1 //===- ValueTracking.cpp - Walk computations to compute properties --------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains routines that help analyze properties that chains of 10 // computations have. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/ValueTracking.h" 15 #include "llvm/ADT/APFloat.h" 16 #include "llvm/ADT/APInt.h" 17 #include "llvm/ADT/ArrayRef.h" 18 #include "llvm/ADT/None.h" 19 #include "llvm/ADT/Optional.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/SmallSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/StringRef.h" 25 #include "llvm/ADT/iterator_range.h" 26 #include "llvm/Analysis/AliasAnalysis.h" 27 #include "llvm/Analysis/AssumeBundleQueries.h" 28 #include "llvm/Analysis/AssumptionCache.h" 29 #include "llvm/Analysis/GuardUtils.h" 30 #include "llvm/Analysis/InstructionSimplify.h" 31 #include "llvm/Analysis/Loads.h" 32 #include "llvm/Analysis/LoopInfo.h" 33 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 34 #include "llvm/Analysis/TargetLibraryInfo.h" 35 #include "llvm/IR/Argument.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/BasicBlock.h" 38 #include "llvm/IR/Constant.h" 39 #include "llvm/IR/ConstantRange.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DerivedTypes.h" 42 #include "llvm/IR/DiagnosticInfo.h" 43 #include "llvm/IR/Dominators.h" 44 #include "llvm/IR/Function.h" 45 #include "llvm/IR/GetElementPtrTypeIterator.h" 46 #include "llvm/IR/GlobalAlias.h" 47 #include "llvm/IR/GlobalValue.h" 48 #include "llvm/IR/GlobalVariable.h" 49 #include "llvm/IR/InstrTypes.h" 50 #include "llvm/IR/Instruction.h" 51 #include "llvm/IR/Instructions.h" 52 #include "llvm/IR/IntrinsicInst.h" 53 #include "llvm/IR/Intrinsics.h" 54 #include "llvm/IR/IntrinsicsAArch64.h" 55 #include "llvm/IR/IntrinsicsX86.h" 56 #include "llvm/IR/LLVMContext.h" 57 #include "llvm/IR/Metadata.h" 58 #include "llvm/IR/Module.h" 59 #include "llvm/IR/Operator.h" 60 #include "llvm/IR/PatternMatch.h" 61 #include "llvm/IR/Type.h" 62 #include "llvm/IR/User.h" 63 #include "llvm/IR/Value.h" 64 #include "llvm/Support/Casting.h" 65 #include "llvm/Support/CommandLine.h" 66 #include "llvm/Support/Compiler.h" 67 #include "llvm/Support/ErrorHandling.h" 68 #include "llvm/Support/KnownBits.h" 69 #include "llvm/Support/MathExtras.h" 70 #include <algorithm> 71 #include <array> 72 #include <cassert> 73 #include <cstdint> 74 #include <iterator> 75 #include <utility> 76 77 using namespace llvm; 78 using namespace llvm::PatternMatch; 79 80 // Controls the number of uses of the value searched for possible 81 // dominating comparisons. 82 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", 83 cl::Hidden, cl::init(20)); 84 85 /// Returns the bitwidth of the given scalar or pointer type. For vector types, 86 /// returns the element type's bitwidth. 87 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { 88 if (unsigned BitWidth = Ty->getScalarSizeInBits()) 89 return BitWidth; 90 91 return DL.getPointerTypeSizeInBits(Ty); 92 } 93 94 namespace { 95 96 // Simplifying using an assume can only be done in a particular control-flow 97 // context (the context instruction provides that context). If an assume and 98 // the context instruction are not in the same block then the DT helps in 99 // figuring out if we can use it. 100 struct Query { 101 const DataLayout &DL; 102 AssumptionCache *AC; 103 const Instruction *CxtI; 104 const DominatorTree *DT; 105 106 // Unlike the other analyses, this may be a nullptr because not all clients 107 // provide it currently. 108 OptimizationRemarkEmitter *ORE; 109 110 /// Set of assumptions that should be excluded from further queries. 111 /// This is because of the potential for mutual recursion to cause 112 /// computeKnownBits to repeatedly visit the same assume intrinsic. The 113 /// classic case of this is assume(x = y), which will attempt to determine 114 /// bits in x from bits in y, which will attempt to determine bits in y from 115 /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call 116 /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo 117 /// (all of which can call computeKnownBits), and so on. 118 std::array<const Value *, MaxAnalysisRecursionDepth> Excluded; 119 120 /// If true, it is safe to use metadata during simplification. 