1 //===- ValueTracking.cpp - Walk computations to compute properties --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains routines that help analyze properties that chains of 11 // computations have. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Analysis/ValueTracking.h" 16 #include "llvm/ADT/Optional.h" 17 #include "llvm/ADT/SmallPtrSet.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/MemoryBuiltins.h" 21 #include "llvm/Analysis/Loads.h" 22 #include "llvm/Analysis/LoopInfo.h" 23 #include "llvm/Analysis/VectorUtils.h" 24 #include "llvm/IR/CallSite.h" 25 #include "llvm/IR/ConstantRange.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DataLayout.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/GetElementPtrTypeIterator.h" 30 #include "llvm/IR/GlobalAlias.h" 31 #include "llvm/IR/GlobalVariable.h" 32 #include "llvm/IR/Instructions.h" 33 #include "llvm/IR/IntrinsicInst.h" 34 #include "llvm/IR/LLVMContext.h" 35 #include "llvm/IR/Metadata.h" 36 #include "llvm/IR/Operator.h" 37 #include "llvm/IR/PatternMatch.h" 38 #include "llvm/IR/Statepoint.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/MathExtras.h" 41 #include <algorithm> 42 #include <array> 43 #include <cstring> 44 using namespace llvm; 45 using namespace llvm::PatternMatch; 46 47 const unsigned MaxDepth = 6; 48 49 // Controls the number of uses of the value searched for possible 50 // dominating comparisons. 51 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", 52 cl::Hidden, cl::init(20)); 53 54 /// Returns the bitwidth of the given scalar or pointer type (if unknown returns 55 /// 0). For vector types, returns the element type's bitwidth. 56 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { 57 if (unsigned BitWidth = Ty->getScalarSizeInBits()) 58 return BitWidth; 59 60 return DL.getPointerTypeSizeInBits(Ty); 61 } 62 63 namespace { 64 // Simplifying using an assume can only be done in a particular control-flow 65 // context (the context instruction provides that context). If an assume and 66 // the context instruction are not in the same block then the DT helps in 67 // figuring out if we can use it. 68 struct Query { 69 const DataLayout &DL; 70 AssumptionCache *AC; 71 const Instruction *CxtI; 72 const DominatorTree *DT; 73 74 /// Set of assumptions that should be excluded from further queries. 75 /// This is because of the potential for mutual recursion to cause 76 /// computeKnownBits to repeatedly visit the same assume intrinsic. The 77 /// classic case of this is assume(x = y), which will attempt to determine 78 /// bits in x from bits in y, which will attempt to determine bits in y from 79 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call 80 /// isKnownNonZero, which calls computeKnownBits and ComputeSignBit and 81 /// isKnownToBeAPowerOfTwo (all of which can call computeKnownBits), and so 82 /// on. 83 std::array<const Value*, MaxDepth> Excluded; 84 unsigned NumExcluded; 85 86 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, 87 const DominatorTree *DT) 88 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), NumExcluded(0) {} 89 90 Query(const Query &Q, const Value *NewExcl) 91 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), NumExcluded(Q.NumExcluded) { 92 Excluded = Q.Excluded; 93 Excluded[NumExcluded++] = NewExcl; 94 assert(NumExcluded <= Excluded.size()); 95 } 96 97 bool isExcluded(const Value *Value) const { 98 if (NumExcluded == 0) 99 return false; 100 auto End = Excluded.begin() + NumExcluded; 101 return std::find(Excluded.begin(), End, Value) != End; 102 } 103 }; 104 } // end anonymous namespace 105 106 // Given the provided Value and, potentially, a context instruction, return 107 // the preferred context instruction (if any). 108 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { 109 // If we've been provided with a context instruction, then use that (provided 110 // it has been inserted). 111 if (CxtI && CxtI->getParent()) 112 return CxtI; 113 114 // If the value is really an already-inserted instruction, then use that. 115 CxtI = dyn_cast<Instruction>(V); 116 if (CxtI && CxtI->getParent()) 117 return CxtI; 118 119 return nullptr; 120 } 121 122 static void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 123 unsigned Depth, const Query &Q); 124 125 void llvm::computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 126 const DataLayout &DL, unsigned Depth, 127 AssumptionCache *AC, const Instruction *CxtI, 128 const DominatorTree *DT) { 129 ::computeKnownBits(V, KnownZero, KnownOne, Depth, 130 Query(DL, AC, safeCxtI(V, CxtI), DT)); 131 } 132 133 bool llvm::haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL, 134 AssumptionCache *AC, const Instruction *CxtI, 135 const DominatorTree *DT) { 136 assert(LHS->getType() == RHS->getType() && 137 "LHS and RHS should have the same type"); 138 assert(LHS->getType()->isIntOrIntVectorTy() && 139 "LHS and RHS should be integers"); 140 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); 141 APInt LHSKnownZero(IT->getBitWidth(), 0), LHSKnownOne(IT->getBitWidth(), 0); 142 APInt RHSKnownZero(IT->getBitWidth(), 0), RHSKnownOne(IT->getBitWidth(), 0); 143 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, 0, AC, CxtI, DT); 144 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, 0, AC, CxtI, DT); 145 return (LHSKnownZero | RHSKnownZero).isAllOnesValue(); 146 } 147 148 static void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 149 unsigned Depth, const Query &Q); 150 151 void llvm::ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 152 const DataLayout &DL, unsigned Depth, 153 AssumptionCache *AC, const Instruction *CxtI, 154 const DominatorTree *DT) { 155 ::ComputeSignBit(V, KnownZero, KnownOne, Depth, 156 Query(DL, AC, safeCxtI(V, CxtI), DT)); 157 } 158 159 static bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth, 160 const Query &Q); 161 162 bool llvm::isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, bool OrZero, 163 unsigned Depth, AssumptionCache *AC, 164 const Instruction *CxtI, 165 const DominatorTree *DT) { 166 return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth, 167 Query(DL, AC, safeCxtI(V, CxtI), DT)); 168 } 169 170 static bool isKnownNonZero(Value *V, unsigned Depth, const Query &Q); 171 172 bool llvm::isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth, 173 AssumptionCache *AC, const Instruction *CxtI, 174 const DominatorTree *DT) { 175 return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT)); 176 } 177 178 bool llvm::isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth, 179 AssumptionCache *AC, const Instruction *CxtI, 180 const DominatorTree *DT) { 181 bool NonNegative, Negative; 182 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT); 183 return NonNegative; 184 } 185 186 bool llvm::isKnownPositive(Value *V, const DataLayout &DL, unsigned Depth, 187 AssumptionCache *AC, const Instruction *CxtI, 188 const DominatorTree *DT) { 189 if (auto *CI = dyn_cast<ConstantInt>(V)) 190 return CI->getValue().isStrictlyPositive(); 191 192 // TODO: We'd doing two recursive queries here. We should factor this such 193 // that only a single query is needed. 194 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) && 195 isKnownNonZero(V, DL, Depth, AC, CxtI, DT); 196 } 197 198 bool llvm::isKnownNegative(Value *V, const DataLayout &DL, unsigned Depth, 199 AssumptionCache *AC, const Instruction *CxtI, 200 const DominatorTree *DT) { 201 bool NonNegative, Negative; 202 ComputeSignBit(V, NonNegative, Negative, DL, Depth, AC, CxtI, DT); 203 return Negative; 204 } 205 206 static bool isKnownNonEqual(Value *V1, Value *V2, const Query &Q); 207 208 bool llvm::isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL, 209 AssumptionCache *AC, const Instruction *CxtI, 210 const DominatorTree *DT) { 211 return ::isKnownNonEqual(V1, V2, Query(DL, AC, 212 safeCxtI(V1, safeCxtI(V2, CxtI)), 213 DT)); 214 } 215 216 static bool MaskedValueIsZero(Value *V, const APInt &Mask, unsigned Depth, 217 const Query &Q); 218 219 bool llvm::MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL, 220 unsigned Depth, AssumptionCache *AC, 221 const Instruction *CxtI, const DominatorTree *DT) { 222 return ::MaskedValueIsZero(V, Mask, Depth, 223 Query(DL, AC, safeCxtI(V, CxtI), DT)); 224 } 225 226 static unsigned ComputeNumSignBits(Value *V, unsigned Depth, const Query &Q); 227 228 unsigned llvm::ComputeNumSignBits(Value *V, const DataLayout &DL, 229 unsigned Depth, AssumptionCache *AC, 230 const Instruction *CxtI, 231 const DominatorTree *DT) { 232 return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT)); 233 } 234 235 static void computeKnownBitsAddSub(bool Add, Value *Op0, Value *Op1, bool NSW, 236 APInt &KnownZero, APInt &KnownOne, 237 APInt &KnownZero2, APInt &KnownOne2, 238 unsigned Depth, const Query &Q) { 239 if (!Add) { 240 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(Op0)) { 241 // We know that the top bits of C-X are clear if X contains less bits 242 // than C (i.e. no wrap-around can happen). For example, 20-X is 243 // positive if we can prove that X is >= 0 and < 16. 244 if (!CLHS->getValue().isNegative()) { 245 unsigned BitWidth = KnownZero.getBitWidth(); 246 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros(); 247 // NLZ can't be BitWidth with no sign bit 248 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1); 249 computeKnownBits(Op1, KnownZero2, KnownOne2, Depth + 1, Q); 250 251 // If all of the MaskV bits are known to be zero, then we know the 252 // output top bits are zero, because we now know that the output is 253 // from [0-C]. 254 if ((KnownZero2 & MaskV) == MaskV) { 255 unsigned NLZ2 = CLHS->getValue().countLeadingZeros(); 256 // Top bits known zero. 257 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2); 258 } 259 } 260 } 261 } 262 263 unsigned BitWidth = KnownZero.getBitWidth(); 264 265 // If an initial sequence of bits in the result is not needed, the 266 // corresponding bits in the operands are not needed. 267 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0); 268 computeKnownBits(Op0, LHSKnownZero, LHSKnownOne, Depth + 1, Q); 269 computeKnownBits(Op1, KnownZero2, KnownOne2, Depth + 1, Q); 270 271 // Carry in a 1 for a subtract, rather than a 0. 272 APInt CarryIn(BitWidth, 0); 273 if (!Add) { 274 // Sum = LHS + ~RHS + 1 275 std::swap(KnownZero2, KnownOne2); 276 CarryIn.setBit(0); 277 } 278 279 APInt PossibleSumZero = ~LHSKnownZero + ~KnownZero2 + CarryIn; 280 APInt PossibleSumOne = LHSKnownOne + KnownOne2 + CarryIn; 281 282 // Compute known bits of the carry. 283 APInt CarryKnownZero = ~(PossibleSumZero ^ LHSKnownZero ^ KnownZero2); 284 APInt CarryKnownOne = PossibleSumOne ^ LHSKnownOne ^ KnownOne2; 285 286 // Compute set of known bits (where all three relevant bits are known). 287 APInt LHSKnown = LHSKnownZero | LHSKnownOne; 288 APInt RHSKnown = KnownZero2 | KnownOne2; 289 APInt CarryKnown = CarryKnownZero | CarryKnownOne; 290 APInt Known = LHSKnown & RHSKnown & CarryKnown; 291 292 assert((PossibleSumZero & Known) == (PossibleSumOne & Known) && 293 "known bits of sum differ"); 294 295 // Compute known bits of the result. 296 KnownZero = ~PossibleSumOne & Known; 297 KnownOne = PossibleSumOne & Known; 298 299 // Are we still trying to solve for the sign bit? 300 if (!Known.isNegative()) { 301 if (NSW) { 302 // Adding two non-negative numbers, or subtracting a negative number from 303 // a non-negative one, can't wrap into negative. 304 if (LHSKnownZero.isNegative() && KnownZero2.isNegative()) 305 KnownZero |= APInt::getSignBit(BitWidth); 306 // Adding two negative numbers, or subtracting a non-negative number from 307 // a negative one, can't wrap into non-negative. 308 else if (LHSKnownOne.isNegative() && KnownOne2.isNegative()) 309 KnownOne |= APInt::getSignBit(BitWidth); 310 } 311 } 312 } 313 314 static void computeKnownBitsMul(Value *Op0, Value *Op1, bool NSW, 315 APInt &KnownZero, APInt &KnownOne, 316 APInt &KnownZero2, APInt &KnownOne2, 317 unsigned Depth, const Query &Q) { 318 unsigned BitWidth = KnownZero.getBitWidth(); 319 computeKnownBits(Op1, KnownZero, KnownOne, Depth + 1, Q); 320 computeKnownBits(Op0, KnownZero2, KnownOne2, Depth + 1, Q); 321 322 bool isKnownNegative = false; 323 bool isKnownNonNegative = false; 324 // If the multiplication is known not to overflow, compute the sign bit. 325 if (NSW) { 326 if (Op0 == Op1) { 327 // The product of a number with itself is non-negative. 328 isKnownNonNegative = true; 329 } else { 330 bool isKnownNonNegativeOp1 = KnownZero.isNegative(); 331 bool isKnownNonNegativeOp0 = KnownZero2.isNegative(); 332 bool isKnownNegativeOp1 = KnownOne.isNegative(); 333 bool isKnownNegativeOp0 = KnownOne2.isNegative(); 334 // The product of two numbers with the same sign is non-negative. 335 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || 336 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); 337 // The product of a negative number and a non-negative number is either 338 // negative or zero. 339 if (!isKnownNonNegative) 340 isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 && 341 isKnownNonZero(Op0, Depth, Q)) || 342 (isKnownNegativeOp0 && isKnownNonNegativeOp1 && 343 isKnownNonZero(Op1, Depth, Q)); 344 } 345 } 346 347 // If low bits are zero in either operand, output low known-0 bits. 348 // Also compute a conservative estimate for high known-0 bits. 349 // More trickiness is possible, but this is sufficient for the 350 // interesting case of alignment computation. 351 KnownOne.clearAllBits(); 352 unsigned TrailZ = KnownZero.countTrailingOnes() + 353 KnownZero2.countTrailingOnes(); 354 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() + 355 KnownZero2.countLeadingOnes(), 356 BitWidth) - BitWidth; 357 358 TrailZ = std::min(TrailZ, BitWidth); 359 LeadZ = std::min(LeadZ, BitWidth); 360 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) | 361 APInt::getHighBitsSet(BitWidth, LeadZ); 362 363 // Only make use of no-wrap flags if we failed to compute the sign bit 364 // directly. This matters if the multiplication always overflows, in 365 // which case we prefer to follow the result of the direct computation, 366 // though as the program is invoking undefined behaviour we can choose 367 // whatever we like here. 368 if (isKnownNonNegative && !KnownOne.isNegative()) 369 KnownZero.setBit(BitWidth - 1); 370 else if (isKnownNegative && !KnownZero.isNegative()) 371 KnownOne.setBit(BitWidth - 1); 372 } 373 374 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, 375 APInt &KnownZero, 376 APInt &KnownOne) { 377 unsigned BitWidth = KnownZero.getBitWidth(); 378 unsigned NumRanges = Ranges.getNumOperands() / 2; 379 assert(NumRanges >= 1); 380 381 KnownZero.setAllBits(); 382 KnownOne.setAllBits(); 383 384 for (unsigned i = 0; i < NumRanges; ++i) { 385 ConstantInt *Lower = 386 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); 387 ConstantInt *Upper = 388 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); 389 ConstantRange Range(Lower->getValue(), Upper->getValue()); 390 391 // The first CommonPrefixBits of all values in Range are equal. 392 unsigned CommonPrefixBits = 393 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros(); 394 395 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); 396 KnownOne &= Range.getUnsignedMax() & Mask; 397 KnownZero &= ~Range.getUnsignedMax() & Mask; 398 } 399 } 400 401 static bool isEphemeralValueOf(Instruction *I, const Value *E) { 402 SmallVector<const Value *, 16> WorkSet(1, I); 403 SmallPtrSet<const Value *, 32> Visited; 404 SmallPtrSet<const Value *, 16> EphValues; 405 406 // The instruction defining an assumption's condition itself is always 407 // considered ephemeral to that assumption (even if it has other 408 // non-ephemeral users). See r246696's test case for an example. 409 if (std::find(I->op_begin(), I->op_end(), E) != I->op_end()) 410 return true; 411 412 while (!WorkSet.empty()) { 413 const Value *V = WorkSet.pop_back_val(); 414 if (!Visited.insert(V).second) 415 continue; 416 417 // If all uses of this value are ephemeral, then so is this value. 418 if (std::all_of(V->user_begin(), V->user_end(), 419 [&](const User *U) { return EphValues.count(U); })) { 420 if (V == E) 421 return true; 422 423 EphValues.insert(V); 424 if (const User *U = dyn_cast<User>(V)) 425 for (User::const_op_iterator J = U->op_begin(), JE = U->op_end(); 426 J != JE; ++J) { 427 if (isSafeToSpeculativelyExecute(*J)) 428 WorkSet.push_back(*J); 429 } 430 } 431 } 432 433 return false; 434 } 435 436 // Is this an intrinsic that cannot be speculated but also cannot trap? 437 static bool isAssumeLikeIntrinsic(const Instruction *I) { 438 if (const CallInst *CI = dyn_cast<CallInst>(I)) 439 if (Function *F = CI->getCalledFunction()) 440 switch (F->getIntrinsicID()) { 441 default: break; 442 // FIXME: This list is repeated from NoTTI::getIntrinsicCost. 443 case Intrinsic::assume: 444 case Intrinsic::dbg_declare: 445 case Intrinsic::dbg_value: 446 case Intrinsic::invariant_start: 447 case Intrinsic::invariant_end: 448 case Intrinsic::lifetime_start: 449 case Intrinsic::lifetime_end: 450 case Intrinsic::objectsize: 451 case Intrinsic::ptr_annotation: 452 case Intrinsic::var_annotation: 453 return true; 454 } 455 456 return false; 457 } 458 459 static bool isValidAssumeForContext(Value *V, const Instruction *CxtI, 460 const DominatorTree *DT) { 461 Instruction *Inv = cast<Instruction>(V); 462 463 // There are two restrictions on the use of an assume: 464 // 1. The assume must dominate the context (or the control flow must 465 // reach the assume whenever it reaches the context). 466 // 2. The context must not be in the assume's set of ephemeral values 467 // (otherwise we will use the assume to prove that the condition 468 // feeding the assume is trivially true, thus causing the removal of 469 // the assume). 470 471 if (DT) { 472 if (DT->dominates(Inv, CxtI)) { 473 return true; 474 } else if (Inv->getParent() == CxtI->getParent()) { 475 // The context comes first, but they're both in the same block. Make sure 476 // there is nothing in between that might interrupt the control flow. 477 for (BasicBlock::const_iterator I = 478 std::next(BasicBlock::const_iterator(CxtI)), 479 IE(Inv); I != IE; ++I) 480 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I)) 481 return false; 482 483 return !isEphemeralValueOf(Inv, CxtI); 484 } 485 486 return false; 487 } 488 489 // When we don't have a DT, we do a limited search... 490 if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { 491 return true; 492 } else if (Inv->getParent() == CxtI->getParent()) { 493 // Search forward from the assume until we reach the context (or the end 494 // of the block); the common case is that the assume will come first. 495 for (BasicBlock::iterator I = std::next(BasicBlock::iterator(Inv)), 496 IE = Inv->getParent()->end(); I != IE; ++I) 497 if (&*I == CxtI) 498 return true; 499 500 // The context must come first... 501 for (BasicBlock::const_iterator I = 502 std::next(BasicBlock::const_iterator(CxtI)), 503 IE(Inv); I != IE; ++I) 504 if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I)) 505 return false; 506 507 return !isEphemeralValueOf(Inv, CxtI); 508 } 509 510 return false; 511 } 512 513 bool llvm::isValidAssumeForContext(const Instruction *I, 514 const Instruction *CxtI, 515 const DominatorTree *DT) { 516 return ::isValidAssumeForContext(const_cast<Instruction *>(I), CxtI, DT); 517 } 518 519 static void computeKnownBitsFromAssume(Value *V, APInt &KnownZero, 520 APInt &KnownOne, unsigned Depth, 521 const Query &Q) { 522 // Use of assumptions is context-sensitive. If we don't have a context, we 523 // cannot use them! 524 if (!Q.AC || !Q.CxtI) 525 return; 526 527 unsigned BitWidth = KnownZero.getBitWidth(); 528 529 for (auto &AssumeVH : Q.AC->assumptions()) { 530 if (!AssumeVH) 531 continue; 532 CallInst *I = cast<CallInst>(AssumeVH); 533 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && 534 "Got assumption for the wrong function!"); 535 if (Q.isExcluded(I)) 536 continue; 537 538 // Warning: This loop can end up being somewhat performance sensetive. 