << __func__ << ": solution (wrap): " << X << '\n'); |
2998 | return X; |
2999 | } |
3000 | |
3001 | std::optional<unsigned> |
3002 | llvm::APIntOps::GetMostSignificantDifferentBit(const APInt &A, const APInt &B) { |
3003 | assert(A.getBitWidth() == B.getBitWidth() && "Must have the same bitwidth" ); |
3004 | if (A == B) |
3005 | return std::nullopt; |
3006 | return A.getBitWidth() - ((A ^ B).countl_zero() + 1); |
3007 | } |
3008 | |
3009 | APInt llvm::APIntOps::ScaleBitMask(const APInt &A, unsigned NewBitWidth, |
3010 | bool MatchAllBits) { |
3011 | unsigned OldBitWidth = A.getBitWidth(); |
3012 | assert((((OldBitWidth % NewBitWidth) == 0) || |
3013 | ((NewBitWidth % OldBitWidth) == 0)) && |
3014 | "One size should be a multiple of the other one. " |
3015 | "Can't do fractional scaling." ); |
3016 | |
3017 | // Check for matching bitwidths. |
3018 | if (OldBitWidth == NewBitWidth) |
3019 | return A; |
3020 | |
3021 | APInt NewA = APInt::getZero(numBits: NewBitWidth); |
3022 | |
3023 | // Check for null input. |
3024 | if (A.isZero()) |
3025 | return NewA; |
3026 | |
3027 | if (NewBitWidth > OldBitWidth) { |
3028 | // Repeat bits. |
3029 | unsigned Scale = NewBitWidth / OldBitWidth; |
3030 | for (unsigned i = 0; i != OldBitWidth; ++i) |
3031 | if (A[i]) |
3032 | NewA.setBits(loBit: i * Scale, hiBit: (i + 1) * Scale); |
3033 | } else { |
3034 | unsigned Scale = OldBitWidth / NewBitWidth; |
3035 | for (unsigned i = 0; i != NewBitWidth; ++i) { |
3036 | if (MatchAllBits) { |
3037 | if (A.extractBits(numBits: Scale, bitPosition: i * Scale).isAllOnes()) |
3038 | NewA.setBit(i); |
3039 | } else { |
3040 | if (!A.extractBits(numBits: Scale, bitPosition: i * Scale).isZero()) |
3041 | NewA.setBit(i); |
3042 | } |
3043 | } |
3044 | } |
3045 | |
3046 | return NewA; |
3047 | } |
3048 | |
3049 | /// StoreIntToMemory - Fills the StoreBytes bytes of memory starting from Dst |
3050 | /// with the integer held in IntVal. |
3051 | void llvm::StoreIntToMemory(const APInt &IntVal, uint8_t *Dst, |
3052 | unsigned StoreBytes) { |
3053 | assert((IntVal.getBitWidth()+7)/8 >= StoreBytes && "Integer too small!" ); |
3054 | const uint8_t *Src = (const uint8_t *)IntVal.getRawData(); |
3055 | |
3056 | if (sys::IsLittleEndianHost) { |
3057 | // Little-endian host - the source is ordered from LSB to MSB. Order the |
3058 | // destination from LSB to MSB: Do a straight copy. |
3059 | memcpy(dest: Dst, src: Src, n: StoreBytes); |
3060 | } else { |
3061 | // Big-endian host - the source is an array of 64 bit words ordered from |
3062 | // LSW to MSW. Each word is ordered from MSB to LSB. Order the destination |
3063 | // from MSB to LSB: Reverse the word order, but not the bytes in a word. |
3064 | while (StoreBytes > sizeof(uint64_t)) { |
3065 | StoreBytes -= sizeof(uint64_t); |
3066 | // May not be aligned so use memcpy. |
3067 | memcpy(dest: Dst + StoreBytes, src: Src, n: sizeof(uint64_t)); |
3068 | Src += sizeof(uint64_t); |
3069 | } |
3070 | |
3071 | memcpy(dest: Dst, src: Src + sizeof(uint64_t) - StoreBytes, n: StoreBytes); |
3072 | } |
3073 | } |
3074 | |
3075 | /// LoadIntFromMemory - Loads the integer stored in the LoadBytes bytes starting |
3076 | /// from Src into IntVal, which is assumed to be wide enough and to hold zero. |