121 InstrInfoQuery IIQ; 122 123 unsigned NumExcluded = 0; 124 125 Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, 126 const DominatorTree *DT, bool UseInstrInfo, 127 OptimizationRemarkEmitter *ORE = nullptr) 128 : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {} 129 130 Query(const Query &Q, const Value *NewExcl) 131 : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE), IIQ(Q.IIQ), 132 NumExcluded(Q.NumExcluded) { 133 Excluded = Q.Excluded; 134 Excluded[NumExcluded++] = NewExcl; 135 assert(NumExcluded <= Excluded.size()); 136 } 137 138 bool isExcluded(const Value *Value) const { 139 if (NumExcluded == 0) 140 return false; 141 auto End = Excluded.begin() + NumExcluded; 142 return std::find(Excluded.begin(), End, Value) != End; 143 } 144 }; 145 146 } // end anonymous namespace 147 148 // Given the provided Value and, potentially, a context instruction, return 149 // the preferred context instruction (if any). 150 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { 151 // If we've been provided with a context instruction, then use that (provided 152 // it has been inserted). 153 if (CxtI && CxtI->getParent()) 154 return CxtI; 155 156 // If the value is really an already-inserted instruction, then use that. 157 CxtI = dyn_cast<Instruction>(V); 158 if (CxtI && CxtI->getParent()) 159 return CxtI; 160 161 return nullptr; 162 } 163 164 static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, 165 const APInt &DemandedElts, 166 APInt &DemandedLHS, APInt &DemandedRHS) { 167 // The length of scalable vectors is unknown at compile time, thus we 168 // cannot check their values 169 if (isa<ScalableVectorType>(Shuf->getType())) 170 return false; 171 172 int NumElts = 173 cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements(); 174 int NumMaskElts = cast<FixedVectorType>(Shuf->getType())->getNumElements(); 175 DemandedLHS = DemandedRHS = APInt::getNullValue(NumElts); 176 if (DemandedElts.isNullValue()) 177 return true; 178 // Simple case of a shuffle with zeroinitializer. 179 if (all_of(Shuf->getShuffleMask(), [](int Elt) { return Elt == 0; })) { 180 DemandedLHS.setBit(0); 181 return true; 182 } 183 for (int i = 0; i != NumMaskElts; ++i) { 184 if (!DemandedElts[i]) 185 continue; 186 int M = Shuf->getMaskValue(i); 187 assert(M < (NumElts * 2) && "Invalid shuffle mask constant"); 188 189 // For undef elements, we don't know anything about the common state of 190 // the shuffle result. 191 if (M == -1) 192 return false; 193 if (M < NumElts) 194 DemandedLHS.setBit(M % NumElts); 195 else 196 DemandedRHS.setBit(M % NumElts); 197 } 198 199 return true; 200 } 201 202 static void computeKnownBits(const Value *V, const APInt &DemandedElts, 203 KnownBits &Known, unsigned Depth, const Query &Q); 204 205 static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, 206 const Query &Q) { 207 // FIXME: We currently have no way to represent the DemandedElts of a scalable 208 // vector 209 if (isa<ScalableVectorType>(V->getType())) { 210 Known.resetAll(); 211 return; 212 } 213 214 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 215 APInt DemandedElts = 216 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 217 computeKnownBits(V, DemandedElts, Known, Depth, Q); 218 } 219 220 void llvm::computeKnownBits(const Value *V, KnownBits &Known, 221 const DataLayout &DL, unsigned Depth, 222 AssumptionCache *AC, const Instruction *CxtI, 223 const DominatorTree *DT, 224 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { 225 ::computeKnownBits(V, Known, Depth, 226 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 227 } 228 229 void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, 230 KnownBits &Known, const DataLayout &DL, 231 unsigned Depth, AssumptionCache *AC, 232 const Instruction *CxtI, const DominatorTree *DT, 233 OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { 234 ::computeKnownBits(V, DemandedElts, Known, Depth, 235 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 236 } 237 238 static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, 239 unsigned Depth, const Query &Q); 240 241 static KnownBits computeKnownBits(const Value *V, unsigned Depth, 242 