539 // We're running this loop for once for each value queried resulting in a 540 // runtime of ~O(#assumes * #values). 541 542 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 543 "must be an assume intrinsic"); 544 545 Value *Arg = I->getArgOperand(0); 546 547 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 548 assert(BitWidth == 1 && "assume operand is not i1?"); 549 KnownZero.clearAllBits(); 550 KnownOne.setAllBits(); 551 return; 552 } 553 554 // The remaining tests are all recursive, so bail out if we hit the limit. 555 if (Depth == MaxDepth) 556 continue; 557 558 Value *A, *B; 559 auto m_V = m_CombineOr(m_Specific(V), 560 m_CombineOr(m_PtrToInt(m_Specific(V)), 561 m_BitCast(m_Specific(V)))); 562 563 CmpInst::Predicate Pred; 564 ConstantInt *C; 565 // assume(v = a) 566 if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) && 567 Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 568 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 569 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 570 KnownZero |= RHSKnownZero; 571 KnownOne |= RHSKnownOne; 572 // assume(v & b = a) 573 } else if (match(Arg, 574 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && 575 Pred == ICmpInst::ICMP_EQ && 576 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 577 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 578 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 579 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0); 580 computeKnownBits(B, MaskKnownZero, MaskKnownOne, Depth+1, Query(Q, I)); 581 582 // For those bits in the mask that are known to be one, we can propagate 583 // known bits from the RHS to V. 584 KnownZero |= RHSKnownZero & MaskKnownOne; 585 KnownOne |= RHSKnownOne & MaskKnownOne; 586 // assume(~(v & b) = a) 587 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), 588 m_Value(A))) && 589 Pred == ICmpInst::ICMP_EQ && 590 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 591 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 592 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 593 APInt MaskKnownZero(BitWidth, 0), MaskKnownOne(BitWidth, 0); 594 computeKnownBits(B, MaskKnownZero, MaskKnownOne, Depth+1, Query(Q, I)); 595 596 // For those bits in the mask that are known to be one, we can propagate 597 // inverted known bits from the RHS to V. 598 KnownZero |= RHSKnownOne & MaskKnownOne; 599 KnownOne |= RHSKnownZero & MaskKnownOne; 600 // assume(v | b = a) 601 } else if (match(Arg, 602 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && 603 Pred == ICmpInst::ICMP_EQ && 604 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 605 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 606 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 607 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 608 computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I)); 609 610 // For those bits in B that are known to be zero, we can propagate known 611 // bits from the RHS to V. 612 KnownZero |= RHSKnownZero & BKnownZero; 613 KnownOne |= RHSKnownOne & BKnownZero; 614 // assume(~(v | b) = a) 615 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), 616 m_Value(A))) && 617 Pred == ICmpInst::ICMP_EQ && 618 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 619 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 620 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 621 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 622 computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I)); 623 624 // For those bits in B that are known to be zero, we can propagate 625 // inverted known bits from the RHS to V. 626 KnownZero |= RHSKnownOne & BKnownZero; 627 KnownOne |= RHSKnownZero & BKnownZero; 628 // assume(v ^ b = a) 629 } else if (match(Arg, 630 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && 631 Pred == ICmpInst::ICMP_EQ && 632 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 633 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 634 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 635 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 636 computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I)); 637 638 // For those bits in B that are known to be zero, we can propagate known 639 // bits from the RHS to V. For those bits in B that are known to be one, 640 // we can propagate inverted known bits from the RHS to V. 641 KnownZero |= RHSKnownZero & BKnownZero; 642 KnownOne |= RHSKnownOne & BKnownZero; 643 KnownZero |= RHSKnownOne & BKnownOne; 644 KnownOne |= RHSKnownZero & BKnownOne; 645 // assume(~(v ^ b) = a) 646 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), 647 m_Value(A))) && 648 Pred == ICmpInst::ICMP_EQ && 649 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 650 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 651 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 652 APInt BKnownZero(BitWidth, 0), BKnownOne(BitWidth, 0); 653 computeKnownBits(B, BKnownZero, BKnownOne, Depth+1, Query(Q, I)); 654 655 // For those bits in B that are known to be zero, we can propagate 656 // inverted known bits from the RHS to V. For those bits in B that are 657 // known to be one, we can propagate known bits from the RHS to V. 658 KnownZero |= RHSKnownOne & BKnownZero; 659 KnownOne |= RHSKnownZero & BKnownZero; 660 KnownZero |= RHSKnownZero & BKnownOne; 661 KnownOne |= RHSKnownOne & BKnownOne; 662 // assume(v << c = a) 663 } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), 664 m_Value(A))) && 665 Pred == ICmpInst::ICMP_EQ && 666 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 667 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 668 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 669 // For those bits in RHS that are known, we can propagate them to known 670 // bits in V shifted to the right by C. 671 KnownZero |= RHSKnownZero.lshr(C->getZExtValue()); 672 KnownOne |= RHSKnownOne.lshr(C->getZExtValue()); 673 // assume(~(v << c) = a) 674 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), 675 m_Value(A))) && 676 Pred == ICmpInst::ICMP_EQ && 677 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 678 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 679 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 680 // For those bits in RHS that are known, we can propagate them inverted 681 // to known bits in V shifted to the right by C. 682 KnownZero |= RHSKnownOne.lshr(C->getZExtValue()); 683 KnownOne |= RHSKnownZero.lshr(C->getZExtValue()); 684 // assume(v >> c = a) 685 } else if (match(Arg, 686 m_c_ICmp(Pred, m_CombineOr(m_LShr(m_V, m_ConstantInt(C)), 687 m_AShr(m_V, m_ConstantInt(C))), 688 m_Value(A))) && 689 Pred == ICmpInst::ICMP_EQ && 690 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 691 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 692 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 693 // For those bits in RHS that are known, we can propagate them to known 694 // bits in V shifted to the right by C. 695 KnownZero |= RHSKnownZero << C->getZExtValue(); 696 KnownOne |= RHSKnownOne << C->getZExtValue(); 697 // assume(~(v >> c) = a) 698 } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_CombineOr( 699 m_LShr(m_V, m_ConstantInt(C)), 700 m_AShr(m_V, m_ConstantInt(C)))), 701 m_Value(A))) && 702 Pred == ICmpInst::ICMP_EQ && 703 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 704 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 705 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 706 // For those bits in RHS that are known, we can propagate them inverted 707 // to known bits in V shifted to the right by C. 708 KnownZero |= RHSKnownOne << C->getZExtValue(); 709 KnownOne |= RHSKnownZero << C->getZExtValue(); 710 // assume(v >=_s c) where c is non-negative 711 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 712 Pred == ICmpInst::ICMP_SGE && 713 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 714 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 715 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 716 717 if (RHSKnownZero.isNegative()) { 718 // We know that the sign bit is zero. 719 KnownZero |= APInt::getSignBit(BitWidth); 720 } 721 // assume(v >_s c) where c is at least -1. 722 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 723 Pred == ICmpInst::ICMP_SGT && 724 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 725 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 726 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 727 728 if (RHSKnownOne.isAllOnesValue() || RHSKnownZero.isNegative()) { 729 // We know that the sign bit is zero. 730 KnownZero |= APInt::getSignBit(BitWidth); 731 } 732 // assume(v <=_s c) where c is negative 733 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 734 Pred == ICmpInst::ICMP_SLE && 735 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 736 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 737 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 738 739 if (RHSKnownOne.isNegative()) { 740 // We know that the sign bit is one. 741 KnownOne |= APInt::getSignBit(BitWidth); 742 } 743 // assume(v <_s c) where c is non-positive 744 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 745 Pred == ICmpInst::ICMP_SLT && 746 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 747 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 748 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 749 750 if (RHSKnownZero.isAllOnesValue() || RHSKnownOne.isNegative()) { 751 // We know that the sign bit is one. 752 KnownOne |= APInt::getSignBit(BitWidth); 753 } 754 // assume(v <=_u c) 755 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 756 Pred == ICmpInst::ICMP_ULE && 757 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 758 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 759 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 760 761 // Whatever high bits in c are zero are known to be zero. 762 KnownZero |= 763 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()); 764 // assume(v <_u c) 765 } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) && 766 Pred == ICmpInst::ICMP_ULT && 767 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 768 APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0); 769 computeKnownBits(A, RHSKnownZero, RHSKnownOne, Depth+1, Query(Q, I)); 770 771 // Whatever high bits in c are zero are known to be zero (if c is a power 772 // of 2, then one more). 773 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I))) 774 KnownZero |= 775 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()+1); 776 else 777 KnownZero |= 778 APInt::getHighBitsSet(BitWidth, RHSKnownZero.countLeadingOnes()); 779 } 780 } 781 } 782 783 // Compute known bits from a shift operator, including those with a 784 // non-constant shift amount. KnownZero and KnownOne are the outputs of this 785 // function. KnownZero2 and KnownOne2 are pre-allocated temporaries with the 786 // same bit width as KnownZero and KnownOne. KZF and KOF are operator-specific 787 // functors that, given the known-zero or known-one bits respectively, and a 788 // shift amount, compute the implied known-zero or known-one bits of the shift 789 // operator's result respectively for that shift amount. The results from calling 790 // KZF and KOF are conservatively combined for all permitted shift amounts. 791 template <typename KZFunctor, typename KOFunctor> 792 static void computeKnownBitsFromShiftOperator(Operator *I, 793 APInt &KnownZero, APInt &KnownOne, 794 APInt &KnownZero2, APInt &KnownOne2, 795 unsigned Depth, const Query &Q, KZFunctor KZF, KOFunctor KOF) { 796 unsigned BitWidth = KnownZero.getBitWidth(); 797 798 if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { 799 unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1); 800 801 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 802 KnownZero = KZF(KnownZero, ShiftAmt); 803 KnownOne = KOF(KnownOne, ShiftAmt); 804 return; 805 } 806 807 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q); 808 809 // Note: We cannot use KnownZero.getLimitedValue() here, because if 810 // BitWidth > 64 and any upper bits are known, we'll end up returning the 811 // limit value (which implies all bits are known). 812 uint64_t ShiftAmtKZ = KnownZero.zextOrTrunc(64).getZExtValue(); 813 uint64_t ShiftAmtKO = KnownOne.zextOrTrunc(64).getZExtValue(); 814 815 // It would be more-clearly correct to use the two temporaries for this 816 // calculation. Reusing the APInts here to prevent unnecessary allocations. 817 KnownZero.clearAllBits(); 818 KnownOne.clearAllBits(); 819 820 // If we know the shifter operand is nonzero, we can sometimes infer more 821 // known bits. However this is expensive to compute, so be lazy about it and 822 // only compute it when absolutely necessary. 823 Optional<bool> ShifterOperandIsNonZero; 824 825 // Early exit if we can't constrain any well-defined shift amount. 826 if (!(ShiftAmtKZ & (BitWidth - 1)) && !(ShiftAmtKO & (BitWidth - 1))) { 827 ShifterOperandIsNonZero = 828 isKnownNonZero(I->getOperand(1), Depth + 1, Q); 829 if (!*ShifterOperandIsNonZero) 830 return; 831 } 832 833 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 834 835 KnownZero = KnownOne = APInt::getAllOnesValue(BitWidth); 836 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { 837 // Combine the shifted known input bits only for those shift amounts 838 // compatible with its known constraints. 839 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) 840 continue; 841 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) 842 continue; 843 // If we know the shifter is nonzero, we may be able to infer more known 844 // bits. This check is sunk down as far as possible to avoid the expensive 845 // call to isKnownNonZero if the cheaper checks above fail. 846 if (ShiftAmt == 0) { 847 if (!ShifterOperandIsNonZero.hasValue()) 848 ShifterOperandIsNonZero = 849 isKnownNonZero(I->getOperand(1), Depth + 1, Q); 850 if (*ShifterOperandIsNonZero) 851 continue; 852 } 853 854 KnownZero &= KZF(KnownZero2, ShiftAmt); 855 KnownOne &= KOF(KnownOne2, ShiftAmt); 856 } 857 858 // If there are no compatible shift amounts, then we've proven that the shift 859 // amount must be >= the BitWidth, and the result is undefined. We could 860 // return anything we'd like, but we need to make sure the sets of known bits 861 // stay disjoint (it should be better for some other code to actually 862 // propagate the undef than to pick a value here using known bits). 863 if ((KnownZero & KnownOne) != 0) { 864 KnownZero.clearAllBits(); 865 KnownOne.clearAllBits(); 866 } 867 } 868 869 static void computeKnownBitsFromOperator(Operator *I, APInt &KnownZero, 870 APInt &KnownOne, unsigned Depth, 871 const Query &Q) { 872 unsigned BitWidth = KnownZero.getBitWidth(); 873 874 APInt KnownZero2(KnownZero), KnownOne2(KnownOne); 875 switch (I->getOpcode()) { 876 default: break; 877 case Instruction::Load: 878 if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range)) 879 computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne); 880 break; 881 case Instruction::And: { 882 // If either the LHS or the RHS are Zero, the result is zero. 883 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q); 884 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 885 886 // Output known-1 bits are only known if set in both the LHS & RHS. 887 KnownOne &= KnownOne2; 888 // Output known-0 are known to be clear if zero in either the LHS | RHS. 889 KnownZero |= KnownZero2; 890 891 // and(x, add (x, -1)) is a common idiom that always clears the low bit; 892 // here we handle the more general case of adding any odd number by 893 // matching the form add(x, add(x, y)) where y is odd. 894 // TODO: This could be generalized to clearing any bit set in y where the 895 // following bit is known to be unset in y. 896 Value *Y = nullptr; 897 if (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)), 898 m_Value(Y))) || 899 match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)), 900 m_Value(Y)))) { 901 APInt KnownZero3(BitWidth, 0), KnownOne3(BitWidth, 0); 902 computeKnownBits(Y, KnownZero3, KnownOne3, Depth + 1, Q); 903 if (KnownOne3.countTrailingOnes() > 0) 904 KnownZero |= APInt::getLowBitsSet(BitWidth, 1); 905 } 906 break; 907 } 908 case Instruction::Or: { 909 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q); 910 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 911 912 // Output known-0 bits are only known if clear in both the LHS & RHS. 913 KnownZero &= KnownZero2; 914 // Output known-1 are known to be set if set in either the LHS | RHS. 915 KnownOne |= KnownOne2; 916 break; 917 } 918 case Instruction::Xor: { 919 computeKnownBits(I->getOperand(1), KnownZero, KnownOne, Depth + 1, Q); 920 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 921 922 // Output known-0 bits are known if clear or set in both the LHS & RHS. 923 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 924 // Output known-1 are known to be set if set in only one of the LHS, RHS. 925 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 926 KnownZero = KnownZeroOut; 927 break; 928 } 929 case Instruction::Mul: { 930 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 931 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, KnownZero, 932 KnownOne, KnownZero2, KnownOne2, Depth, Q); 933 break; 934 } 935 case Instruction::UDiv: { 936 // For the purposes of computing leading zeros we can conservatively 937 // treat a udiv as a logical right shift by the power of 2 known to 938 // be less than the denominator. 939 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 940 unsigned LeadZ = KnownZero2.countLeadingOnes(); 941 942 KnownOne2.clearAllBits(); 943 KnownZero2.clearAllBits(); 944 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q); 945 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros(); 946 if (RHSUnknownLeadingOnes != BitWidth) 947 LeadZ = std::min(BitWidth, 948 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1); 949 950 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ); 951 break; 952 } 953 case Instruction::Select: 954 computeKnownBits(I->getOperand(2), KnownZero, KnownOne, Depth + 1, Q); 955 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q); 956 957 // Only known if known in both the LHS and RHS. 958 KnownOne &= KnownOne2; 959 KnownZero &= KnownZero2; 960 break; 961 case Instruction::FPTrunc: 962 case Instruction::FPExt: 963 case Instruction::FPToUI: 964 case Instruction::FPToSI: 965 case Instruction::SIToFP: 966 case Instruction::UIToFP: 967 break; // Can't work with floating point. 968 case Instruction::PtrToInt: 969 case Instruction::IntToPtr: 970 case Instruction::AddrSpaceCast: // Pointers could be different sizes. 971 // FALL THROUGH and handle them the same as zext/trunc. 