3077 | void llvm::LoadIntFromMemory(APInt &IntVal, const uint8_t *Src, |
3078 | unsigned LoadBytes) { |
3079 | assert((IntVal.getBitWidth()+7)/8 >= LoadBytes && "Integer too small!" ); |
3080 | uint8_t *Dst = reinterpret_cast<uint8_t *>( |
3081 | const_cast<uint64_t *>(IntVal.getRawData())); |
3082 | |
3083 | if (sys::IsLittleEndianHost) |
3084 | // Little-endian host - the destination must be ordered from LSB to MSB. |
3085 | // The source is ordered from LSB to MSB: Do a straight copy. |
3086 | memcpy(dest: Dst, src: Src, n: LoadBytes); |
3087 | else { |
3088 | // Big-endian - the destination is an array of 64 bit words ordered from |
3089 | // LSW to MSW. Each word must be ordered from MSB to LSB. The source is |
3090 | // ordered from MSB to LSB: Reverse the word order, but not the bytes in |
3091 | // a word. |
3092 | while (LoadBytes > sizeof(uint64_t)) { |
3093 | LoadBytes -= sizeof(uint64_t); |
3094 | // May not be aligned so use memcpy. |
3095 | memcpy(dest: Dst, src: Src + LoadBytes, n: sizeof(uint64_t)); |
3096 | Dst += sizeof(uint64_t); |
3097 | } |
3098 | |
3099 | memcpy(dest: Dst + sizeof(uint64_t) - LoadBytes, src: Src, n: LoadBytes); |
3100 | } |
3101 | } |
3102 | |
3103 | APInt APIntOps::avgFloorS(const APInt &C1, const APInt &C2) { |
3104 | // Return floor((C1 + C2) / 2) |
3105 | return (C1 & C2) + (C1 ^ C2).ashr(ShiftAmt: 1); |
3106 | } |
3107 | |
3108 | APInt APIntOps::avgFloorU(const APInt &C1, const APInt &C2) { |
3109 | // Return floor((C1 + C2) / 2) |
3110 | return (C1 & C2) + (C1 ^ C2).lshr(shiftAmt: 1); |
3111 | } |
3112 | |
3113 | APInt APIntOps::avgCeilS(const APInt &C1, const APInt &C2) { |
3114 | // Return ceil((C1 + C2) / 2) |
3115 | return (C1 | C2) - (C1 ^ C2).ashr(ShiftAmt: 1); |
3116 | } |
3117 | |
3118 | APInt APIntOps::avgCeilU(const APInt &C1, const APInt &C2) { |
3119 | // Return ceil((C1 + C2) / 2) |
3120 | return (C1 | C2) - (C1 ^ C2).lshr(shiftAmt: 1); |
3121 | } |
3122 | |
3123 | APInt APIntOps::mulhs(const APInt &C1, const APInt &C2) { |
3124 | assert(C1.getBitWidth() == C2.getBitWidth() && "Unequal bitwidths" ); |
3125 | unsigned FullWidth = C1.getBitWidth() * 2; |
3126 | APInt C1Ext = C1.sext(Width: FullWidth); |
3127 | APInt C2Ext = C2.sext(Width: FullWidth); |
3128 | return (C1Ext * C2Ext).extractBits(numBits: C1.getBitWidth(), bitPosition: C1.getBitWidth()); |
3129 | } |
3130 | |
3131 | APInt APIntOps::mulhu(const APInt &C1, const APInt &C2) { |
3132 | assert(C1.getBitWidth() == C2.getBitWidth() && "Unequal bitwidths" ); |
3133 | unsigned FullWidth = C1.getBitWidth() * 2; |
3134 | APInt C1Ext = C1.zext(width: FullWidth); |
3135 | APInt C2Ext = C2.zext(width: FullWidth); |
3136 | return (C1Ext * C2Ext).extractBits(numBits: C1.getBitWidth(), bitPosition: C1.getBitWidth()); |
3137 | } |
3138 | |
3139 | APInt APIntOps::pow(const APInt &X, int64_t N) { |
3140 | assert(N >= 0 && "negative exponents not supported." ); |
3141 | APInt Acc = APInt(X.getBitWidth(), 1); |
3142 | if (N == 0) |
3143 | return Acc; |
3144 | APInt Base = X; |
3145 | int64_t RemainingExponent = N; |
3146 | while (RemainingExponent > 0) { |
3147 | while (RemainingExponent % 2 == 0) { |
3148 | Base *= Base; |
3149 | RemainingExponent /= 2; |
3150 | } |
3151 | --RemainingExponent; |
3152 | Acc *= Base; |
3153 | } |
3154 | return Acc; |
3155 | } |
3156 | |