const Query &Q); 243 244 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, 245 unsigned Depth, AssumptionCache *AC, 246 const Instruction *CxtI, 247 const DominatorTree *DT, 248 OptimizationRemarkEmitter *ORE, 249 bool UseInstrInfo) { 250 return ::computeKnownBits( 251 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 252 } 253 254 KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, 255 const DataLayout &DL, unsigned Depth, 256 AssumptionCache *AC, const Instruction *CxtI, 257 const DominatorTree *DT, 258 OptimizationRemarkEmitter *ORE, 259 bool UseInstrInfo) { 260 return ::computeKnownBits( 261 V, DemandedElts, Depth, 262 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); 263 } 264 265 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS, 266 const DataLayout &DL, AssumptionCache *AC, 267 const Instruction *CxtI, const DominatorTree *DT, 268 bool UseInstrInfo) { 269 assert(LHS->getType() == RHS->getType() && 270 "LHS and RHS should have the same type"); 271 assert(LHS->getType()->isIntOrIntVectorTy() && 272 "LHS and RHS should be integers"); 273 // Look for an inverted mask: (X & ~M) op (Y & M). 274 Value *M; 275 if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) && 276 match(RHS, m_c_And(m_Specific(M), m_Value()))) 277 return true; 278 if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) && 279 match(LHS, m_c_And(m_Specific(M), m_Value()))) 280 return true; 281 IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); 282 KnownBits LHSKnown(IT->getBitWidth()); 283 KnownBits RHSKnown(IT->getBitWidth()); 284 computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); 285 computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); 286 return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue(); 287 } 288 289 bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) { 290 for (const User *U : CxtI->users()) { 291 if (const ICmpInst *IC = dyn_cast<ICmpInst>(U)) 292 if (IC->isEquality()) 293 if (Constant *C = dyn_cast<Constant>(IC->getOperand(1))) 294 if (C->isNullValue()) 295 continue; 296 return false; 297 } 298 return true; 299 } 300 301 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, 302 const Query &Q); 303 304 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, 305 bool OrZero, unsigned Depth, 306 AssumptionCache *AC, const Instruction *CxtI, 307 const DominatorTree *DT, bool UseInstrInfo) { 308 return ::isKnownToBeAPowerOfTwo( 309 V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 310 } 311 312 static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, 313 unsigned Depth, const Query &Q); 314 315 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q); 316 317 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth, 318 AssumptionCache *AC, const Instruction *CxtI, 319 const DominatorTree *DT, bool UseInstrInfo) { 320 return ::isKnownNonZero(V, Depth, 321 Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 322 } 323 324 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL, 325 unsigned Depth, AssumptionCache *AC, 326 const Instruction *CxtI, const DominatorTree *DT, 327 bool UseInstrInfo) { 328 KnownBits Known = 329 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); 330 return Known.isNonNegative(); 331 } 332 333 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth, 334 AssumptionCache *AC, const Instruction *CxtI, 335 const DominatorTree *DT, bool UseInstrInfo) { 336 if (auto *CI = dyn_cast<ConstantInt>(V)) 337 return CI->getValue().isStrictlyPositive(); 338 339 // TODO: We'd doing two recursive queries here. We should factor this such 340 // that only a single query is needed. 