972 case Instruction::ZExt: 973 case Instruction::Trunc: { 974 Type *SrcTy = I->getOperand(0)->getType(); 975 976 unsigned SrcBitWidth; 977 // Note that we handle pointer operands here because of inttoptr/ptrtoint 978 // which fall through here. 979 SrcBitWidth = Q.DL.getTypeSizeInBits(SrcTy->getScalarType()); 980 981 assert(SrcBitWidth && "SrcBitWidth can't be zero"); 982 KnownZero = KnownZero.zextOrTrunc(SrcBitWidth); 983 KnownOne = KnownOne.zextOrTrunc(SrcBitWidth); 984 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 985 KnownZero = KnownZero.zextOrTrunc(BitWidth); 986 KnownOne = KnownOne.zextOrTrunc(BitWidth); 987 // Any top bits are known to be zero. 988 if (BitWidth > SrcBitWidth) 989 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 990 break; 991 } 992 case Instruction::BitCast: { 993 Type *SrcTy = I->getOperand(0)->getType(); 994 if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 995 // TODO: For now, not handling conversions like: 996 // (bitcast i64 %x to <2 x i32>) 997 !I->getType()->isVectorTy()) { 998 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 999 break; 1000 } 1001 break; 1002 } 1003 case Instruction::SExt: { 1004 // Compute the bits in the result that are not present in the input. 1005 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 1006 1007 KnownZero = KnownZero.trunc(SrcBitWidth); 1008 KnownOne = KnownOne.trunc(SrcBitWidth); 1009 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 1010 KnownZero = KnownZero.zext(BitWidth); 1011 KnownOne = KnownOne.zext(BitWidth); 1012 1013 // If the sign bit of the input is known set or clear, then we know the 1014 // top bits of the result. 1015 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero 1016 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 1017 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set 1018 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth); 1019 break; 1020 } 1021 case Instruction::Shl: { 1022 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0 1023 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) { 1024 return (KnownZero << ShiftAmt) | 1025 APInt::getLowBitsSet(BitWidth, ShiftAmt); // Low bits known 0. 1026 }; 1027 1028 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) { 1029 return KnownOne << ShiftAmt; 1030 }; 1031 1032 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne, 1033 KnownZero2, KnownOne2, Depth, Q, KZF, 1034 KOF); 1035 break; 1036 } 1037 case Instruction::LShr: { 1038 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1039 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) { 1040 return APIntOps::lshr(KnownZero, ShiftAmt) | 1041 // High bits known zero. 1042 APInt::getHighBitsSet(BitWidth, ShiftAmt); 1043 }; 1044 1045 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) { 1046 return APIntOps::lshr(KnownOne, ShiftAmt); 1047 }; 1048 1049 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne, 1050 KnownZero2, KnownOne2, Depth, Q, KZF, 1051 KOF); 1052 break; 1053 } 1054 case Instruction::AShr: { 1055 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1056 auto KZF = [BitWidth](const APInt &KnownZero, unsigned ShiftAmt) { 1057 return APIntOps::ashr(KnownZero, ShiftAmt); 1058 }; 1059 1060 auto KOF = [BitWidth](const APInt &KnownOne, unsigned ShiftAmt) { 1061 return APIntOps::ashr(KnownOne, ShiftAmt); 1062 }; 1063 1064 computeKnownBitsFromShiftOperator(I, KnownZero, KnownOne, 1065 KnownZero2, KnownOne2, Depth, Q, KZF, 1066 KOF); 1067 break; 1068 } 1069 case Instruction::Sub: { 1070 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 1071 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, 1072 KnownZero, KnownOne, KnownZero2, KnownOne2, Depth, 1073 Q); 1074 break; 1075 } 1076 case Instruction::Add: { 1077 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 1078 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, 1079 KnownZero, KnownOne, KnownZero2, KnownOne2, Depth, 1080 Q); 1081 break; 1082 } 1083 case Instruction::SRem: 1084 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) { 1085 APInt RA = Rem->getValue().abs(); 1086 if (RA.isPowerOf2()) { 1087 APInt LowBits = RA - 1; 1088 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, 1089 Q); 1090 1091 // The low bits of the first operand are unchanged by the srem. 1092 KnownZero = KnownZero2 & LowBits; 1093 KnownOne = KnownOne2 & LowBits; 1094 1095 // If the first operand is non-negative or has all low bits zero, then 1096 // the upper bits are all zero. 1097 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits)) 1098 KnownZero |= ~LowBits; 1099 1100 // If the first operand is negative and not all low bits are zero, then 1101 // the upper bits are all one. 1102 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0)) 1103 KnownOne |= ~LowBits; 1104 1105 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1106 } 1107 } 1108 1109 // The sign bit is the LHS's sign bit, except when the result of the 1110 // remainder is zero. 1111 if (KnownZero.isNonNegative()) { 1112 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0); 1113 computeKnownBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth + 1, 1114 Q); 1115 // If it's known zero, our sign bit is also zero. 1116 if (LHSKnownZero.isNegative()) 1117 KnownZero.setBit(BitWidth - 1); 1118 } 1119 1120 break; 1121 case Instruction::URem: { 1122 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) { 1123 const APInt &RA = Rem->getValue(); 1124 if (RA.isPowerOf2()) { 1125 APInt LowBits = (RA - 1); 1126 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 1127 KnownZero |= ~LowBits; 1128 KnownOne &= LowBits; 1129 break; 1130 } 1131 } 1132 1133 // Since the result is less than or equal to either operand, any leading 1134 // zero bits in either operand must also exist in the result. 1135 computeKnownBits(I->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 1136 computeKnownBits(I->getOperand(1), KnownZero2, KnownOne2, Depth + 1, Q); 1137 1138 unsigned Leaders = std::max(KnownZero.countLeadingOnes(), 1139 KnownZero2.countLeadingOnes()); 1140 KnownOne.clearAllBits(); 1141 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders); 1142 break; 1143 } 1144 1145 case Instruction::Alloca: { 1146 AllocaInst *AI = cast<AllocaInst>(I); 1147 unsigned Align = AI->getAlignment(); 1148 if (Align == 0) 1149 Align = Q.DL.getABITypeAlignment(AI->getAllocatedType()); 1150 1151 if (Align > 0) 1152 KnownZero = APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align)); 1153 break; 1154 } 1155 case Instruction::GetElementPtr: { 1156 // Analyze all of the subscripts of this getelementptr instruction 1157 // to determine if we can prove known low zero bits. 1158 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0); 1159 computeKnownBits(I->getOperand(0), LocalKnownZero, LocalKnownOne, Depth + 1, 1160 Q); 1161 unsigned TrailZ = LocalKnownZero.countTrailingOnes(); 1162 1163 gep_type_iterator GTI = gep_type_begin(I); 1164 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { 1165 Value *Index = I->getOperand(i); 1166 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 1167 // Handle struct member offset arithmetic. 1168 1169 // Handle case when index is vector zeroinitializer 1170 Constant *CIndex = cast<Constant>(Index); 1171 if (CIndex->isZeroValue()) 1172 continue; 1173 1174 if (CIndex->getType()->isVectorTy()) 1175 Index = CIndex->getSplatValue(); 1176 1177 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); 1178 const StructLayout *SL = Q.DL.getStructLayout(STy); 1179 uint64_t Offset = SL->getElementOffset(Idx); 1180 TrailZ = std::min<unsigned>(TrailZ, 1181 countTrailingZeros(Offset)); 1182 } else { 1183 // Handle array index arithmetic. 1184 Type *IndexedTy = GTI.getIndexedType(); 1185 if (!IndexedTy->isSized()) { 1186 TrailZ = 0; 1187 break; 1188 } 1189 unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits(); 1190 uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy); 1191 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0); 1192 computeKnownBits(Index, LocalKnownZero, LocalKnownOne, Depth + 1, Q); 1193 TrailZ = std::min(TrailZ, 1194 unsigned(countTrailingZeros(TypeSize) + 1195 LocalKnownZero.countTrailingOnes())); 1196 } 1197 } 1198 1199 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ); 1200 break; 1201 } 1202 case Instruction::PHI: { 1203 PHINode *P = cast<PHINode>(I); 1204 // Handle the case of a simple two-predecessor recurrence PHI. 1205 // There's a lot more that could theoretically be done here, but 1206 // this is sufficient to catch some interesting cases. 1207 if (P->getNumIncomingValues() == 2) { 1208 for (unsigned i = 0; i != 2; ++i) { 1209 Value *L = P->getIncomingValue(i); 1210 Value *R = P->getIncomingValue(!i); 1211 Operator *LU = dyn_cast<Operator>(L); 1212 if (!LU) 1213 continue; 1214 unsigned Opcode = LU->getOpcode(); 1215 // Check for operations that have the property that if 1216 // both their operands have low zero bits, the result 1217 // will have low zero bits. 1218 if (Opcode == Instruction::Add || 1219 Opcode == Instruction::Sub || 1220 Opcode == Instruction::And || 1221 Opcode == Instruction::Or || 1222 Opcode == Instruction::Mul) { 1223 Value *LL = LU->getOperand(0); 1224 Value *LR = LU->getOperand(1); 1225 // Find a recurrence. 1226 if (LL == I) 1227 L = LR; 1228 else if (LR == I) 1229 L = LL; 1230 else 1231 break; 1232 // Ok, we have a PHI of the form L op= R. Check for low 1233 // zero bits. 1234 computeKnownBits(R, KnownZero2, KnownOne2, Depth + 1, Q); 1235 1236 // We need to take the minimum number of known bits 1237 APInt KnownZero3(KnownZero), KnownOne3(KnownOne); 1238 computeKnownBits(L, KnownZero3, KnownOne3, Depth + 1, Q); 1239 1240 KnownZero = APInt::getLowBitsSet(BitWidth, 1241 std::min(KnownZero2.countTrailingOnes(), 1242 KnownZero3.countTrailingOnes())); 1243 break; 1244 } 1245 } 1246 } 1247 1248 // Unreachable blocks may have zero-operand PHI nodes. 1249 if (P->getNumIncomingValues() == 0) 1250 break; 1251 1252 // Otherwise take the unions of the known bit sets of the operands, 1253 // taking conservative care to avoid excessive recursion. 1254 if (Depth < MaxDepth - 1 && !KnownZero && !KnownOne) { 1255 // Skip if every incoming value references to ourself. 1256 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue())) 1257 break; 1258 1259 KnownZero = APInt::getAllOnesValue(BitWidth); 1260 KnownOne = APInt::getAllOnesValue(BitWidth); 1261 for (Value *IncValue : P->incoming_values()) { 1262 // Skip direct self references. 1263 if (IncValue == P) continue; 1264 1265 KnownZero2 = APInt(BitWidth, 0); 1266 KnownOne2 = APInt(BitWidth, 0); 1267 // Recurse, but cap the recursion to one level, because we don't 1268 // want to waste time spinning around in loops. 1269 computeKnownBits(IncValue, KnownZero2, KnownOne2, MaxDepth - 1, Q); 1270 KnownZero &= KnownZero2; 1271 KnownOne &= KnownOne2; 1272 // If all bits have been ruled out, there's no need to check 1273 // more operands. 1274 if (!KnownZero && !KnownOne) 1275 break; 1276 } 1277 } 1278 break; 1279 } 1280 case Instruction::Call: 1281 case Instruction::Invoke: 1282 // If range metadata is attached to this call, set known bits from that, 1283 // and then intersect with known bits based on other properties of the 1284 // function. 1285 if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range)) 1286 computeKnownBitsFromRangeMetadata(*MD, KnownZero, KnownOne); 1287 if (Value *RV = CallSite(I).getReturnedArgOperand()) { 1288 computeKnownBits(RV, KnownZero2, KnownOne2, Depth + 1, Q); 1289 KnownZero |= KnownZero2; 1290 KnownOne |= KnownOne2; 1291 } 1292 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1293 switch (II->getIntrinsicID()) { 1294 default: break; 1295 case Intrinsic::bswap: 1296 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 1297 KnownZero |= KnownZero2.byteSwap(); 1298 KnownOne |= KnownOne2.byteSwap(); 1299 break; 1300 case Intrinsic::ctlz: 1301 case Intrinsic::cttz: { 1302 unsigned LowBits = Log2_32(BitWidth)+1; 1303 // If this call is undefined for 0, the result will be less than 2^n. 1304 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1305 LowBits -= 1; 1306 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1307 break; 1308 } 1309 case Intrinsic::ctpop: { 1310 computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, Depth + 1, Q); 1311 // We can bound the space the count needs. Also, bits known to be zero 1312 // can't contribute to the population. 1313 unsigned BitsPossiblySet = BitWidth - KnownZero2.countPopulation(); 1314 unsigned LeadingZeros = 1315 APInt(BitWidth, BitsPossiblySet).countLeadingZeros(); 1316 assert(LeadingZeros <= BitWidth); 1317 KnownZero |= APInt::getHighBitsSet(BitWidth, LeadingZeros); 1318 KnownOne &= ~KnownZero; 1319 // TODO: we could bound KnownOne using the lower bound on the number 1320 // of bits which might be set provided by popcnt KnownOne2. 1321 break; 1322 } 1323 case Intrinsic::x86_sse42_crc32_64_64: 1324 KnownZero |= APInt::getHighBitsSet(64, 32); 1325 break; 1326 } 1327 } 1328 break; 1329 case Instruction::ExtractValue: 1330 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { 1331 ExtractValueInst *EVI = cast<ExtractValueInst>(I); 1332 if (EVI->getNumIndices() != 1) break; 1333 if (EVI->getIndices()[0] == 0) { 1334 switch (II->getIntrinsicID()) { 1335 default: break; 1336 case Intrinsic::uadd_with_overflow: 1337 case Intrinsic::sadd_with_overflow: 1338 computeKnownBitsAddSub(true, II->getArgOperand(0), 1339 II->getArgOperand(1), false, KnownZero, 1340 KnownOne, KnownZero2, KnownOne2, Depth, Q); 1341 break; 1342 case Intrinsic::usub_with_overflow: 1343 case Intrinsic::ssub_with_overflow: 1344 computeKnownBitsAddSub(false, II->getArgOperand(0), 1345 II->getArgOperand(1), false, KnownZero, 1346 KnownOne, KnownZero2, KnownOne2, Depth, Q); 1347 break; 1348 case Intrinsic::umul_with_overflow: 1349 case Intrinsic::smul_with_overflow: 1350 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, 1351 KnownZero, KnownOne, KnownZero2, KnownOne2, Depth, 1352 Q); 1353 break; 1354 } 1355 } 1356 } 1357 } 1358 } 1359 1360 /// Determine which bits of V are known to be either zero or one and return 1361 /// them in the KnownZero/KnownOne bit sets. 1362 /// 1363 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that 1364 /// we cannot optimize based on the assumption that it is zero without changing 1365 /// it to be an explicit zero. If we don't change it to zero, other code could 1366 /// optimized based on the contradictory assumption that it is non-zero. 1367 /// Because instcombine aggressively folds operations with undef args anyway, 1368 /// this won't lose us code quality. 1369 /// 1370 /// This function is defined on values with integer type, values with pointer 1371 /// type, and vectors of integers. In the case 1372 /// where V is a vector, known zero, and known one values are the 1373 /// same width as the vector element, and the bit is set only if it is true 1374 /// for all of the elements in the vector. 1375 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 1376 unsigned Depth, const Query &Q) { 1377 assert(V && "No Value?"); 1378 assert(Depth <= MaxDepth && "Limit Search Depth"); 1379 unsigned BitWidth = KnownZero.getBitWidth(); 1380 1381 assert((V->getType()->isIntOrIntVectorTy() || 1382 V->getType()->getScalarType()->isPointerTy()) && 1383 "Not integer or pointer type!"); 1384 assert((Q.DL.getTypeSizeInBits(V->getType()->getScalarType()) == BitWidth) && 1385 (!V->getType()->isIntOrIntVectorTy() || 1386 V->getType()->getScalarSizeInBits() == BitWidth) && 1387 KnownZero.getBitWidth() == BitWidth && 1388 KnownOne.getBitWidth() == BitWidth && 1389 "V, KnownOne and KnownZero should have same BitWidth"); 1390 1391 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 1392 // We know all of the bits for a constant! 1393 KnownOne = CI->getValue(); 1394 KnownZero = ~KnownOne; 1395 return; 1396 } 1397 // Null and aggregate-zero are all-zeros. 1398 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { 1399 KnownOne.clearAllBits(); 1400 KnownZero = APInt::getAllOnesValue(BitWidth); 1401 return; 1402 } 1403 // Handle a constant vector by taking the intersection of the known bits of 1404 // each element. 1405 if (ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) { 1406 // We know that CDS must be a vector of integers. Take the intersection of 1407 // each element. 1408 KnownZero.setAllBits(); KnownOne.setAllBits(); 1409 APInt Elt(KnownZero.getBitWidth(), 0); 1410 for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { 1411 Elt = CDS->getElementAsInteger(i); 1412 KnownZero &= ~Elt; 1413 KnownOne &= Elt; 1414 } 1415 return; 1416 } 1417 1418 if (auto *CV = dyn_cast<ConstantVector>(V)) { 1419 // We know that CV must be a vector of integers. Take the intersection of 1420 // each element. 1421 KnownZero.setAllBits(); KnownOne.setAllBits(); 1422 APInt Elt(KnownZero.getBitWidth(), 0); 1423 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { 1424 Constant *Element = CV->getAggregateElement(i); 1425 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); 1426 if (!ElementCI) { 1427 KnownZero.clearAllBits(); 1428 KnownOne.clearAllBits(); 1429 return; 1430 } 1431 Elt = ElementCI->getValue(); 1432 KnownZero &= ~Elt; 1433 KnownOne &= Elt; 1434 } 1435 return; 1436 } 1437 1438 // Start out not knowing anything. 1439 KnownZero.clearAllBits(); KnownOne.clearAllBits(); 1440 1441 // Limit search depth. 1442 // All recursive calls that increase depth must come after this. 1443 if (Depth == MaxDepth) 1444 return; 1445 1446 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has 1447 // the bits of its aliasee. 1448 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 1449 if (!GA->isInterposable()) 1450 computeKnownBits(GA->getAliasee(), KnownZero, KnownOne, Depth + 1, Q); 1451 return; 1452 } 1453 1454 if (Operator *I = dyn_cast<Operator>(V)) 1455 computeKnownBitsFromOperator(I, KnownZero, KnownOne, Depth, Q); 1456 1457 // Aligned pointers have trailing zeros - refine KnownZero set 1458 if (V->getType()->isPointerTy()) { 1459 unsigned Align = V->getPointerAlignment(Q.DL); 1460 if (Align) 1461 KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align)); 1462 } 1463 1464 // computeKnownBitsFromAssume strictly refines KnownZero and 1465 // KnownOne. Therefore, we run them after computeKnownBitsFromOperator. 1466 1467 // Check whether a nearby assume intrinsic can determine some known bits. 1468 computeKnownBitsFromAssume(V, KnownZero, KnownOne, Depth, Q); 1469 1470 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1471 } 1472 1473 /// Determine whether the sign bit is known to be zero or one. 1474 /// Convenience wrapper around computeKnownBits. 1475 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 1476 unsigned Depth, const Query &Q) { 1477 unsigned BitWidth = getBitWidth(V->getType(), Q.DL); 1478 if (!BitWidth) { 1479 KnownZero = false; 1480 KnownOne = false; 1481 return; 1482 } 1483 APInt ZeroBits(BitWidth, 0); 1484 APInt OneBits(BitWidth, 0); 1485 computeKnownBits(V, ZeroBits, OneBits, Depth, Q); 1486 KnownOne = OneBits[BitWidth - 1]; 1487 KnownZero = ZeroBits[BitWidth - 1]; 1488 } 1489 1490 /// Return true if the given value is known to have exactly one 1491 /// bit set when defined. For vectors return true if every element is known to 1492 /// be a power of two when defined. Supports values with integer or pointer 1493 /// types and vectors of integers. 1494 bool isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth, 1495 const Query &Q) { 1496 if (Constant *C = dyn_cast<Constant>(V)) { 1497 if (C->isNullValue()) 1498 return OrZero; 1499 1500 const APInt *ConstIntOrConstSplatInt; 1501 if (match(C, m_APInt(ConstIntOrConstSplatInt))) 1502 return ConstIntOrConstSplatInt->isPowerOf2(); 1503 } 1504 1505 // 1 << X is clearly a power of two if the one is not shifted off the end. If 1506 // it is shifted off the end then the result is undefined. 1507 if (match(V, m_Shl(m_One(), m_Value()))) 1508 return true; 1509 1510 // (signbit) >>l X is clearly a power of two if the one is not shifted off the 1511 // bottom. If it is shifted off the bottom then the result is undefined. 1512 if (match(V, m_LShr(m_SignBit(), m_Value()))) 1513 return true; 1514 1515 // The remaining tests are all recursive, so bail out if we hit the limit. 1516 if (Depth++ == MaxDepth) 1517 return false; 1518 1519 Value *X = nullptr, *Y = nullptr; 1520 // A shift left or a logical shift right of a power of two is a power of two 1521 // or zero. 1522 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || 1523 match(V, m_LShr(m_Value(X), m_Value())))) 1524 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); 1525 1526 if (ZExtInst *ZI = dyn_cast<ZExtInst>(V)) 1527 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); 1528 1529 if (SelectInst *SI = dyn_cast<SelectInst>(V)) 1530 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && 1531 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); 1532 1533 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { 1534 // A power of two and'd with anything is a power of two or zero. 1535 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || 1536 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) 1537 return true; 1538 // X & (-X) is always a power of two or zero. 1539 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) 1540 return true; 1541 return false; 1542 } 1543 1544 // Adding a power-of-two or zero to the same power-of-two or zero yields 1545 // either the original power-of-two, a larger power-of-two or zero. 1546 if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 1547 OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); 1548 if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) { 1549 if (match(X, m_And(m_Specific(Y), m_Value())) || 1550 match(X, m_And(m_Value(), m_Specific(Y)))) 1551 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) 1552 return true; 1553 if (match(Y, m_And(m_Specific(X), m_Value())) || 1554 match(Y, m_And(m_Value(), m_Specific(X)))) 1555 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) 1556 return true; 1557 1558 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 1559 APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0); 1560 computeKnownBits(X, LHSZeroBits, LHSOneBits, Depth, Q); 1561 1562 APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0); 1563 computeKnownBits(Y, RHSZeroBits, RHSOneBits, Depth, Q); 1564 // If i8 V is a power of two or zero: 1565 // ZeroBits: 1 1 1 0 1 1 1 1 1566 // ~ZeroBits: 0 0 0 1 0 0 0 0 1567 if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2()) 1568 // If OrZero isn't set, we cannot give back a zero result. 1569 // Make sure either the LHS or RHS has a bit set. 1570 if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue()) 1571 return true; 1572 } 1573 } 1574 1575 // An exact divide or right shift can only shift off zero bits, so the result 1576 // is a power of two only if the first operand is a power of two and not 1577 // copying a sign bit (sdiv int_min, 2). 1578 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || 1579 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { 1580 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, 1581 Depth, Q); 1582 } 1583 1584 return false; 1585 } 1586 1587 /// \brief Test whether a GEP's result is known to be non-null. 1588 /// 1589 /// Uses properties inherent in a GEP to try to determine whether it is known 1590 /// to be non-null. 1591 /// 1592 /// Currently this routine does not support vector GEPs. 1593 static bool isGEPKnownNonNull(GEPOperator *GEP, unsigned Depth, 1594 const Query &Q) { 1595 if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0) 1596 return false; 1597 1598 // FIXME: Support vector-GEPs. 1599 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP"); 1600 1601 // If the base pointer is non-null, we cannot walk to a null address with an 1602 // inbounds GEP in address space zero. 1603 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) 1604 return true; 1605 1606 // Walk the GEP operands and see if any operand introduces a non-zero offset. 1607 // If so, then the GEP cannot produce a null pointer, as doing so would 1608 // inherently violate the inbounds contract within address space zero. 1609 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 1610 GTI != GTE; ++GTI) { 1611 // Struct types are easy -- they must always be indexed by a constant. 1612 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 1613 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); 1614 unsigned ElementIdx = OpC->getZExtValue(); 1615 const StructLayout *SL = Q.DL.getStructLayout(STy); 1616 uint64_t ElementOffset = SL->getElementOffset(ElementIdx); 1617 if (ElementOffset > 0) 1618 return true; 1619 continue; 1620 } 1621 1622 // If we have a zero-sized type, the index doesn't matter. Keep looping. 1623 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0) 1624 continue; 1625 1626 // Fast path the constant operand case both for efficiency and so we don't 1627 // increment Depth when just zipping down an all-constant GEP. 1628 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { 1629 if (!OpC->isZero()) 1630 return true; 1631 continue; 1632 } 1633 1634 // We post-increment Depth here because while isKnownNonZero increments it 1635 // as well, when we pop back up that increment won't persist. We don't want 1636 // to recurse 10k times just because we have 10k GEP operands. We don't 1637 // bail completely out because we want to handle constant GEPs regardless 1638 // of depth. 1639 if (Depth++ >= MaxDepth) 1640 continue; 1641 1642 if (isKnownNonZero(GTI.getOperand(), Depth, Q)) 1643 return true; 1644 } 1645 1646 return false; 1647 } 1648 1649 /// Does the 'Range' metadata (which must be a valid MD_range operand list) 1650 /// ensure that the value it's attached to is never Value? 'RangeType' is 1651 /// is the type of the value described by the range. 1652 static bool rangeMetadataExcludesValue(MDNode* Ranges, const APInt& Value) { 1653 const unsigned NumRanges = Ranges->getNumOperands() / 2; 1654 assert(NumRanges >= 1); 1655 for (unsigned i = 0; i < NumRanges; ++i) { 1656 ConstantInt *Lower = 1657 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); 1658 ConstantInt *Upper = 1659 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); 1660 ConstantRange Range(Lower->getValue(), Upper->getValue()); 1661 if (Range.contains(Value)) 1662 return false; 1663 } 1664 return true; 1665 } 1666 1667 /// Return true if the given value is known to be non-zero when defined. 1668 /// For vectors return true if every element is known to be non-zero when 1669 /// defined. Supports values with integer or pointer type and vectors of 1670 /// integers. 1671 bool isKnownNonZero(Value *V, unsigned Depth, const Query &Q) { 1672 if (auto *C = dyn_cast<Constant>(V)) { 1673 if (C->isNullValue()) 1674 return false; 1675 if (isa<ConstantInt>(C)) 1676 // Must be non-zero due to null test above. 1677 return true; 1678 1679 // For constant vectors, check that all elements are undefined or known 1680 // non-zero to determine that the whole vector is known non-zero. 1681 if (auto *VecTy = dyn_cast<VectorType>(C->getType())) { 1682 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { 1683 Constant *Elt = C->getAggregateElement(i); 1684 if (!Elt || Elt->isNullValue()) 1685 return false; 1686 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) 1687 return false; 1688 } 1689 return true; 1690 } 1691 1692 return false; 1693 } 1694 1695 if (auto *I = dyn_cast<Instruction>(V)) { 1696 if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) { 1697 // If the possible ranges don't contain zero, then the value is 1698 // definitely non-zero. 1699 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { 1700 const APInt ZeroValue(Ty->getBitWidth(), 0); 1701 if (rangeMetadataExcludesValue(Ranges, ZeroValue)) 1702 return true; 1703 } 1704 } 1705 } 1706 1707 // The remaining tests are all recursive, so bail out if we hit the limit. 1708 if (Depth++ >= MaxDepth) 1709 return false; 1710 1711 // Check for pointer simplifications. 1712 if (V->getType()->isPointerTy()) { 1713 if (isKnownNonNull(V)) 1714 return true; 1715 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 1716 if (isGEPKnownNonNull(GEP, Depth, Q)) 1717 return true; 1718 } 1719 1720 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); 1721 1722 // X | Y != 0 if X != 0 or Y != 0. 1723 Value *X = nullptr, *Y = nullptr; 1724 if (match(V, m_Or(m_Value(X), m_Value(Y)))) 1725 return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q); 1726 1727 // ext X != 0 if X != 0. 1728 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) 1729 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q); 1730 1731 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined 1732 // if the lowest bit is shifted off the end. 1733 if (BitWidth && match(V, m_Shl(m_Value(X), m_Value(Y)))) { 1734 // shl nuw can't remove any non-zero bits. 1735 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 1736 if (BO->hasNoUnsignedWrap()) 1737 return isKnownNonZero(X, Depth, Q); 1738 1739 APInt KnownZero(BitWidth, 0); 1740 APInt KnownOne(BitWidth, 0); 1741 computeKnownBits(X, KnownZero, KnownOne, Depth, Q); 1742 if (KnownOne[0]) 1743 return true; 1744 } 1745 // shr X, Y != 0 if X is negative. Note that the value of the shift is not 1746 // defined if the sign bit is shifted off the end. 1747 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { 1748 // shr exact can only shift out zero bits. 1749 PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); 1750 if (BO->isExact()) 1751 return isKnownNonZero(X, Depth, Q); 1752 1753 bool XKnownNonNegative, XKnownNegative; 1754 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q); 1755 if (XKnownNegative) 1756 return true; 1757 1758 // If the shifter operand is a constant, and all of the bits shifted 1759 // out are known to be zero, and X is known non-zero then at least one 1760 // non-zero bit must remain. 1761 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) { 1762 APInt KnownZero(BitWidth, 0); 1763 APInt KnownOne(BitWidth, 0); 1764 computeKnownBits(X, KnownZero, KnownOne, Depth, Q); 1765 1766 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1); 1767 // Is there a known one in the portion not shifted out? 1768 if (KnownOne.countLeadingZeros() < BitWidth - ShiftVal) 1769 return true; 1770 // Are all the bits to be shifted out known zero? 1771 if (KnownZero.countTrailingOnes() >= ShiftVal) 1772 return isKnownNonZero(X, Depth, Q); 1773 } 1774 } 1775 // div exact can only produce a zero if the dividend is zero. 1776 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { 1777 return isKnownNonZero(X, Depth, Q); 1778 } 1779 // X + Y. 1780 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 1781 bool XKnownNonNegative, XKnownNegative; 1782 bool YKnownNonNegative, YKnownNegative; 1783 ComputeSignBit(X, XKnownNonNegative, XKnownNegative, Depth, Q); 1784 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Depth, Q); 1785 1786 // If X and Y are both non-negative (as signed values) then their sum is not 1787 // zero unless both X and Y are zero. 1788 if (XKnownNonNegative && YKnownNonNegative) 1789 if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q)) 1790 return true; 1791 1792 // If X and Y are both negative (as signed values) then their sum is not 1793 // zero unless both X and Y equal INT_MIN. 1794 if (BitWidth && XKnownNegative && YKnownNegative) { 1795 APInt KnownZero(BitWidth, 0); 1796 APInt KnownOne(BitWidth, 0); 1797 APInt Mask = APInt::getSignedMaxValue(BitWidth); 1798 // The sign bit of X is set. If some other bit is set then X is not equal 1799 // to INT_MIN. 1800 computeKnownBits(X, KnownZero, KnownOne, Depth, Q); 1801 if ((KnownOne & Mask) != 0) 1802 return true; 1803 // The sign bit of Y is set. If some other bit is set then Y is not equal 1804 // to INT_MIN. 1805 computeKnownBits(Y, KnownZero, KnownOne, Depth, Q); 1806 if ((KnownOne & Mask) != 0) 1807 return true; 1808 } 1809 1810 // The sum of a non-negative number and a power of two is not zero. 1811 if (XKnownNonNegative && 1812 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) 1813 return true; 1814 if (YKnownNonNegative && 1815 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) 1816 return true; 1817 } 1818 // X * Y. 1819 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { 1820 OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 1821 // If X and Y are non-zero then so is X * Y as long as the multiplication 1822 // does not overflow. 1823 if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) && 1824 isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q)) 1825 return true; 1826 } 1827 // (C ? X : Y) != 0 if X != 0 and Y != 0. 1828 else if (SelectInst *SI = dyn_cast<SelectInst>(V)) { 1829 if (isKnownNonZero(SI->getTrueValue(), Depth, Q) && 1830 isKnownNonZero(SI->getFalseValue(), Depth, Q)) 1831 return true; 1832 } 1833 // PHI 1834 else if (PHINode *PN = dyn_cast<PHINode>(V)) { 1835 // Try and detect a recurrence that monotonically increases from a 1836 // starting value, as these are common as induction variables. 1837 if (PN->getNumIncomingValues() == 2) { 1838 Value *Start = PN->getIncomingValue(0); 1839 Value *Induction = PN->getIncomingValue(1); 1840 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start)) 1841 std::swap(Start, Induction); 1842 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) { 1843 if (!C->isZero() && !C->isNegative()) { 1844 ConstantInt *X; 1845 if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) || 1846 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) && 1847 !X->isNegative()) 1848 return true; 1849 } 1850 } 1851 } 1852 // Check if all incoming values are non-zero constant. 1853 bool AllNonZeroConstants = all_of(PN->operands(), [](Value *V) { 1854 return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZeroValue(); 1855 }); 1856 if (AllNonZeroConstants) 1857 return true; 1858 } 1859 1860 if (!BitWidth) return false; 1861 APInt KnownZero(BitWidth, 0); 1862 APInt KnownOne(BitWidth, 0); 1863 computeKnownBits(V, KnownZero, KnownOne, Depth, Q); 1864 return KnownOne != 0; 1865 } 1866 1867 /// Return true if V2 == V1 + X, where X is known non-zero. 1868 static bool isAddOfNonZero(Value *V1, Value *V2, const Query &Q) { 1869 BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); 1870 if (!BO || BO->getOpcode() != Instruction::Add) 1871 return false; 1872 Value *Op = nullptr; 1873 if (V2 == BO->getOperand(0)) 1874 Op = BO->getOperand(1); 1875 else if (V2 == BO->getOperand(1)) 1876 Op = BO->getOperand(0); 1877 else 1878 return false; 1879 return isKnownNonZero(Op, 0, Q); 1880 } 1881 1882 /// Return true if it is known that V1 != V2. 1883 static bool isKnownNonEqual(Value *V1, Value *V2, const Query &Q) { 1884 if (V1->getType()->isVectorTy() || V1 == V2) 1885 return false; 1886 if (V1->getType() != V2->getType()) 1887 // We can't look through casts yet. 1888 return false; 1889 if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q)) 1890 return true; 1891 1892 if (IntegerType *Ty = dyn_cast<IntegerType>(V1->getType())) { 1893 // Are any known bits in V1 contradictory to known bits in V2? If V1 1894 // has a known zero where V2 has a known one, they must not be equal. 1895 auto BitWidth = Ty->getBitWidth(); 1896 APInt KnownZero1(BitWidth, 0); 1897 APInt KnownOne1(BitWidth, 0); 1898 computeKnownBits(V1, KnownZero1, KnownOne1, 0, Q); 1899 APInt KnownZero2(BitWidth, 0); 1900 APInt KnownOne2(BitWidth, 0); 1901 computeKnownBits(V2, KnownZero2, KnownOne2, 0, Q); 1902 1903 auto OppositeBits = (KnownZero1 & KnownOne2) | (KnownZero2 & KnownOne1); 1904 if (OppositeBits.getBoolValue()) 1905 return true; 1906 } 1907 return false; 1908 } 1909 1910 /// Return true if 'V & Mask' is known to be zero. We use this predicate to 1911 /// simplify operations downstream. Mask is known to be zero for bits that V 1912 /// cannot have. 1913 /// 1914 /// This function is defined on values with integer type, values with pointer 1915 /// type, and vectors of integers. In the case 1916 /// where V is a vector, the mask, known zero, and known one values are the 1917 /// same width as the vector element, and the bit is set only if it is true 1918 /// for all of the elements in the vector. 1919 bool MaskedValueIsZero(Value *V, const APInt &Mask, unsigned Depth, 1920 const Query &Q) { 1921 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0); 1922 computeKnownBits(V, KnownZero, KnownOne, Depth, Q); 1923 return (KnownZero & Mask) == Mask; 1924 } 1925 1926 /// For vector constants, loop over the elements and find the constant with the 1927 /// minimum number of sign bits. Return 0 if the value is not a vector constant 1928 /// or if any element was not analyzed; otherwise, return the count for the 1929 /// element with the minimum number of sign bits. 1930 static unsigned computeNumSignBitsVectorConstant(Value *V, unsigned TyBits) { 1931 auto *CV = dyn_cast<Constant>(V); 1932 if (!CV || !CV->getType()->isVectorTy()) 1933 return 0; 1934 1935 unsigned MinSignBits = TyBits; 1936 unsigned NumElts = CV->getType()->getVectorNumElements(); 1937 for (unsigned i = 0; i != NumElts; ++i) { 1938 // If we find a non-ConstantInt, bail out. 1939 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); 1940 if (!Elt) 1941 return 0; 1942 1943 // If the sign bit is 1, flip the bits, so we always count leading zeros. 1944 APInt EltVal = Elt->getValue(); 1945 if (EltVal.isNegative()) 1946 EltVal = ~EltVal; 1947 MinSignBits = std::min(MinSignBits, EltVal.countLeadingZeros()); 1948 } 1949 1950 return MinSignBits; 1951 } 1952 1953 /// Return the number of times the sign bit of the register is replicated into 1954 /// the other bits. We know that at least 1 bit is always equal to the sign bit 1955 /// (itself), but other cases can give us information. For example, immediately 1956 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each 1957 /// other, so we return 3. For vectors, return the number of sign bits for the 1958 /// vector element with the mininum number of known sign bits. 1959 unsigned ComputeNumSignBits(Value *V, unsigned Depth, const Query &Q) { 1960 unsigned TyBits = Q.DL.getTypeSizeInBits(V->getType()->getScalarType()); 1961 unsigned Tmp, Tmp2; 1962 unsigned FirstAnswer = 1; 1963 1964 // Note that ConstantInt is handled by the general computeKnownBits case 1965 // below. 1966 1967 if (Depth == 6) 1968 return 1; // Limit search depth. 1969 1970 Operator *U = dyn_cast<Operator>(V); 1971 switch (Operator::getOpcode(V)) { 1972 default: break; 1973 case Instruction::SExt: 1974 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); 1975 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; 1976 1977 case Instruction::SDiv: { 1978 const APInt *Denominator; 1979 // sdiv X, C -> adds log(C) sign bits. 1980 if (match(U->getOperand(1), m_APInt(Denominator))) { 1981 1982 // Ignore non-positive denominator. 1983 if (!Denominator->isStrictlyPositive()) 1984 break; 1985 1986 // Calculate the incoming numerator bits. 1987 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 1988 1989 // Add floor(log(C)) bits to the numerator bits. 1990 return std::min(TyBits, NumBits + Denominator->logBase2()); 1991 } 1992 break; 1993 } 1994 1995 case Instruction::SRem: { 1996 const APInt *Denominator; 1997 // srem X, C -> we know that the result is within [-C+1,C) when C is a 1998 // positive constant. This let us put a lower bound on the number of sign 1999 // bits. 2000 if (match(U->getOperand(1), m_APInt(Denominator))) { 2001 2002 // Ignore non-positive denominator. 