341 return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) && 342 isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo); 343 } 344 345 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth, 346 AssumptionCache *AC, const Instruction *CxtI, 347 const DominatorTree *DT, bool UseInstrInfo) { 348 KnownBits Known = 349 computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); 350 return Known.isNegative(); 351 } 352 353 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, 354 const Query &Q); 355 356 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, 357 const DataLayout &DL, AssumptionCache *AC, 358 const Instruction *CxtI, const DominatorTree *DT, 359 bool UseInstrInfo) { 360 return ::isKnownNonEqual(V1, V2, 0, 361 Query(DL, AC, safeCxtI(V1, safeCxtI(V2, CxtI)), DT, 362 UseInstrInfo, /*ORE=*/nullptr)); 363 } 364 365 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, 366 const Query &Q); 367 368 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, 369 const DataLayout &DL, unsigned Depth, 370 AssumptionCache *AC, const Instruction *CxtI, 371 const DominatorTree *DT, bool UseInstrInfo) { 372 return ::MaskedValueIsZero( 373 V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 374 } 375 376 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, 377 unsigned Depth, const Query &Q); 378 379 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, 380 const Query &Q) { 381 // FIXME: We currently have no way to represent the DemandedElts of a scalable 382 // vector 383 if (isa<ScalableVectorType>(V->getType())) 384 return 1; 385 386 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 387 APInt DemandedElts = 388 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 389 return ComputeNumSignBits(V, DemandedElts, Depth, Q); 390 } 391 392 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, 393 unsigned Depth, AssumptionCache *AC, 394 const Instruction *CxtI, 395 const DominatorTree *DT, bool UseInstrInfo) { 396 return ::ComputeNumSignBits( 397 V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); 398 } 399 400 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, 401 bool NSW, const APInt &DemandedElts, 402 KnownBits &KnownOut, KnownBits &Known2, 403 unsigned Depth, const Query &Q) { 404 computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q); 405 406 // If one operand is unknown and we have no nowrap information, 407 // the result will be unknown independently of the second operand. 408 if (KnownOut.isUnknown() && !NSW) 409 return; 410 411 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); 412 KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut); 413 } 414 415 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, 416 const APInt &DemandedElts, KnownBits &Known, 417 KnownBits &Known2, unsigned Depth, 418 const Query &Q) { 419 computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q); 420 computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); 421 422 bool isKnownNegative = false; 423 bool isKnownNonNegative = false; 424 // If the multiplication is known not to overflow, compute the sign bit. 425 if (NSW) { 426 if (Op0 == Op1) { 427 // The product of a number with itself is non-negative. 428 isKnownNonNegative = true; 429 } else { 430 bool isKnownNonNegativeOp1 = Known.isNonNegative(); 431 bool isKnownNonNegativeOp0 = Known2.isNonNegative(); 432 bool isKnownNegativeOp1 = Known.isNegative(); 433 bool isKnownNegativeOp0 = Known2.isNegative(); 434 // The product of two numbers with the same sign is non-negative. 435 isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || 436 (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); 437 // The product of a negative number and a non-negative number is either 438 // negative or zero. 439 if (!isKnownNonNegative) 440 isKnownNegative = 441 (isKnownNegativeOp1 && isKnownNonNegativeOp0 && 442 Known2.isNonZero()) || 443 (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); 444 } 445 } 446 447 Known = KnownBits::computeForMul(Known, Known2); 448 449 // Only make use of no-wrap flags if we failed to compute the sign bit 450 // directly. This matters if the multiplication always overflows, in 451 // which case we prefer to follow the result of the direct computation, 452 // though as the program is invoking undefined behaviour we can choose 453 // whatever we like here. 454 if (isKnownNonNegative && !Known.isNegative()) 455 Known.makeNonNegative(); 456 else if (isKnownNegative && !Known.isNonNegative()) 457 Known.makeNegative(); 458 } 459 460 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, 461 KnownBits &Known) { 462 unsigned BitWidth = Known.getBitWidth(); 463 unsigned NumRanges = Ranges.getNumOperands() / 2; 464 assert(NumRanges >= 1); 465 466 Known.Zero.setAllBits(); 467 Known.One.setAllBits(); 468 469 for (unsigned i = 0; i < NumRanges; ++i) { 470 ConstantInt *Lower = 471 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); 472 ConstantInt *Upper = 473 mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); 474 ConstantRange Range(Lower->getValue(), Upper->getValue()); 475 476 // The first CommonPrefixBits of all values in Range are equal. 