2003 if (!Denominator->isStrictlyPositive()) 2004 break; 2005 2006 // Calculate the incoming numerator bits. SRem by a positive constant 2007 // can't lower the number of sign bits. 2008 unsigned NumrBits = 2009 ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2010 2011 // Calculate the leading sign bit constraints by examining the 2012 // denominator. Given that the denominator is positive, there are two 2013 // cases: 2014 // 2015 // 1. the numerator is positive. The result range is [0,C) and [0,C) u< 2016 // (1 << ceilLogBase2(C)). 2017 // 2018 // 2. the numerator is negative. Then the result range is (-C,0] and 2019 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). 2020 // 2021 // Thus a lower bound on the number of sign bits is `TyBits - 2022 // ceilLogBase2(C)`. 2023 2024 unsigned ResBits = TyBits - Denominator->ceilLogBase2(); 2025 return std::max(NumrBits, ResBits); 2026 } 2027 break; 2028 } 2029 2030 case Instruction::AShr: { 2031 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2032 // ashr X, C -> adds C sign bits. Vectors too. 2033 const APInt *ShAmt; 2034 if (match(U->getOperand(1), m_APInt(ShAmt))) { 2035 Tmp += ShAmt->getZExtValue(); 2036 if (Tmp > TyBits) Tmp = TyBits; 2037 } 2038 return Tmp; 2039 } 2040 case Instruction::Shl: { 2041 const APInt *ShAmt; 2042 if (match(U->getOperand(1), m_APInt(ShAmt))) { 2043 // shl destroys sign bits. 2044 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2045 Tmp2 = ShAmt->getZExtValue(); 2046 if (Tmp2 >= TyBits || // Bad shift. 2047 Tmp2 >= Tmp) break; // Shifted all sign bits out. 2048 return Tmp - Tmp2; 2049 } 2050 break; 2051 } 2052 case Instruction::And: 2053 case Instruction::Or: 2054 case Instruction::Xor: // NOT is handled here. 2055 // Logical binary ops preserve the number of sign bits at the worst. 2056 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2057 if (Tmp != 1) { 2058 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2059 FirstAnswer = std::min(Tmp, Tmp2); 2060 // We computed what we know about the sign bits as our first 2061 // answer. Now proceed to the generic code that uses 2062 // computeKnownBits, and pick whichever answer is better. 2063 } 2064 break; 2065 2066 case Instruction::Select: 2067 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2068 if (Tmp == 1) return 1; // Early out. 2069 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); 2070 return std::min(Tmp, Tmp2); 2071 2072 case Instruction::Add: 2073 // Add can have at most one carry bit. Thus we know that the output 2074 // is, at worst, one more bit than the inputs. 2075 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2076 if (Tmp == 1) return 1; // Early out. 2077 2078 // Special case decrementing a value (ADD X, -1): 2079 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) 2080 if (CRHS->isAllOnesValue()) { 2081 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2082 computeKnownBits(U->getOperand(0), KnownZero, KnownOne, Depth + 1, Q); 2083 2084 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2085 // sign bits set. 2086 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue()) 2087 return TyBits; 2088 2089 // If we are subtracting one from a positive number, there is no carry 2090 // out of the result. 2091 if (KnownZero.isNegative()) 2092 return Tmp; 2093 } 2094 2095 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2096 if (Tmp2 == 1) return 1; 2097 return std::min(Tmp, Tmp2)-1; 2098 2099 case Instruction::Sub: 2100 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2101 if (Tmp2 == 1) return 1; 2102 2103 // Handle NEG. 2104 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) 2105 if (CLHS->isNullValue()) { 2106 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2107 computeKnownBits(U->getOperand(1), KnownZero, KnownOne, Depth + 1, Q); 2108 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2109 // sign bits set. 2110 if ((KnownZero | APInt(TyBits, 1)).isAllOnesValue()) 2111 return TyBits; 2112 2113 // If the input is known to be positive (the sign bit is known clear), 2114 // the output of the NEG has the same number of sign bits as the input. 2115 if (KnownZero.isNegative()) 2116 return Tmp2; 2117 2118 // Otherwise, we treat this like a SUB. 2119 } 2120 2121 // Sub can have at most one carry bit. Thus we know that the output 2122 // is, at worst, one more bit than the inputs. 2123 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2124 if (Tmp == 1) return 1; // Early out. 2125 return std::min(Tmp, Tmp2)-1; 2126 2127 case Instruction::PHI: { 2128 PHINode *PN = cast<PHINode>(U); 2129 unsigned NumIncomingValues = PN->getNumIncomingValues(); 2130 // Don't analyze large in-degree PHIs. 2131 if (NumIncomingValues > 4) break; 2132 // Unreachable blocks may have zero-operand PHI nodes. 2133 if (NumIncomingValues == 0) break; 2134 2135 // Take the minimum of all incoming values. This can't infinitely loop 2136 // because of our depth threshold. 2137 Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q); 2138 for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) { 2139 if (Tmp == 1) return Tmp; 2140 Tmp = std::min( 2141 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q)); 2142 } 2143 return Tmp; 2144 } 2145 2146 case Instruction::Trunc: 2147 // FIXME: it's tricky to do anything useful for this, but it is an important 2148 // case for targets like X86. 2149 break; 2150 } 2151 2152 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2153 // use this information. 2154 2155 // If we can examine all elements of a vector constant successfully, we're 2156 // done (we can't do any better than that). If not, keep trying. 2157 if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits)) 2158 return VecSignBits; 2159 2160 APInt KnownZero(TyBits, 0), KnownOne(TyBits, 0); 2161 computeKnownBits(V, KnownZero, KnownOne, Depth, Q); 2162 2163 // If we know that the sign bit is either zero or one, determine the number of 2164 // identical bits in the top of the input value. 2165 if (KnownZero.isNegative()) 2166 return std::max(FirstAnswer, KnownZero.countLeadingOnes()); 2167 2168 if (KnownOne.isNegative()) 2169 return std::max(FirstAnswer, KnownOne.countLeadingOnes()); 2170 2171 // computeKnownBits gave us no extra information about the top bits. 2172 return FirstAnswer; 2173 } 2174 2175 /// This function computes the integer multiple of Base that equals V. 2176 /// If successful, it returns true and returns the multiple in 2177 /// Multiple. If unsuccessful, it returns false. It looks 2178 /// through SExt instructions only if LookThroughSExt is true. 2179 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, 2180 bool LookThroughSExt, unsigned Depth) { 2181 const unsigned MaxDepth = 6; 2182 2183 assert(V && "No Value?"); 2184 assert(Depth <= MaxDepth && "Limit Search Depth"); 2185 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!"); 2186 2187 Type *T = V->getType(); 2188 2189 ConstantInt *CI = dyn_cast<ConstantInt>(V); 2190 2191 if (Base == 0) 2192 return false; 2193 2194 if (Base == 1) { 2195 Multiple = V; 2196 return true; 2197 } 2198 2199 ConstantExpr *CO = dyn_cast<ConstantExpr>(V); 2200 Constant *BaseVal = ConstantInt::get(T, Base); 2201 if (CO && CO == BaseVal) { 2202 // Multiple is 1. 2203 Multiple = ConstantInt::get(T, 1); 2204 return true; 2205 } 2206 2207 if (CI && CI->getZExtValue() % Base == 0) { 2208 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base); 2209 return true; 2210 } 2211 2212 if (Depth == MaxDepth) return false; // Limit search depth. 2213 2214 Operator *I = dyn_cast<Operator>(V); 2215 if (!I) return false; 2216 2217 switch (I->getOpcode()) { 2218 default: break; 2219 case Instruction::SExt: 2220 if (!LookThroughSExt) return false; 2221 // otherwise fall through to ZExt 2222 case Instruction::ZExt: 2223 return ComputeMultiple(I->getOperand(0), Base, Multiple, 2224 LookThroughSExt, Depth+1); 2225 case Instruction::Shl: 2226 case Instruction::Mul: { 2227 Value *Op0 = I->getOperand(0); 2228 Value *Op1 = I->getOperand(1); 2229 2230 if (I->getOpcode() == Instruction::Shl) { 2231 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1); 2232 if (!Op1CI) return false; 2233 // Turn Op0 << Op1 into Op0 * 2^Op1 2234 APInt Op1Int = Op1CI->getValue(); 2235 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1); 2236 APInt API(Op1Int.getBitWidth(), 0); 2237 API.setBit(BitToSet); 2238 Op1 = ConstantInt::get(V->getContext(), API); 2239 } 2240 2241 Value *Mul0 = nullptr; 2242 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) { 2243 if (Constant *Op1C = dyn_cast<Constant>(Op1)) 2244 if (Constant *MulC = dyn_cast<Constant>(Mul0)) { 2245 if (Op1C->getType()->getPrimitiveSizeInBits() < 2246 MulC->getType()->getPrimitiveSizeInBits()) 2247 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType()); 2248 if (Op1C->getType()->getPrimitiveSizeInBits() > 2249 MulC->getType()->getPrimitiveSizeInBits()) 2250 MulC = ConstantExpr::getZExt(MulC, Op1C->getType()); 2251 2252 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1) 2253 Multiple = ConstantExpr::getMul(MulC, Op1C); 2254 return true; 2255 } 2256 2257 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0)) 2258 if (Mul0CI->getValue() == 1) { 2259 // V == Base * Op1, so return Op1 2260 Multiple = Op1; 2261 return true; 2262 } 2263 } 2264 2265 Value *Mul1 = nullptr; 2266 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) { 2267 if (Constant *Op0C = dyn_cast<Constant>(Op0)) 2268 if (Constant *MulC = dyn_cast<Constant>(Mul1)) { 2269 if (Op0C->getType()->getPrimitiveSizeInBits() < 2270 MulC->getType()->getPrimitiveSizeInBits()) 2271 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType()); 2272 if (Op0C->getType()->getPrimitiveSizeInBits() > 2273 MulC->getType()->getPrimitiveSizeInBits()) 2274 MulC = ConstantExpr::getZExt(MulC, Op0C->getType()); 2275 2276 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0) 2277 Multiple = ConstantExpr::getMul(MulC, Op0C); 2278 return true; 2279 } 2280 2281 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1)) 2282 if (Mul1CI->getValue() == 1) { 2283 // V == Base * Op0, so return Op0 2284 Multiple = Op0; 2285 return true; 2286 } 2287 } 2288 } 2289 } 2290 2291 // We could not determine if V is a multiple of Base. 2292 return false; 2293 } 2294 2295 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS, 2296 const TargetLibraryInfo *TLI) { 2297 const Function *F = ICS.getCalledFunction(); 2298 if (!F) 2299 return Intrinsic::not_intrinsic; 2300 2301 if (F->isIntrinsic()) 2302 return F->getIntrinsicID(); 2303 2304 if (!TLI) 2305 return Intrinsic::not_intrinsic; 2306 2307 LibFunc::Func Func; 2308 // We're going to make assumptions on the semantics of the functions, check 2309 // that the target knows that it's available in this environment and it does 2310 // not have local linkage. 2311 if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func)) 2312 return Intrinsic::not_intrinsic; 2313 2314 if (!ICS.onlyReadsMemory()) 2315 return Intrinsic::not_intrinsic; 2316 2317 // Otherwise check if we have a call to a function that can be turned into a 2318 // vector intrinsic. 2319 switch (Func) { 2320 default: 2321 break; 2322 case LibFunc::sin: 2323 case LibFunc::sinf: 2324 case LibFunc::sinl: 2325 return Intrinsic::sin; 2326 case LibFunc::cos: 2327 case LibFunc::cosf: 2328 case LibFunc::cosl: 2329 return Intrinsic::cos; 2330 case LibFunc::exp: 2331 case LibFunc::expf: 2332 case LibFunc::expl: 2333 return Intrinsic::exp; 2334 case LibFunc::exp2: 2335 case LibFunc::exp2f: 2336 case LibFunc::exp2l: 2337 return Intrinsic::exp2; 2338 case LibFunc::log: 2339 case LibFunc::logf: 2340 case LibFunc::logl: 2341 return Intrinsic::log; 2342 case LibFunc::log10: 2343 case LibFunc::log10f: 2344 case LibFunc::log10l: 2345 return Intrinsic::log10; 2346 case LibFunc::log2: 2347 case LibFunc::log2f: 2348 case LibFunc::log2l: 2349 return Intrinsic::log2; 2350 case LibFunc::fabs: 2351 case LibFunc::fabsf: 2352 case LibFunc::fabsl: 2353 return Intrinsic::fabs; 2354 case LibFunc::fmin: 2355 case LibFunc::fminf: 2356 case LibFunc::fminl: 2357 return Intrinsic::minnum; 2358 case LibFunc::fmax: 2359 case LibFunc::fmaxf: 2360 case LibFunc::fmaxl: 2361 return Intrinsic::maxnum; 2362 case LibFunc::copysign: 2363 case LibFunc::copysignf: 2364 case LibFunc::copysignl: 2365 return Intrinsic::copysign; 2366 case LibFunc::floor: 2367 case LibFunc::floorf: 2368 case LibFunc::floorl: 2369 return Intrinsic::floor; 2370 case LibFunc::ceil: 2371 case LibFunc::ceilf: 2372 case LibFunc::ceill: 2373 return Intrinsic::ceil; 2374 case LibFunc::trunc: 2375 case LibFunc::truncf: 2376 case LibFunc::truncl: 2377 return Intrinsic::trunc; 2378 case LibFunc::rint: 2379 case LibFunc::rintf: 2380 case LibFunc::rintl: 2381 return Intrinsic::rint; 2382 case LibFunc::nearbyint: 2383 case LibFunc::nearbyintf: 2384 case LibFunc::nearbyintl: 2385 return Intrinsic::nearbyint; 2386 case LibFunc::round: 2387 case LibFunc::roundf: 2388 case LibFunc::roundl: 2389 return Intrinsic::round; 2390 case LibFunc::pow: 2391 case LibFunc::powf: 2392 case LibFunc::powl: 2393 return Intrinsic::pow; 2394 case LibFunc::sqrt: 2395 case LibFunc::sqrtf: 2396 case LibFunc::sqrtl: 2397 if (ICS->hasNoNaNs()) 2398 return Intrinsic::sqrt; 2399 return Intrinsic::not_intrinsic; 2400 } 2401 2402 return Intrinsic::not_intrinsic; 2403 } 2404 2405 /// Return true if we can prove that the specified FP value is never equal to 2406 /// -0.0. 2407 /// 2408 /// NOTE: this function will need to be revisited when we support non-default 2409 /// rounding modes! 2410 /// 2411 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, 2412 unsigned Depth) { 2413 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) 2414 return !CFP->getValueAPF().isNegZero(); 2415 2416 // FIXME: Magic number! At the least, this should be given a name because it's 2417 // used similarly in CannotBeOrderedLessThanZero(). A better fix may be to 2418 // expose it as a parameter, so it can be used for testing / experimenting. 2419 if (Depth == 6) 2420 return false; // Limit search depth. 2421 2422 const Operator *I = dyn_cast<Operator>(V); 2423 if (!I) return false; 2424 2425 // Check if the nsz fast-math flag is set 2426 if (const FPMathOperator *FPO = dyn_cast<FPMathOperator>(I)) 2427 if (FPO->hasNoSignedZeros()) 2428 return true; 2429 2430 // (add x, 0.0) is guaranteed to return +0.0, not -0.0. 2431 if (I->getOpcode() == Instruction::FAdd) 2432 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(1))) 2433 if (CFP->isNullValue()) 2434 return true; 2435 2436 // sitofp and uitofp turn into +0.0 for zero. 2437 if (isa<SIToFPInst>(I) || isa<UIToFPInst>(I)) 2438 return true; 2439 2440 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 2441 Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI); 2442 switch (IID) { 2443 default: 2444 break; 2445 // sqrt(-0.0) = -0.0, no other negative results are possible. 2446 case Intrinsic::sqrt: 2447 return CannotBeNegativeZero(CI->getArgOperand(0), TLI, Depth + 1); 2448 // fabs(x) != -0.0 2449 case Intrinsic::fabs: 2450 return true; 2451 } 2452 } 2453 2454 return false; 2455 } 2456 2457 bool llvm::CannotBeOrderedLessThanZero(const Value *V, 2458 const TargetLibraryInfo *TLI, 2459 unsigned Depth) { 2460 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) 2461 return !CFP->getValueAPF().isNegative() || CFP->getValueAPF().isZero(); 2462 2463 // FIXME: Magic number! At the least, this should be given a name because it's 2464 // used similarly in CannotBeNegativeZero(). A better fix may be to 2465 // expose it as a parameter, so it can be used for testing / experimenting. 2466 if (Depth == 6) 2467 return false; // Limit search depth. 2468 2469 const Operator *I = dyn_cast<Operator>(V); 2470 if (!I) return false; 2471 2472 switch (I->getOpcode()) { 2473 default: break; 2474 // Unsigned integers are always nonnegative. 2475 case Instruction::UIToFP: 2476 return true; 2477 case Instruction::FMul: 2478 // x*x is always non-negative or a NaN. 2479 if (I->getOperand(0) == I->getOperand(1)) 2480 return true; 2481 // Fall through 2482 case Instruction::FAdd: 2483 case Instruction::FDiv: 2484 case Instruction::FRem: 2485 return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) && 2486 CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1); 2487 case Instruction::Select: 2488 return CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1) && 2489 CannotBeOrderedLessThanZero(I->getOperand(2), TLI, Depth + 1); 2490 case Instruction::FPExt: 2491 case Instruction::FPTrunc: 2492 // Widening/narrowing never change sign. 2493 return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1); 2494 case Instruction::Call: 2495 Intrinsic::ID IID = getIntrinsicForCallSite(cast<CallInst>(I), TLI); 2496 switch (IID) { 2497 default: 2498 break; 2499 case Intrinsic::maxnum: 2500 return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) || 2501 CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1); 2502 case Intrinsic::minnum: 2503 return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1) && 2504 CannotBeOrderedLessThanZero(I->getOperand(1), TLI, Depth + 1); 2505 case Intrinsic::exp: 2506 case Intrinsic::exp2: 2507 case Intrinsic::fabs: 2508 case Intrinsic::sqrt: 2509 return true; 2510 case Intrinsic::powi: 2511 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 2512 // powi(x,n) is non-negative if n is even. 2513 if (CI->getBitWidth() <= 64 && CI->getSExtValue() % 2u == 0) 2514 return true; 2515 } 2516 return CannotBeOrderedLessThanZero(I->getOperand(0), TLI, Depth + 1); 2517 case Intrinsic::fma: 2518 case Intrinsic::fmuladd: 2519 // x*x+y is non-negative if y is non-negative. 2520 return I->getOperand(0) == I->getOperand(1) && 2521 CannotBeOrderedLessThanZero(I->getOperand(2), TLI, Depth + 1); 2522 } 2523 break; 2524 } 2525 return false; 2526 } 2527 2528 /// If the specified value can be set by repeating the same byte in memory, 2529 /// return the i8 value that it is represented with. This is 2530 /// true for all i8 values obviously, but is also true for i32 0, i32 -1, 2531 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated 2532 /// byte store (e.g. i16 0x1234), return null. 2533 Value *llvm::isBytewiseValue(Value *V) { 2534 // All byte-wide stores are splatable, even of arbitrary variables. 2535 if (V->getType()->isIntegerTy(8)) return V; 2536 2537 // Handle 'null' ConstantArrayZero etc. 2538 if (Constant *C = dyn_cast<Constant>(V)) 2539 if (C->isNullValue()) 2540 return Constant::getNullValue(Type::getInt8Ty(V->getContext())); 2541 2542 // Constant float and double values can be handled as integer values if the 2543 // corresponding integer value is "byteable". An important case is 0.0. 2544 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 2545 if (CFP->getType()->isFloatTy()) 2546 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext())); 2547 if (CFP->getType()->isDoubleTy()) 2548 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext())); 2549 // Don't handle long double formats, which have strange constraints. 2550 } 2551 2552 // We can handle constant integers that are multiple of 8 bits. 2553 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 2554 if (CI->getBitWidth() % 8 == 0) { 2555 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!"); 2556 2557 if (!CI->getValue().isSplat(8)) 2558 return nullptr; 2559 return ConstantInt::get(V->getContext(), CI->getValue().trunc(8)); 2560 } 2561 } 2562 2563 // A ConstantDataArray/Vector is splatable if all its members are equal and 2564 // also splatable. 2565 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) { 2566 Value *Elt = CA->getElementAsConstant(0); 2567 Value *Val = isBytewiseValue(Elt); 2568 if (!Val) 2569 return nullptr; 2570 2571 for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I) 2572 if (CA->getElementAsConstant(I) != Elt) 2573 return nullptr; 2574 2575 return Val; 2576 } 2577 2578 // Conceptually, we could handle things like: 2579 // %a = zext i8 %X to i16 2580 // %b = shl i16 %a, 8 2581 // %c = or i16 %a, %b 2582 // but until there is an example that actually needs this, it doesn't seem 2583 // worth worrying about. 2584 return nullptr; 2585 } 2586 2587 2588 // This is the recursive version of BuildSubAggregate. It takes a few different 2589 // arguments. Idxs is the index within the nested struct From that we are 2590 // looking at now (which is of type IndexedType). IdxSkip is the number of 2591 // indices from Idxs that should be left out when inserting into the resulting 2592 // struct. To is the result struct built so far, new insertvalue instructions 2593 // build on that. 2594 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, 2595 SmallVectorImpl<unsigned> &Idxs, 2596 unsigned IdxSkip, 2597 Instruction *InsertBefore) { 2598 llvm::StructType *STy = dyn_cast<llvm::StructType>(IndexedType); 2599 if (STy) { 2600 // Save the original To argument so we can modify it 2601 Value *OrigTo = To; 2602 // General case, the type indexed by Idxs is a struct 2603 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2604 // Process each struct element recursively 2605 Idxs.push_back(i); 2606 Value *PrevTo = To; 2607 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, 2608 InsertBefore); 2609 Idxs.pop_back(); 2610 if (!To) { 2611 // Couldn't find any inserted value for this index? Cleanup 2612 while (PrevTo != OrigTo) { 2613 InsertValueInst* Del = cast<InsertValueInst>(PrevTo); 2614 PrevTo = Del->getAggregateOperand(); 2615 Del->eraseFromParent(); 2616 } 2617 // Stop processing elements 2618 break; 2619 } 2620 } 2621 // If we successfully found a value for each of our subaggregates 2622 if (To) 2623 return To; 2624 } 2625 // Base case, the type indexed by SourceIdxs is not a struct, or not all of 2626 // the struct's elements had a value that was inserted directly. In the latter 2627 // case, perhaps we can't determine each of the subelements individually, but 2628 // we might be able to find the complete struct somewhere. 2629 2630 // Find the value that is at that particular spot 2631 Value *V = FindInsertedValue(From, Idxs); 2632 2633 if (!V) 2634 return nullptr; 2635 2636 // Insert the value in the new (sub) aggregrate 2637 return llvm::InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), 2638 "tmp", InsertBefore); 2639 } 2640 2641 // This helper takes a nested struct and extracts a part of it (which is again a 2642 // struct) into a new value. For example, given the struct: 2643 // { a, { b, { c, d }, e } } 2644 // and the indices "1, 1" this returns 2645 // { c, d }. 2646 // 2647 // It does this by inserting an insertvalue for each element in the resulting 2648 // struct, as opposed to just inserting a single struct. This will only work if 2649 // each of the elements of the substruct are known (ie, inserted into From by an 2650 // insertvalue instruction somewhere). 2651 // 2652 // All inserted insertvalue instructions are inserted before InsertBefore 2653 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, 2654 Instruction *InsertBefore) { 2655 assert(InsertBefore && "Must have someplace to insert!"); 2656 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), 2657 idx_range); 2658 Value *To = UndefValue::get(IndexedType); 2659 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); 2660 unsigned IdxSkip = Idxs.size(); 2661 2662 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); 2663 } 2664 2665 /// Given an aggregrate and an sequence of indices, see if 2666 /// the scalar value indexed is already around as a register, for example if it 2667 /// were inserted directly into the aggregrate. 2668 /// 2669 /// If InsertBefore is not null, this function will duplicate (modified) 2670 /// insertvalues when a part of a nested struct is extracted. 2671 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, 2672 Instruction *InsertBefore) { 2673 // Nothing to index? Just return V then (this is useful at the end of our 2674 // recursion). 2675 if (idx_range.empty()) 2676 return V; 2677 // We have indices, so V should have an indexable type. 2678 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && 2679 "Not looking at a struct or array?"); 2680 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && 2681 "Invalid indices for type?"); 2682 2683 if (Constant *C = dyn_cast<Constant>(V)) { 2684 C = C->getAggregateElement(idx_range[0]); 2685 if (!C) return nullptr; 2686 return FindInsertedValue(C, idx_range.slice(1), InsertBefore); 2687 } 2688 2689 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { 2690 // Loop the indices for the insertvalue instruction in parallel with the 2691 // requested indices 2692 const unsigned *req_idx = idx_range.begin(); 2693 for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); 2694 i != e; ++i, ++req_idx) { 2695 if (req_idx == idx_range.end()) { 2696 // We can't handle this without inserting insertvalues 2697 if (!InsertBefore) 2698 return nullptr; 2699 2700 // The requested index identifies a part of a nested aggregate. Handle 2701 // this specially. For example, 2702 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 2703 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 2704 // %C = extractvalue {i32, { i32, i32 } } %B, 1 2705 // This can be changed into 2706 // %A = insertvalue {i32, i32 } undef, i32 10, 0 2707 // %C = insertvalue {i32, i32 } %A, i32 11, 1 2708 // which allows the unused 0,0 element from the nested struct to be 2709 // removed. 2710 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), 2711 InsertBefore); 2712 } 2713 2714 // This insert value inserts something else than what we are looking for. 2715 // See if the (aggregate) value inserted into has the value we are 2716 // looking for, then. 2717 if (*req_idx != *i) 2718 return FindInsertedValue(I->getAggregateOperand(), idx_range, 2719 InsertBefore); 2720 } 2721 // If we end up here, the indices of the insertvalue match with those 2722 // requested (though possibly only partially). Now we recursively look at 2723 // the inserted value, passing any remaining indices. 2724 return FindInsertedValue(I->getInsertedValueOperand(), 2725 makeArrayRef(req_idx, idx_range.end()), 2726 InsertBefore); 2727 } 2728 2729 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { 2730 // If we're extracting a value from an aggregate that was extracted from 2731 // something else, we can extract from that something else directly instead. 2732 // However, we will need to chain I's indices with the requested indices. 2733 2734 // Calculate the number of indices required 2735 unsigned size = I->getNumIndices() + idx_range.size(); 2736 // Allocate some space to put the new indices in 2737 SmallVector<unsigned, 5> Idxs; 2738 Idxs.reserve(size); 2739 // Add indices from the extract value instruction 2740 Idxs.append(I->idx_begin(), I->idx_end()); 2741 2742 // Add requested indices 2743 Idxs.append(idx_range.begin(), idx_range.end()); 2744 2745 assert(Idxs.size() == size 2746 && "Number of indices added not correct?"); 2747 2748 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); 2749 } 2750 // Otherwise, we don't know (such as, extracting from a function return value 2751 // or load instruction) 2752 return nullptr; 2753 } 2754 2755 /// Analyze the specified pointer to see if it can be expressed as a base 2756 /// pointer plus a constant offset. Return the base and offset to the caller. 2757 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 2758 const DataLayout &DL) { 2759 unsigned BitWidth = DL.getPointerTypeSizeInBits(Ptr->getType()); 2760 APInt ByteOffset(BitWidth, 0); 2761 2762 // We walk up the defs but use a visited set to handle unreachable code. In 2763 // that case, we stop after accumulating the cycle once (not that it 2764 // matters). 2765 SmallPtrSet<Value *, 16> Visited; 2766 while (Visited.insert(Ptr).second) { 2767 if (Ptr->getType()->isVectorTy()) 2768 break; 2769 2770 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 2771 APInt GEPOffset(BitWidth, 0); 2772 if (!GEP->accumulateConstantOffset(DL, GEPOffset)) 2773 break; 2774 2775 ByteOffset += GEPOffset; 2776 2777 Ptr = GEP->getPointerOperand(); 2778 } else if (Operator::getOpcode(Ptr) == Instruction::BitCast || 2779 Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) { 2780 Ptr = cast<Operator>(Ptr)->getOperand(0); 2781 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) { 2782 if (GA->isInterposable()) 2783 break; 2784 Ptr = GA->getAliasee(); 2785 } else { 2786 break; 2787 } 2788 } 2789 Offset = ByteOffset.getSExtValue(); 2790 return Ptr; 2791 } 2792 2793 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP) { 2794 // Make sure the GEP has exactly three arguments. 2795 if (GEP->getNumOperands() != 3) 2796 return false; 2797 2798 // Make sure the index-ee is a pointer to array of i8. 2799 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); 2800 if (!AT || !AT->getElementType()->isIntegerTy(8)) 2801 return false; 2802 2803 // Check to make sure that the first operand of the GEP is an integer and 2804 // has value 0 so that we are sure we're indexing into the initializer. 2805 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); 2806 if (!FirstIdx || !FirstIdx->isZero()) 2807 return false; 2808 2809 return true; 2810 } 2811 2812 /// This function computes the length of a null-terminated C string pointed to 2813 /// by V. If successful, it returns true and returns the string in Str. 2814 /// If unsuccessful, it returns false. 2815 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, 2816 uint64_t Offset, bool TrimAtNul) { 2817 assert(V); 2818 2819 // Look through bitcast instructions and geps. 2820 V = V->stripPointerCasts(); 2821 2822 // If the value is a GEP instruction or constant expression, treat it as an 2823 // offset. 2824 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 2825 // The GEP operator should be based on a pointer to string constant, and is 2826 // indexing into the string constant. 2827 if (!isGEPBasedOnPointerToString(GEP)) 2828 return false; 2829 2830 // If the second index isn't a ConstantInt, then this is a variable index 2831 // into the array. If this occurs, we can't say anything meaningful about 2832 // the string. 2833 uint64_t StartIdx = 0; 2834 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) 2835 StartIdx = CI->getZExtValue(); 2836 else 2837 return false; 2838 return getConstantStringInfo(GEP->getOperand(0), Str, StartIdx + Offset, 2839 TrimAtNul); 2840 } 2841 2842 // The GEP instruction, constant or instruction, must reference a global 2843 // variable that is a constant and is initialized. The referenced constant 2844 // initializer is the array that we'll use for optimization. 2845 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); 2846 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) 2847 return false; 2848 2849 // Handle the all-zeros case. 2850 if (GV->getInitializer()->isNullValue()) { 2851 // This is a degenerate case. The initializer is constant zero so the 2852 // length of the string must be zero. 2853 Str = ""; 2854 return true; 2855 } 2856 2857 // This must be a ConstantDataArray. 2858 const auto *Array = dyn_cast<ConstantDataArray>(GV->getInitializer()); 2859 if (!Array || !Array->isString()) 2860 return false; 2861 2862 // Get the number of elements in the array. 2863 uint64_t NumElts = Array->getType()->getArrayNumElements(); 2864 2865 // Start out with the entire array in the StringRef. 2866 Str = Array->getAsString(); 2867 2868 if (Offset > NumElts) 2869 return false; 2870 2871 // Skip over 'offset' bytes. 2872 Str = Str.substr(Offset); 2873 2874 if (TrimAtNul) { 2875 // Trim off the \0 and anything after it. If the array is not nul 2876 // terminated, we just return the whole end of string. The client may know 2877 // some other way that the string is length-bound. 2878 Str = Str.substr(0, Str.find('\0')); 2879 } 2880 return true; 2881 } 2882 2883 // These next two are very similar to the above, but also look through PHI 2884 // nodes. 2885 // TODO: See if we can integrate these two together. 2886 2887 /// If we can compute the length of the string pointed to by 2888 /// the specified pointer, return 'len+1'. If we can't, return 0. 2889 static uint64_t GetStringLengthH(Value *V, SmallPtrSetImpl<PHINode*> &PHIs) { 2890 // Look through noop bitcast instructions. 2891 V = V->stripPointerCasts(); 2892 2893 // If this is a PHI node, there are two cases: either we have already seen it 2894 // or we haven't. 2895 if (PHINode *PN = dyn_cast<PHINode>(V)) { 2896 if (!PHIs.insert(PN).second) 2897 return ~0ULL; // already in the set. 2898 2899 // If it was new, see if all the input strings are the same length. 2900 uint64_t LenSoFar = ~0ULL; 2901 for (Value *IncValue : PN->incoming_values()) { 2902 uint64_t Len = GetStringLengthH(IncValue, PHIs); 2903 if (Len == 0) return 0; // Unknown length -> unknown. 2904 2905 if (Len == ~0ULL) continue; 2906 2907 if (Len != LenSoFar && LenSoFar != ~0ULL) 2908 return 0; // Disagree -> unknown. 2909 LenSoFar = Len; 2910 } 2911 2912 // Success, all agree. 2913 return LenSoFar; 2914 } 2915 2916 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) 2917 if (SelectInst *SI = dyn_cast<SelectInst>(V)) { 2918 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs); 2919 if (Len1 == 0) return 0; 2920 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs); 2921 if (Len2 == 0) return 0; 2922 if (Len1 == ~0ULL) return Len2; 2923 if (Len2 == ~0ULL) return Len1; 2924 if (Len1 != Len2) return 0; 2925 return Len1; 2926 } 2927 2928 // Otherwise, see if we can read the string. 2929 StringRef StrData; 2930 if (!getConstantStringInfo(V, StrData)) 2931 return 0; 2932 2933 return StrData.size()+1; 2934 } 2935 2936 /// If we can compute the length of the string pointed to by 2937 /// the specified pointer, return 'len+1'. If we can't, return 0. 2938 uint64_t llvm::GetStringLength(Value *V) { 2939 if (!V->getType()->isPointerTy()) return 0; 2940 2941 SmallPtrSet<PHINode*, 32> PHIs; 2942 uint64_t Len = GetStringLengthH(V, PHIs); 2943 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return 2944 // an empty string as a length. 2945 return Len == ~0ULL ? 1 : Len; 2946 } 2947 2948 /// \brief \p PN defines a loop-variant pointer to an object. Check if the 2949 /// previous iteration of the loop was referring to the same object as \p PN. 2950 static bool isSameUnderlyingObjectInLoop(PHINode *PN, LoopInfo *LI) { 2951 // Find the loop-defined value. 2952 Loop *L = LI->getLoopFor(PN->getParent()); 2953 if (PN->getNumIncomingValues() != 2) 2954 return true; 2955 2956 // Find the value from previous iteration. 2957 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); 2958 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 2959 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); 2960 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 2961 return true; 2962 2963 // If a new pointer is loaded in the loop, the pointer references a different 2964 // object in every iteration. E.g.: 2965 // for (i) 2966 // int *p = a[i]; 2967 // ... 2968 if (auto *Load = dyn_cast<LoadInst>(PrevValue)) 2969 if (!L->isLoopInvariant(Load->getPointerOperand())) 2970 return false; 2971 return true; 2972 } 2973 2974 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL, 2975 unsigned MaxLookup) { 2976 if (!V->getType()->isPointerTy()) 2977 return V; 2978 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { 2979 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 2980 V = GEP->getPointerOperand(); 2981 } else if (Operator::getOpcode(V) == Instruction::BitCast || 2982 Operator::getOpcode(V) == Instruction::AddrSpaceCast) { 2983 V = cast<Operator>(V)->getOperand(0); 2984 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 2985 if (GA->isInterposable()) 2986 return V; 2987 V = GA->getAliasee(); 2988 } else { 2989 if (auto CS = CallSite(V)) 2990 if (Value *RV = CS.getReturnedArgOperand()) { 2991 V = RV; 2992 continue; 2993 } 2994 2995 // See if InstructionSimplify knows any relevant tricks. 2996 if (Instruction *I = dyn_cast<Instruction>(V)) 2997 // TODO: Acquire a DominatorTree and AssumptionCache and use them. 2998 if (Value *Simplified = SimplifyInstruction(I, DL, nullptr)) { 2999 V = Simplified; 3000 continue; 3001 } 3002 3003 return V; 3004 } 3005 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 3006 } 3007 return V; 3008 } 3009 3010 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects, 3011 const DataLayout &DL, LoopInfo *LI, 3012 unsigned MaxLookup) { 3013 SmallPtrSet<Value *, 4> Visited; 3014 SmallVector<Value *, 4> Worklist; 3015 Worklist.push_back(V); 3016 do { 3017 Value *P = Worklist.pop_back_val(); 3018 P = GetUnderlyingObject(P, DL, MaxLookup); 3019 3020 if (!Visited.insert(P).second) 3021 continue; 3022 3023 if (SelectInst *SI = dyn_cast<SelectInst>(P)) { 3024 Worklist.push_back(SI->getTrueValue()); 3025 Worklist.push_back(SI->getFalseValue()); 3026 continue; 3027 } 3028 3029 if (PHINode *PN = dyn_cast<PHINode>(P)) { 3030 // If this PHI changes the underlying object in every iteration of the 3031 // loop, don't look through it. Consider: 3032 // int **A; 3033 // for (i) { 3034 // Prev = Curr; // Prev = PHI (Prev_0, Curr) 3035 // Curr = A[i]; 3036 // *Prev, *Curr; 3037 // 3038 // Prev is tracking Curr one iteration behind so they refer to different 3039 // underlying objects. 3040 if (!LI || !LI->isLoopHeader(PN->getParent()) || 3041 isSameUnderlyingObjectInLoop(PN, LI)) 3042 for (Value *IncValue : PN->incoming_values()) 3043 Worklist.push_back(IncValue); 3044 continue; 3045 } 3046 3047 Objects.push_back(P); 3048 } while (!Worklist.empty()); 3049 } 3050 3051 /// Return true if the only users of this pointer are lifetime markers. 3052 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { 3053 for (const User *U : V->users()) { 3054 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 3055 if (!II) return false; 3056 3057 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 3058 II->getIntrinsicID() != Intrinsic::lifetime_end) 3059 return false; 3060 } 3061 return true; 3062 } 3063 3064 bool llvm::isSafeToSpeculativelyExecute(const Value *V, 3065 const Instruction *CtxI, 3066 const DominatorTree *DT) { 3067 const Operator *Inst = dyn_cast<Operator>(V); 3068 if (!Inst) 3069 return false; 3070 3071 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) 3072 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) 3073 if (C->canTrap()) 3074 return false; 3075 3076 switch (Inst->getOpcode()) { 3077 default: 3078 return true; 3079 case Instruction::UDiv: 3080 case Instruction::URem: { 3081 // x / y is undefined if y == 0. 3082 const APInt *V; 3083 if (match(Inst->getOperand(1), m_APInt(V))) 3084 return *V != 0; 3085 return false; 3086 } 3087 case Instruction::SDiv: 3088 case Instruction::SRem: { 3089 // x / y is undefined if y == 0 or x == INT_MIN and y == -1 3090 const APInt *Numerator, *Denominator; 3091 if (!match(Inst->getOperand(1), m_APInt(Denominator))) 3092 return false; 3093 // We cannot hoist this division if the denominator is 0. 