477 unsigned CommonPrefixBits = 478 (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros(); 479 APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); 480 APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth); 481 Known.One &= UnsignedMax & Mask; 482 Known.Zero &= ~UnsignedMax & Mask; 483 } 484 } 485 486 static bool isEphemeralValueOf(const Instruction *I, const Value *E) { 487 SmallVector<const Value *, 16> WorkSet(1, I); 488 SmallPtrSet<const Value *, 32> Visited; 489 SmallPtrSet<const Value *, 16> EphValues; 490 491 // The instruction defining an assumption's condition itself is always 492 // considered ephemeral to that assumption (even if it has other 493 // non-ephemeral users). See r246696's test case for an example. 494 if (is_contained(I->operands(), E)) 495 return true; 496 497 while (!WorkSet.empty()) { 498 const Value *V = WorkSet.pop_back_val(); 499 if (!Visited.insert(V).second) 500 continue; 501 502 // If all uses of this value are ephemeral, then so is this value. 503 if (llvm::all_of(V->users(), [&](const User *U) { 504 return EphValues.count(U); 505 })) { 506 if (V == E) 507 return true; 508 509 if (V == I || isSafeToSpeculativelyExecute(V)) { 510 EphValues.insert(V); 511 if (const User *U = dyn_cast<User>(V)) 512 append_range(WorkSet, U->operands()); 513 } 514 } 515 } 516 517 return false; 518 } 519 520 // Is this an intrinsic that cannot be speculated but also cannot trap? 521 bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { 522 if (const CallInst *CI = dyn_cast<CallInst>(I)) 523 if (Function *F = CI->getCalledFunction()) 524 switch (F->getIntrinsicID()) { 525 default: break; 526 // FIXME: This list is repeated from NoTTI::getIntrinsicCost. 527 case Intrinsic::assume: 528 case Intrinsic::sideeffect: 529 case Intrinsic::pseudoprobe: 530 case Intrinsic::dbg_declare: 531 case Intrinsic::dbg_value: 532 case Intrinsic::dbg_label: 533 case Intrinsic::invariant_start: 534 case Intrinsic::invariant_end: 535 case Intrinsic::lifetime_start: 536 case Intrinsic::lifetime_end: 537 case Intrinsic::experimental_noalias_scope_decl: 538 case Intrinsic::objectsize: 539 case Intrinsic::ptr_annotation: 540 case Intrinsic::var_annotation: 541 return true; 542 } 543 544 return false; 545 } 546 547 bool llvm::isValidAssumeForContext(const Instruction *Inv, 548 const Instruction *CxtI, 549 const DominatorTree *DT) { 550 // There are two restrictions on the use of an assume: 551 // 1. The assume must dominate the context (or the control flow must 552 // reach the assume whenever it reaches the context). 553 // 2. The context must not be in the assume's set of ephemeral values 554 // (otherwise we will use the assume to prove that the condition 555 // feeding the assume is trivially true, thus causing the removal of 556 // the assume). 557 558 if (Inv->getParent() == CxtI->getParent()) { 559 // If Inv and CtxI are in the same block, check if the assume (Inv) is first 560 // in the BB. 561 if (Inv->comesBefore(CxtI)) 562 return true; 563 564 // Don't let an assume affect itself - this would cause the problems 565 // `isEphemeralValueOf` is trying to prevent, and it would also make 566 // the loop below go out of bounds. 567 if (Inv == CxtI) 568 return false; 569 570 // The context comes first, but they're both in the same block. 571 // Make sure there is nothing in between that might interrupt 572 // the control flow, not even CxtI itself. 573 for (BasicBlock::const_iterator I(CxtI), IE(Inv); I != IE; ++I) 574 if (!isGuaranteedToTransferExecutionToSuccessor(&*I)) 575 return false; 576 577 return !isEphemeralValueOf(Inv, CxtI); 578 } 579 580 // Inv and CxtI are in different blocks. 581 if (DT) { 582 if (DT->dominates(Inv, CxtI)) 583 return true; 584 } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { 585 // We don't have a DT, but this trivially dominates. 