3094 if (*Denominator == 0) 3095 return false; 3096 // It's safe to hoist if the denominator is not 0 or -1. 3097 if (*Denominator != -1) 3098 return true; 3099 // At this point we know that the denominator is -1. It is safe to hoist as 3100 // long we know that the numerator is not INT_MIN. 3101 if (match(Inst->getOperand(0), m_APInt(Numerator))) 3102 return !Numerator->isMinSignedValue(); 3103 // The numerator *might* be MinSignedValue. 3104 return false; 3105 } 3106 case Instruction::Load: { 3107 const LoadInst *LI = cast<LoadInst>(Inst); 3108 if (!LI->isUnordered() || 3109 // Speculative load may create a race that did not exist in the source. 3110 LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) || 3111 // Speculative load may load data from dirty regions. 3112 LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress)) 3113 return false; 3114 const DataLayout &DL = LI->getModule()->getDataLayout(); 3115 return isDereferenceableAndAlignedPointer(LI->getPointerOperand(), 3116 LI->getAlignment(), DL, CtxI, DT); 3117 } 3118 case Instruction::Call: { 3119 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 3120 switch (II->getIntrinsicID()) { 3121 // These synthetic intrinsics have no side-effects and just mark 3122 // information about their operands. 3123 // FIXME: There are other no-op synthetic instructions that potentially 3124 // should be considered at least *safe* to speculate... 3125 case Intrinsic::dbg_declare: 3126 case Intrinsic::dbg_value: 3127 return true; 3128 3129 case Intrinsic::bswap: 3130 case Intrinsic::ctlz: 3131 case Intrinsic::ctpop: 3132 case Intrinsic::cttz: 3133 case Intrinsic::objectsize: 3134 case Intrinsic::sadd_with_overflow: 3135 case Intrinsic::smul_with_overflow: 3136 case Intrinsic::ssub_with_overflow: 3137 case Intrinsic::uadd_with_overflow: 3138 case Intrinsic::umul_with_overflow: 3139 case Intrinsic::usub_with_overflow: 3140 return true; 3141 // These intrinsics are defined to have the same behavior as libm 3142 // functions except for setting errno. 3143 case Intrinsic::sqrt: 3144 case Intrinsic::fma: 3145 case Intrinsic::fmuladd: 3146 return true; 3147 // These intrinsics are defined to have the same behavior as libm 3148 // functions, and the corresponding libm functions never set errno. 3149 case Intrinsic::trunc: 3150 case Intrinsic::copysign: 3151 case Intrinsic::fabs: 3152 case Intrinsic::minnum: 3153 case Intrinsic::maxnum: 3154 return true; 3155 // These intrinsics are defined to have the same behavior as libm 3156 // functions, which never overflow when operating on the IEEE754 types 3157 // that we support, and never set errno otherwise. 3158 case Intrinsic::ceil: 3159 case Intrinsic::floor: 3160 case Intrinsic::nearbyint: 3161 case Intrinsic::rint: 3162 case Intrinsic::round: 3163 return true; 3164 // TODO: are convert_{from,to}_fp16 safe? 3165 // TODO: can we list target-specific intrinsics here? 3166 default: break; 3167 } 3168 } 3169 return false; // The called function could have undefined behavior or 3170 // side-effects, even if marked readnone nounwind. 3171 } 3172 case Instruction::VAArg: 3173 case Instruction::Alloca: 3174 case Instruction::Invoke: 3175 case Instruction::PHI: 3176 case Instruction::Store: 3177 case Instruction::Ret: 3178 case Instruction::Br: 3179 case Instruction::IndirectBr: 3180 case Instruction::Switch: 3181 case Instruction::Unreachable: 3182 case Instruction::Fence: 3183 case Instruction::AtomicRMW: 3184 case Instruction::AtomicCmpXchg: 3185 case Instruction::LandingPad: 3186 case Instruction::Resume: 3187 case Instruction::CatchSwitch: 3188 case Instruction::CatchPad: 3189 case Instruction::CatchRet: 3190 case Instruction::CleanupPad: 3191 case Instruction::CleanupRet: 3192 return false; // Misc instructions which have effects 3193 } 3194 } 3195 3196 bool llvm::mayBeMemoryDependent(const Instruction &I) { 3197 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I); 3198 } 3199 3200 /// Return true if we know that the specified value is never null. 3201 bool llvm::isKnownNonNull(const Value *V) { 3202 assert(V->getType()->isPointerTy() && "V must be pointer type"); 3203 3204 // Alloca never returns null, malloc might. 3205 if (isa<AllocaInst>(V)) return true; 3206 3207 // A byval, inalloca, or nonnull argument is never null. 3208 if (const Argument *A = dyn_cast<Argument>(V)) 3209 return A->hasByValOrInAllocaAttr() || A->hasNonNullAttr(); 3210 3211 // A global variable in address space 0 is non null unless extern weak. 3212 // Other address spaces may have null as a valid address for a global, 3213 // so we can't assume anything. 3214 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 3215 return !GV->hasExternalWeakLinkage() && 3216 GV->getType()->getAddressSpace() == 0; 3217 3218 // A Load tagged with nonnull metadata is never null. 3219 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) 3220 return LI->getMetadata(LLVMContext::MD_nonnull); 3221 3222 if (auto CS = ImmutableCallSite(V)) 3223 if (CS.isReturnNonNull()) 3224 return true; 3225 3226 return false; 3227 } 3228 3229 static bool isKnownNonNullFromDominatingCondition(const Value *V, 3230 const Instruction *CtxI, 3231 const DominatorTree *DT) { 3232 assert(V->getType()->isPointerTy() && "V must be pointer type"); 3233 3234 unsigned NumUsesExplored = 0; 3235 for (auto *U : V->users()) { 3236 // Avoid massive lists 3237 if (NumUsesExplored >= DomConditionsMaxUses) 3238 break; 3239 NumUsesExplored++; 3240 // Consider only compare instructions uniquely controlling a branch 3241 CmpInst::Predicate Pred; 3242 if (!match(const_cast<User *>(U), 3243 m_c_ICmp(Pred, m_Specific(V), m_Zero())) || 3244 (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)) 3245 continue; 3246 3247 for (auto *CmpU : U->users()) { 3248 if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) { 3249 assert(BI->isConditional() && "uses a comparison!"); 3250 3251 BasicBlock *NonNullSuccessor = 3252 BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0); 3253 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); 3254 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) 3255 return true; 3256 } else if (Pred == ICmpInst::ICMP_NE && 3257 match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) && 3258 DT->dominates(cast<Instruction>(CmpU), CtxI)) { 3259 return true; 3260 } 3261 } 3262 } 3263 3264 return false; 3265 } 3266 3267 bool llvm::isKnownNonNullAt(const Value *V, const Instruction *CtxI, 3268 const DominatorTree *DT) { 3269 if (isKnownNonNull(V)) 3270 return true; 3271 3272 return CtxI ? ::isKnownNonNullFromDominatingCondition(V, CtxI, DT) : false; 3273 } 3274 3275 OverflowResult llvm::computeOverflowForUnsignedMul(Value *LHS, Value *RHS, 3276 const DataLayout &DL, 3277 AssumptionCache *AC, 3278 const Instruction *CxtI, 3279 const DominatorTree *DT) { 3280 // Multiplying n * m significant bits yields a result of n + m significant 3281 // bits. If the total number of significant bits does not exceed the 3282 // result bit width (minus 1), there is no overflow. 3283 // This means if we have enough leading zero bits in the operands 3284 // we can guarantee that the result does not overflow. 3285 // Ref: "Hacker's Delight" by Henry Warren 3286 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 3287 APInt LHSKnownZero(BitWidth, 0); 3288 APInt LHSKnownOne(BitWidth, 0); 3289 APInt RHSKnownZero(BitWidth, 0); 3290 APInt RHSKnownOne(BitWidth, 0); 3291 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, DL, /*Depth=*/0, AC, CxtI, 3292 DT); 3293 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, DL, /*Depth=*/0, AC, CxtI, 3294 DT); 3295 // Note that underestimating the number of zero bits gives a more 3296 // conservative answer. 3297 unsigned ZeroBits = LHSKnownZero.countLeadingOnes() + 3298 RHSKnownZero.countLeadingOnes(); 3299 // First handle the easy case: if we have enough zero bits there's 3300 // definitely no overflow. 3301 if (ZeroBits >= BitWidth) 3302 return OverflowResult::NeverOverflows; 3303 3304 // Get the largest possible values for each operand. 3305 APInt LHSMax = ~LHSKnownZero; 3306 APInt RHSMax = ~RHSKnownZero; 3307 3308 // We know the multiply operation doesn't overflow if the maximum values for 3309 // each operand will not overflow after we multiply them together. 3310 bool MaxOverflow; 3311 LHSMax.umul_ov(RHSMax, MaxOverflow); 3312 if (!MaxOverflow) 3313 return OverflowResult::NeverOverflows; 3314 3315 // We know it always overflows if multiplying the smallest possible values for 3316 // the operands also results in overflow. 3317 bool MinOverflow; 3318 LHSKnownOne.umul_ov(RHSKnownOne, MinOverflow); 3319 if (MinOverflow) 3320 return OverflowResult::AlwaysOverflows; 3321 3322 return OverflowResult::MayOverflow; 3323 } 3324 3325 OverflowResult llvm::computeOverflowForUnsignedAdd(Value *LHS, Value *RHS, 3326 const DataLayout &DL, 3327 AssumptionCache *AC, 3328 const Instruction *CxtI, 3329 const DominatorTree *DT) { 3330 bool LHSKnownNonNegative, LHSKnownNegative; 3331 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0, 3332 AC, CxtI, DT); 3333 if (LHSKnownNonNegative || LHSKnownNegative) { 3334 bool RHSKnownNonNegative, RHSKnownNegative; 3335 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0, 3336 AC, CxtI, DT); 3337 3338 if (LHSKnownNegative && RHSKnownNegative) { 3339 // The sign bit is set in both cases: this MUST overflow. 3340 // Create a simple add instruction, and insert it into the struct. 3341 return OverflowResult::AlwaysOverflows; 3342 } 3343 3344 if (LHSKnownNonNegative && RHSKnownNonNegative) { 3345 // The sign bit is clear in both cases: this CANNOT overflow. 3346 // Create a simple add instruction, and insert it into the struct. 3347 return OverflowResult::NeverOverflows; 3348 } 3349 } 3350 3351 return OverflowResult::MayOverflow; 3352 } 3353 3354 static OverflowResult computeOverflowForSignedAdd( 3355 Value *LHS, Value *RHS, AddOperator *Add, const DataLayout &DL, 3356 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT) { 3357 if (Add && Add->hasNoSignedWrap()) { 3358 return OverflowResult::NeverOverflows; 3359 } 3360 3361 bool LHSKnownNonNegative, LHSKnownNegative; 3362 bool RHSKnownNonNegative, RHSKnownNegative; 3363 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, /*Depth=*/0, 3364 AC, CxtI, DT); 3365 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, /*Depth=*/0, 3366 AC, CxtI, DT); 3367 3368 if ((LHSKnownNonNegative && RHSKnownNegative) || 3369 (LHSKnownNegative && RHSKnownNonNegative)) { 3370 // The sign bits are opposite: this CANNOT overflow. 3371 return OverflowResult::NeverOverflows; 3372 } 3373 3374 // The remaining code needs Add to be available. Early returns if not so. 3375 if (!Add) 3376 return OverflowResult::MayOverflow; 3377 3378 // If the sign of Add is the same as at least one of the operands, this add 3379 // CANNOT overflow. This is particularly useful when the sum is 3380 // @llvm.assume'ed non-negative rather than proved so from analyzing its 3381 // operands. 3382 bool LHSOrRHSKnownNonNegative = 3383 (LHSKnownNonNegative || RHSKnownNonNegative); 3384 bool LHSOrRHSKnownNegative = (LHSKnownNegative || RHSKnownNegative); 3385 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { 3386 bool AddKnownNonNegative, AddKnownNegative; 3387 ComputeSignBit(Add, AddKnownNonNegative, AddKnownNegative, DL, 3388 /*Depth=*/0, AC, CxtI, DT); 3389 if ((AddKnownNonNegative && LHSOrRHSKnownNonNegative) || 3390 (AddKnownNegative && LHSOrRHSKnownNegative)) { 3391 return OverflowResult::NeverOverflows; 3392 } 3393 } 3394 3395 return OverflowResult::MayOverflow; 3396 } 3397 3398 bool llvm::isOverflowIntrinsicNoWrap(IntrinsicInst *II, DominatorTree &DT) { 3399 #ifndef NDEBUG 3400 auto IID = II->getIntrinsicID(); 3401 assert((IID == Intrinsic::sadd_with_overflow || 3402 IID == Intrinsic::uadd_with_overflow || 3403 IID == Intrinsic::ssub_with_overflow || 3404 IID == Intrinsic::usub_with_overflow || 3405 IID == Intrinsic::smul_with_overflow || 3406 IID == Intrinsic::umul_with_overflow) && 3407 "Not an overflow intrinsic!"); 3408 #endif 3409 3410 SmallVector<BranchInst *, 2> GuardingBranches; 3411 SmallVector<ExtractValueInst *, 2> Results; 3412 3413 for (User *U : II->users()) { 3414 if (auto *EVI = dyn_cast<ExtractValueInst>(U)) { 3415 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type"); 3416 3417 if (EVI->getIndices()[0] == 0) 3418 Results.push_back(EVI); 3419 else { 3420 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type"); 3421 3422 for (auto *U : EVI->users()) 3423 if (auto *B = dyn_cast<BranchInst>(U)) { 3424 assert(B->isConditional() && "How else is it using an i1?"); 3425 GuardingBranches.push_back(B); 3426 } 3427 } 3428 } else { 3429 // We are using the aggregate directly in a way we don't want to analyze 3430 // here (storing it to a global, say). 3431 return false; 3432 } 3433 } 3434 3435 auto AllUsesGuardedByBranch = [&](BranchInst *BI) { 3436 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); 3437 if (!NoWrapEdge.isSingleEdge()) 3438 return false; 3439 3440 // Check if all users of the add are provably no-wrap. 3441 for (auto *Result : Results) { 3442 // If the extractvalue itself is not executed on overflow, the we don't 3443 // need to check each use separately, since domination is transitive. 3444 if (DT.dominates(NoWrapEdge, Result->getParent())) 3445 continue; 3446 3447 for (auto &RU : Result->uses()) 3448 if (!DT.dominates(NoWrapEdge, RU)) 3449 return false; 3450 } 3451 3452 return true; 3453 }; 3454 3455 return any_of(GuardingBranches, AllUsesGuardedByBranch); 3456 } 3457 3458 3459 OverflowResult llvm::computeOverflowForSignedAdd(AddOperator *Add, 3460 const DataLayout &DL, 3461 AssumptionCache *AC, 3462 const Instruction *CxtI, 3463 const DominatorTree *DT) { 3464 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), 3465 Add, DL, AC, CxtI, DT); 3466 } 3467 3468 OverflowResult llvm::computeOverflowForSignedAdd(Value *LHS, Value *RHS, 3469 const DataLayout &DL, 3470 AssumptionCache *AC, 3471 const Instruction *CxtI, 3472 const DominatorTree *DT) { 3473 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); 3474 } 3475 3476 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { 3477 // A memory operation returns normally if it isn't volatile. A volatile 3478 // operation is allowed to trap. 3479 // 3480 // An atomic operation isn't guaranteed to return in a reasonable amount of 3481 // time because it's possible for another thread to interfere with it for an 3482 // arbitrary length of time, but programs aren't allowed to rely on that. 3483 if (const LoadInst *LI = dyn_cast<LoadInst>(I)) 3484 return !LI->isVolatile(); 3485 if (const StoreInst *SI = dyn_cast<StoreInst>(I)) 3486 return !SI->isVolatile(); 3487 if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I)) 3488 return !CXI->isVolatile(); 3489 if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) 3490 return !RMWI->isVolatile(); 3491 if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I)) 3492 return !MII->isVolatile(); 3493 3494 // If there is no successor, then execution can't transfer to it. 3495 if (const auto *CRI = dyn_cast<CleanupReturnInst>(I)) 3496 return !CRI->unwindsToCaller(); 3497 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) 3498 return !CatchSwitch->unwindsToCaller(); 3499 if (isa<ResumeInst>(I)) 3500 return false; 3501 if (isa<ReturnInst>(I)) 3502 return false; 3503 3504 // Calls can throw, or contain an infinite loop, or kill the process. 3505 if (CallSite CS = CallSite(const_cast<Instruction*>(I))) { 3506 // Calls which don't write to arbitrary memory are safe. 3507 // FIXME: Ignoring infinite loops without any side-effects is too aggressive, 3508 // but it's consistent with other passes. See http://llvm.org/PR965 . 3509 // FIXME: This isn't aggressive enough; a call which only writes to a 3510 // global is guaranteed to return. 3511 return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() || 3512 match(I, m_Intrinsic<Intrinsic::assume>()); 3513 } 3514 3515 // Other instructions return normally. 3516 return true; 3517 } 3518 3519 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, 3520 const Loop *L) { 3521 // The loop header is guaranteed to be executed for every iteration. 3522 // 3523 // FIXME: Relax this constraint to cover all basic blocks that are 3524 // guaranteed to be executed at every iteration. 3525 if (I->getParent() != L->getHeader()) return false; 3526 3527 for (const Instruction &LI : *L->getHeader()) { 3528 if (&LI == I) return true; 3529 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; 3530 } 3531 llvm_unreachable("Instruction not contained in its own parent basic block."); 3532 } 3533 3534 bool llvm::propagatesFullPoison(const Instruction *I) { 3535 switch (I->getOpcode()) { 3536 case Instruction::Add: 3537 case Instruction::Sub: 3538 case Instruction::Xor: 3539 case Instruction::Trunc: 3540 case Instruction::BitCast: 3541 case Instruction::AddrSpaceCast: 3542 // These operations all propagate poison unconditionally. Note that poison 3543 // is not any particular value, so xor or subtraction of poison with 3544 // itself still yields poison, not zero. 3545 return true; 3546 3547 case Instruction::AShr: 3548 case Instruction::SExt: 3549 // For these operations, one bit of the input is replicated across 3550 // multiple output bits. A replicated poison bit is still poison. 3551 return true; 3552 3553 case Instruction::Shl: { 3554 // Left shift *by* a poison value is poison. The number of 3555 // positions to shift is unsigned, so no negative values are 3556 // possible there. Left shift by zero places preserves poison. So 3557 // it only remains to consider left shift of poison by a positive 3558 // number of places. 3559 // 3560 // A left shift by a positive number of places leaves the lowest order bit 3561 // non-poisoned. However, if such a shift has a no-wrap flag, then we can 3562 // make the poison operand violate that flag, yielding a fresh full-poison 3563 // value. 3564 auto *OBO = cast<OverflowingBinaryOperator>(I); 3565 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap(); 3566 } 3567 3568 case Instruction::Mul: { 3569 // A multiplication by zero yields a non-poison zero result, so we need to 3570 // rule out zero as an operand. Conservatively, multiplication by a 3571 // non-zero constant is not multiplication by zero. 3572 // 3573 // Multiplication by a non-zero constant can leave some bits 3574 // non-poisoned. For example, a multiplication by 2 leaves the lowest 3575 // order bit unpoisoned. So we need to consider that. 3576 // 3577 // Multiplication by 1 preserves poison. If the multiplication has a 3578 // no-wrap flag, then we can make the poison operand violate that flag 3579 // when multiplied by any integer other than 0 and 1. 3580 auto *OBO = cast<OverflowingBinaryOperator>(I); 3581 if (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) { 3582 for (Value *V : OBO->operands()) { 3583 if (auto *CI = dyn_cast<ConstantInt>(V)) { 3584 // A ConstantInt cannot yield poison, so we can assume that it is 3585 // the other operand that is poison. 3586 return !CI->isZero(); 3587 } 3588 } 3589 } 3590 return false; 3591 } 3592 3593 case Instruction::ICmp: 3594 // Comparing poison with any value yields poison. This is why, for 3595 // instance, x s< (x +nsw 1) can be folded to true. 3596 return true; 3597 3598 case Instruction::GetElementPtr: 3599 // A GEP implicitly represents a sequence of additions, subtractions, 3600 // truncations, sign extensions and multiplications. The multiplications 3601 // are by the non-zero sizes of some set of types, so we do not have to be 3602 // concerned with multiplication by zero. If the GEP is in-bounds, then 3603 // these operations are implicitly no-signed-wrap so poison is propagated 3604 // by the arguments above for Add, Sub, Trunc, SExt and Mul. 3605 return cast<GEPOperator>(I)->isInBounds(); 3606 3607 default: 3608 return false; 3609 } 3610 } 3611 3612 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) { 3613 switch (I->getOpcode()) { 3614 case Instruction::Store: 3615 return cast<StoreInst>(I)->getPointerOperand(); 3616 3617 case Instruction::Load: 3618 return cast<LoadInst>(I)->getPointerOperand(); 3619 3620 case Instruction::AtomicCmpXchg: 3621 return cast<AtomicCmpXchgInst>(I)->getPointerOperand(); 3622 3623 case Instruction::AtomicRMW: 3624 return cast<AtomicRMWInst>(I)->getPointerOperand(); 3625 3626 case Instruction::UDiv: 3627 case Instruction::SDiv: 3628 case Instruction::URem: 3629 case Instruction::SRem: 3630 return I->getOperand(1); 3631 3632 default: 3633 return nullptr; 3634 } 3635 } 3636 3637 bool llvm::isKnownNotFullPoison(const Instruction *PoisonI) { 3638 // We currently only look for uses of poison values within the same basic 3639 // block, as that makes it easier to guarantee that the uses will be 3640 // executed given that PoisonI is executed. 3641 // 3642 // FIXME: Expand this to consider uses beyond the same basic block. To do 3643 // this, look out for the distinction between post-dominance and strong 3644 // post-dominance. 3645 const BasicBlock *BB = PoisonI->getParent(); 3646 3647 // Set of instructions that we have proved will yield poison if PoisonI 3648 // does. 3649 SmallSet<const Value *, 16> YieldsPoison; 3650 SmallSet<const BasicBlock *, 4> Visited; 3651 YieldsPoison.insert(PoisonI); 3652 Visited.insert(PoisonI->getParent()); 3653 3654 BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end(); 3655 3656 unsigned Iter = 0; 3657 while (Iter++ < MaxDepth) { 3658 for (auto &I : make_range(Begin, End)) { 3659 if (&I != PoisonI) { 3660 const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I); 3661 if (NotPoison != nullptr && YieldsPoison.count(NotPoison)) 3662 return true; 3663 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 3664 return false; 3665 } 3666 3667 // Mark poison that propagates from I through uses of I. 3668 if (YieldsPoison.count(&I)) { 3669 for (const User *User : I.users()) { 3670 const Instruction *UserI = cast<Instruction>(User); 3671 if (propagatesFullPoison(UserI)) 3672 YieldsPoison.insert(User); 3673 } 3674 } 3675 } 3676 3677 if (auto *NextBB = BB->getSingleSuccessor()) { 3678 if (Visited.insert(NextBB).second) { 3679 BB = NextBB; 3680 Begin = BB->getFirstNonPHI()->getIterator(); 3681 End = BB->end(); 3682 continue; 3683 } 3684 } 3685 3686 break; 3687 }; 3688 return false; 3689 } 3690 3691 static bool isKnownNonNaN(Value *V, FastMathFlags FMF) { 3692 if (FMF.noNaNs()) 3693 return true; 3694 3695 if (auto *C = dyn_cast<ConstantFP>(V)) 3696 return !C->isNaN(); 3697 return false; 3698 } 3699 3700 static bool isKnownNonZero(Value *V) { 3701 if (auto *C = dyn_cast<ConstantFP>(V)) 3702 return !C->isZero(); 3703 return false; 3704 } 3705 3706 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, 3707 FastMathFlags FMF, 3708 Value *CmpLHS, Value *CmpRHS, 3709 Value *TrueVal, Value *FalseVal, 3710 Value *&LHS, Value *&RHS) { 3711 LHS = CmpLHS; 3712 RHS = CmpRHS; 3713 3714 // If the predicate is an "or-equal" (FP) predicate, then signed zeroes may 3715 // return inconsistent results between implementations. 3716 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 3717 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) 3718 // Therefore we behave conservatively and only proceed if at least one of the 3719 // operands is known to not be zero, or if we don't care about signed zeroes. 3720 switch (Pred) { 3721 default: break; 3722 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: 3723 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: 3724 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && 3725 !isKnownNonZero(CmpRHS)) 3726 return {SPF_UNKNOWN, SPNB_NA, false}; 3727 } 3728 3729 SelectPatternNaNBehavior NaNBehavior = SPNB_NA; 3730 bool Ordered = false; 3731 3732 // When given one NaN and one non-NaN input: 3733 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. 3734 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the 3735 // ordered comparison fails), which could be NaN or non-NaN. 3736 // so here we discover exactly what NaN behavior is required/accepted. 3737 if (CmpInst::isFPPredicate(Pred)) { 3738 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); 3739 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); 3740 3741 if (LHSSafe && RHSSafe) { 3742 // Both operands are known non-NaN. 3743 NaNBehavior = SPNB_RETURNS_ANY; 3744 } else if (CmpInst::isOrdered(Pred)) { 3745 // An ordered comparison will return false when given a NaN, so it 3746 // returns the RHS. 3747 Ordered = true; 3748 if (LHSSafe) 3749 // LHS is non-NaN, so if RHS is NaN then NaN will be returned. 3750 NaNBehavior = SPNB_RETURNS_NAN; 3751 else if (RHSSafe) 3752 NaNBehavior = SPNB_RETURNS_OTHER; 3753 else 3754 // Completely unsafe. 3755 return {SPF_UNKNOWN, SPNB_NA, false}; 3756 } else { 3757 Ordered = false; 3758 // An unordered comparison will return true when given a NaN, so it 3759 // returns the LHS. 3760 if (LHSSafe) 3761 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. 3762 NaNBehavior = SPNB_RETURNS_OTHER; 3763 else if (RHSSafe) 3764 NaNBehavior = SPNB_RETURNS_NAN; 3765 else 3766 // Completely unsafe. 3767 return {SPF_UNKNOWN, SPNB_NA, false}; 3768 } 3769 } 3770 3771 if (TrueVal == CmpRHS && FalseVal == CmpLHS) { 3772 std::swap(CmpLHS, CmpRHS); 3773 Pred = CmpInst::getSwappedPredicate(Pred); 3774 if (NaNBehavior == SPNB_RETURNS_NAN) 3775 NaNBehavior = SPNB_RETURNS_OTHER; 3776 else if (NaNBehavior == SPNB_RETURNS_OTHER) 3777 NaNBehavior = SPNB_RETURNS_NAN; 3778 Ordered = !Ordered; 3779 } 3780 3781 // ([if]cmp X, Y) ? X : Y 3782 if (TrueVal == CmpLHS && FalseVal == CmpRHS) { 3783 switch (Pred) { 3784 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. 3785 case ICmpInst::ICMP_UGT: 3786 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; 3787 case ICmpInst::ICMP_SGT: 3788 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; 3789 case ICmpInst::ICMP_ULT: 3790 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; 3791 case ICmpInst::ICMP_SLT: 3792 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; 3793 case FCmpInst::FCMP_UGT: 3794 case FCmpInst::FCMP_UGE: 3795 case FCmpInst::FCMP_OGT: 3796 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; 3797 case FCmpInst::FCMP_ULT: 3798 case FCmpInst::FCMP_ULE: 3799 case FCmpInst::FCMP_OLT: 3800 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; 3801 } 3802 } 3803 3804 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) { 3805 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) || 3806 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) { 3807 3808 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X 3809 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X 3810 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) { 3811 return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false}; 3812 } 3813 3814 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X 3815 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X 3816 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) { 3817 return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false}; 3818 } 3819 } 3820 3821 // Y >s C ? ~Y : ~C == ~Y <s ~C ? ~Y : ~C = SMIN(~Y, ~C) 3822 if (const auto *C2 = dyn_cast<ConstantInt>(FalseVal)) { 3823 if (Pred == ICmpInst::ICMP_SGT && C1->getType() == C2->getType() && 3824 ~C1->getValue() == C2->getValue() && 3825 (match(TrueVal, m_Not(m_Specific(CmpLHS))) || 3826 match(CmpLHS, m_Not(m_Specific(TrueVal))))) { 3827 LHS = TrueVal; 3828 RHS = FalseVal; 3829 return {SPF_SMIN, SPNB_NA, false}; 3830 } 3831 } 3832 } 3833 3834 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5) 3835 3836 return {SPF_UNKNOWN, SPNB_NA, false}; 3837 } 3838 3839 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, 3840 Instruction::CastOps *CastOp) { 3841 CastInst *CI = dyn_cast<CastInst>(V1); 3842 Constant *C = dyn_cast<Constant>(V2); 3843 if (!CI) 3844 return nullptr; 3845 *CastOp = CI->getOpcode(); 3846 3847 if (auto *CI2 = dyn_cast<CastInst>(V2)) { 3848 // If V1 and V2 are both the same cast from the same type, we can look 3849 // through V1. 3850 if (CI2->getOpcode() == CI->getOpcode() && 3851 CI2->getSrcTy() == CI->getSrcTy()) 3852 return CI2->getOperand(0); 3853 return nullptr; 3854 } else if (!C) { 3855 return nullptr; 3856 } 3857 3858 Constant *CastedTo = nullptr; 3859 3860 if (isa<ZExtInst>(CI) && CmpI->isUnsigned()) 3861 CastedTo = ConstantExpr::getTrunc(C, CI->getSrcTy()); 3862 3863 if (isa<SExtInst>(CI) && CmpI->isSigned()) 3864 CastedTo = ConstantExpr::getTrunc(C, CI->getSrcTy(), true); 3865 3866 if (isa<TruncInst>(CI)) 3867 CastedTo = ConstantExpr::getIntegerCast(C, CI->getSrcTy(), CmpI->isSigned()); 3868 3869 if (isa<FPTruncInst>(CI)) 3870 CastedTo = ConstantExpr::getFPExtend(C, CI->getSrcTy(), true); 3871 3872 if (isa<FPExtInst>(CI)) 3873 CastedTo = ConstantExpr::getFPTrunc(C, CI->getSrcTy(), true); 3874 3875 if (isa<FPToUIInst>(CI)) 3876 CastedTo = ConstantExpr::getUIToFP(C, CI->getSrcTy(), true); 3877 3878 if (isa<FPToSIInst>(CI)) 3879 CastedTo = ConstantExpr::getSIToFP(C, CI->getSrcTy(), true); 3880 3881 if (isa<UIToFPInst>(CI)) 3882 CastedTo = ConstantExpr::getFPToUI(C, CI->getSrcTy(), true); 3883 3884 if (isa<SIToFPInst>(CI)) 3885 CastedTo = ConstantExpr::getFPToSI(C, CI->getSrcTy(), true); 3886 3887 if (!CastedTo) 3888 return nullptr; 3889 3890 Constant *CastedBack = 3891 ConstantExpr::getCast(CI->getOpcode(), CastedTo, C->getType(), true); 3892 // Make sure the cast doesn't lose any information. 3893 if (CastedBack != C) 3894 return nullptr; 3895 3896 return CastedTo; 3897 } 3898 3899 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, 3900 Instruction::CastOps *CastOp) { 3901 SelectInst *SI = dyn_cast<SelectInst>(V); 3902 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; 3903 3904 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); 3905 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; 3906 3907 CmpInst::Predicate Pred = CmpI->getPredicate(); 3908 Value *CmpLHS = CmpI->getOperand(0); 3909 Value *CmpRHS = CmpI->getOperand(1); 3910 Value *TrueVal = SI->getTrueValue(); 3911 Value *FalseVal = SI->getFalseValue(); 3912 FastMathFlags FMF; 3913 if (isa<FPMathOperator>(CmpI)) 3914 FMF = CmpI->getFastMathFlags(); 3915 3916 // Bail out early. 3917 if (CmpI->isEquality()) 3918 return {SPF_UNKNOWN, SPNB_NA, false}; 3919 3920 // Deal with type mismatches. 3921 if (CastOp && CmpLHS->getType() != TrueVal->getType()) { 3922 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) 3923 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 3924 cast<CastInst>(TrueVal)->getOperand(0), C, 3925 LHS, RHS); 3926 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) 3927 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 3928 C, cast<CastInst>(FalseVal)->getOperand(0), 3929 LHS, RHS); 3930 } 3931 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, 3932 LHS, RHS); 3933 } 3934 3935 ConstantRange llvm::getConstantRangeFromMetadata(MDNode &Ranges) { 3936 const unsigned NumRanges = Ranges.getNumOperands() / 2; 3937 assert(NumRanges >= 1 && "Must have at least one range!"); 3938 assert(Ranges.getNumOperands() % 2 == 0 && "Must be a sequence of pairs"); 3939 3940 auto *FirstLow = mdconst::extract<ConstantInt>(Ranges.getOperand(0)); 3941 auto *FirstHigh = mdconst::extract<ConstantInt>(Ranges.getOperand(1)); 3942 3943 ConstantRange CR(FirstLow->getValue(), FirstHigh->getValue()); 3944 3945 for (unsigned i = 1; i < NumRanges; ++i) { 3946 auto *Low = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); 3947 auto *High = mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); 3948 3949 // Note: unionWith will potentially create a range that contains values not 3950 // contained in any of the original N ranges. 3951 CR = CR.unionWith(ConstantRange(Low->getValue(), High->getValue())); 3952 } 3953 3954 return CR; 3955 } 3956 3957 /// Return true if "icmp Pred LHS RHS" is always true. 3958 static bool isTruePredicate(CmpInst::Predicate Pred, Value *LHS, Value *RHS, 3959 const DataLayout &DL, unsigned Depth, 3960 AssumptionCache *AC, const Instruction *CxtI, 3961 const DominatorTree *DT) { 3962 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!"); 3963 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) 3964 return true; 3965 3966 switch (Pred) { 3967 default: 3968 return false; 3969 3970 case CmpInst::ICMP_SLE: { 3971 const APInt *C; 3972 3973 // LHS s<= LHS +_{nsw} C if C >= 0 3974 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) 3975 return !C->isNegative(); 3976 return false; 3977 } 3978 3979 case CmpInst::ICMP_ULE: { 3980 const APInt *C; 3981 3982 // LHS u<= LHS +_{nuw} C for any C 3983 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) 3984 return true; 3985 3986 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) 3987 auto MatchNUWAddsToSameValue = [&](Value *A, Value *B, Value *&X, 3988 const APInt *&CA, const APInt *&CB) { 3989 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && 3990 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) 3991 return true; 3992 3993 // If X & C == 0 then (X | C) == X +_{nuw} C 3994 if (match(A, m_Or(m_Value(X), m_APInt(CA))) && 3995 match(B, m_Or(m_Specific(X), m_APInt(CB)))) { 3996 unsigned BitWidth = CA->getBitWidth(); 3997 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 3998 computeKnownBits(X, KnownZero, KnownOne, DL, Depth + 1, AC, CxtI, DT); 3999 4000 if ((KnownZero & *CA) == *CA && (KnownZero & *CB) == *CB) 4001 return true; 4002 } 4003 4004 return false; 4005 }; 4006 4007 Value *X; 4008 const APInt *CLHS, *CRHS; 4009 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) 4010 return CLHS->ule(*CRHS); 4011 4012 return false; 4013 } 4014 } 4015 } 4016 4017 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred 4018 /// ALHS ARHS" is true. Otherwise, return None. 4019 static Optional<bool> 4020 isImpliedCondOperands(CmpInst::Predicate Pred, Value *ALHS, Value *ARHS, 4021 Value *BLHS, Value *BRHS, const DataLayout &DL, 4022 unsigned Depth, AssumptionCache *AC, 4023 const Instruction *CxtI, const DominatorTree *DT) { 4024 switch (Pred) { 4025 default: 4026 return None; 4027 4028 case CmpInst::ICMP_SLT: 4029 case CmpInst::ICMP_SLE: 4030 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth, AC, CxtI, 4031 DT) && 4032 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth, AC, CxtI, DT)) 4033 return true; 4034 return None; 4035 4036 case CmpInst::ICMP_ULT: 4037 case CmpInst::ICMP_ULE: 4038 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth, AC, CxtI, 4039 DT) && 4040 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth, AC, CxtI, DT)) 4041 return true; 4042 return None; 4043 } 4044 } 4045 4046 /// Return true if the operands of the two compares match. IsSwappedOps is true 4047 /// when the operands match, but are swapped. 4048 static bool isMatchingOps(Value *ALHS, Value *ARHS, Value *BLHS, Value *BRHS, 4049 bool &IsSwappedOps) { 4050 4051 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS); 4052 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS); 4053 return IsMatchingOps || IsSwappedOps; 4054 } 4055 4056 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is 4057 /// true. Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS 4058 /// BRHS" is false. Otherwise, return None if we can't infer anything. 4059 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred, 4060 Value *ALHS, Value *ARHS, 4061 CmpInst::Predicate BPred, 4062 Value *BLHS, Value *BRHS, 4063 bool IsSwappedOps) { 4064 // Canonicalize the operands so they're matching. 4065 if (IsSwappedOps) { 4066 std::swap(BLHS, BRHS); 4067 BPred = ICmpInst::getSwappedPredicate(BPred); 4068 } 4069 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred)) 4070 return true; 4071 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred)) 4072 return false; 4073 4074 return None; 4075 } 4076 4077 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is 4078 /// true. Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS 4079 /// C2" is false. Otherwise, return None if we can't infer anything. 4080 static Optional<bool> 4081 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, Value *ALHS, 4082 ConstantInt *C1, CmpInst::Predicate BPred, 4083 Value *BLHS, ConstantInt *C2) { 4084 assert(ALHS == BLHS && "LHS operands must match."); 4085 ConstantRange DomCR = 4086 ConstantRange::makeExactICmpRegion(APred, C1->getValue()); 4087 ConstantRange CR = 4088 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue()); 4089 ConstantRange Intersection = DomCR.intersectWith(CR); 4090 ConstantRange Difference = DomCR.difference(CR); 4091 if (Intersection.isEmptySet()) 4092 return false; 4093 if (Difference.isEmptySet()) 4094 return true; 4095 return None; 4096 } 4097 4098 Optional<bool> llvm::isImpliedCondition(Value *LHS, Value *RHS, 4099 const DataLayout &DL, bool InvertAPred, 4100 unsigned Depth, AssumptionCache *AC, 4101 const Instruction *CxtI, 4102 const DominatorTree *DT) { 4103 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for example. 4104 if (LHS->getType() != RHS->getType()) 4105 return None; 4106 4107 Type *OpTy = LHS->getType(); 4108 assert(OpTy->getScalarType()->isIntegerTy(1)); 4109 4110 // LHS ==> RHS by definition 4111 if (!InvertAPred && LHS == RHS) 4112 return true; 4113 4114 if (OpTy->isVectorTy()) 4115 // TODO: extending the code below to handle vectors 4116 return None; 4117 assert(OpTy->isIntegerTy(1) && "implied by above"); 4118 4119 ICmpInst::Predicate APred, BPred; 4120 Value *ALHS, *ARHS; 4121 Value *BLHS, *BRHS; 4122 4123 if (!match(LHS, m_ICmp(APred, m_Value(ALHS), m_Value(ARHS))) || 4124 !match(RHS, m_ICmp(BPred, m_Value(BLHS), m_Value(BRHS)))) 4125 return None; 4126 4127 if (InvertAPred) 4128 APred = CmpInst::getInversePredicate(APred); 4129 4130 // Can we infer anything when the two compares have matching operands? 4131 bool IsSwappedOps; 4132 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) { 4133 if (Optional<bool> Implication = isImpliedCondMatchingOperands( 4134 APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps)) 4135 return Implication; 4136 // No amount of additional analysis will infer the second condition, so 4137 // early exit. 4138 return None; 4139 } 4140 4141 // Can we infer anything when the LHS operands match and the RHS operands are 4142 // constants (not necessarily matching)? 4143 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) { 4144 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands( 4145 APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS, 4146 cast<ConstantInt>(BRHS))) 4147 return Implication; 4148 // No amount of additional analysis will infer the second condition, so 4149 // early exit. 4150 return None; 4151 } 4152 4153 if (APred == BPred) 4154 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth, AC, 4155 CxtI, DT); 4156 4157 return None; 4158 } 4159