586 return true; 587 } 588 589 return false; 590 } 591 592 static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { 593 // v u> y implies v != 0. 594 if (Pred == ICmpInst::ICMP_UGT) 595 return true; 596 597 // Special-case v != 0 to also handle v != null. 598 if (Pred == ICmpInst::ICMP_NE) 599 return match(RHS, m_Zero()); 600 601 // All other predicates - rely on generic ConstantRange handling. 602 const APInt *C; 603 if (!match(RHS, m_APInt(C))) 604 return false; 605 606 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C); 607 return !TrueValues.contains(APInt::getNullValue(C->getBitWidth())); 608 } 609 610 static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) { 611 // Use of assumptions is context-sensitive. If we don't have a context, we 612 // cannot use them! 613 if (!Q.AC || !Q.CxtI) 614 return false; 615 616 if (Q.CxtI && V->getType()->isPointerTy()) { 617 SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull}; 618 if (!NullPointerIsDefined(Q.CxtI->getFunction(), 619 V->getType()->getPointerAddressSpace())) 620 AttrKinds.push_back(Attribute::Dereferenceable); 621 622 if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC)) 623 return true; 624 } 625 626 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { 627 if (!AssumeVH) 628 continue; 629 CallInst *I = cast<CallInst>(AssumeVH); 630 assert(I->getFunction() == Q.CxtI->getFunction() && 631 "Got assumption for the wrong function!"); 632 if (Q.isExcluded(I)) 633 continue; 634 635 // Warning: This loop can end up being somewhat performance sensitive. 636 // We're running this loop for once for each value queried resulting in a 637 // runtime of ~O(#assumes * #values). 638 639 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 640 "must be an assume intrinsic"); 641 642 Value *RHS; 643 CmpInst::Predicate Pred; 644 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); 645 if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS)))) 646 return false; 647 648 if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT)) 649 return true; 650 } 651 652 return false; 653 } 654 655 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known, 656 unsigned Depth, const Query &Q) { 657 // Use of assumptions is context-sensitive. If we don't have a context, we 658 // cannot use them! 659 if (!Q.AC || !Q.CxtI) 660 return; 661 662 unsigned BitWidth = Known.getBitWidth(); 663 664 // Refine Known set if the pointer alignment is set by assume bundles. 665 if (V->getType()->isPointerTy()) { 666 if (RetainedKnowledge RK = getKnowledgeValidInContext( 667 V, {Attribute::Alignment}, Q.CxtI, Q.DT, Q.AC)) { 668 Known.Zero.setLowBits(Log2_32(RK.ArgValue)); 669 } 670 } 671 672 // Note that the patterns below need to be kept in sync with the code 673 // in AssumptionCache::updateAffectedValues. 674 675 for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { 676 if (!AssumeVH) 677 continue; 678 CallInst *I = cast<CallInst>(AssumeVH); 679 assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && 680 "Got assumption for the wrong function!"); 681 if (Q.isExcluded(I)) 682 continue; 683 684 // Warning: This loop can end up being somewhat performance sensitive. 685 // We're running this loop for once for each value queried resulting in a 686 // runtime of ~O(#assumes * #values). 687 688 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 689 "must be an assume intrinsic"); 690 691 Value *Arg = I->getArgOperand(0); 692 693 if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 694 assert(BitWidth == 1 && "assume operand is not i1?"); 695 Known.setAllOnes(); 696 return; 697 } 698 if (match(Arg, m_Not(m_Specific(V))) && 699 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 700 assert(BitWidth == 1 && "assume operand is not i1?"); 701 Known.setAllZero(); 702 return; 703 } 704 705 // The remaining tests are all recursive, so bail out if we hit the limit. 706 if (Depth == MaxAnalysisRecursionDepth) 707 continue; 708 709 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); 710 if (!Cmp) 711 continue; 712 713 // Note that ptrtoint may change the bitwidth. 714 Value *A, *B; 715 auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); 716 717 CmpInst::Predicate Pred; 718 uint64_t C; 719 switch (Cmp->getPredicate()) { 720 default: 721 break; 722 case ICmpInst::ICMP_EQ: 723 // assume(v = a) 724 if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A))) && 725 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 726 KnownBits RHSKnown = 727 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 728 Known.Zero |= RHSKnown.Zero; 729 Known.One |= RHSKnown.One; 730 // assume(v & b = a) 731 } else if (match(Cmp, 732 m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) && 733 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 734 KnownBits RHSKnown = 735 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 736 KnownBits MaskKnown = 737 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 738 739 // For those bits in the mask that are known to be one, we can propagate 740 // known bits from the RHS to V. 741 Known.Zero |= RHSKnown.Zero & MaskKnown.One; 742 Known.One |= RHSKnown.One & MaskKnown.One; 743 // assume(~(v & b) = a) 744 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), 745 m_Value(A))) && 746 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 747 KnownBits RHSKnown = 748 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 749 KnownBits MaskKnown = 750 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 751 752 // For those bits in the mask that are known to be one, we can propagate 753 // inverted known bits from the RHS to V. 754 Known.Zero |= RHSKnown.One & MaskKnown.One; 755 Known.One |= RHSKnown.Zero & MaskKnown.One; 756 // assume(v | b = a) 757 } else if (match(Cmp, 758 m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) && 759 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 760 KnownBits RHSKnown = 761 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 762 KnownBits BKnown = 763 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 764 765 // For those bits in B that are known to be zero, we can propagate known 766 // bits from the RHS to V. 767 Known.Zero |= RHSKnown.Zero & BKnown.Zero; 768 Known.One |= RHSKnown.One & BKnown.Zero; 769 // assume(~(v | b) = a) 770 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), 771 m_Value(A))) && 772 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 773 KnownBits RHSKnown = 774 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 775 KnownBits BKnown = 776 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 777 778 // For those bits in B that are known to be zero, we can propagate 779 // inverted known bits from the RHS to V. 780 Known.Zero |= RHSKnown.One & BKnown.Zero; 781 Known.One |= RHSKnown.Zero & BKnown.Zero; 782 // assume(v ^ b = a) 783 } else if (match(Cmp, 784 m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) && 785 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 786 KnownBits RHSKnown = 787 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 788 KnownBits BKnown = 789 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 790 791 // For those bits in B that are known to be zero, we can propagate known 792 // bits from the RHS to V. For those bits in B that are known to be one, 793 // we can propagate inverted known bits from the RHS to V. 794 Known.Zero |= RHSKnown.Zero & BKnown.Zero; 795 Known.One |= RHSKnown.One & BKnown.Zero; 796 Known.Zero |= RHSKnown.One & BKnown.One; 797 Known.One |= RHSKnown.Zero & BKnown.One; 798 // assume(~(v ^ b) = a) 799 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), 800 m_Value(A))) && 801 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 802 KnownBits RHSKnown = 803 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 804 KnownBits BKnown = 805 computeKnownBits(B, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 806 807 // For those bits in B that are known to be zero, we can propagate 808 // inverted known bits from the RHS to V. For those bits in B that are 809 // known to be one, we can propagate known bits from the RHS to V. 810 Known.Zero |= RHSKnown.One & BKnown.Zero; 811 Known.One |= RHSKnown.Zero & BKnown.Zero; 812 Known.Zero |= RHSKnown.Zero & BKnown.One; 813 Known.One |= RHSKnown.One & BKnown.One; 814 // assume(v << c = a) 815 } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), 816 m_Value(A))) && 817 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 818 KnownBits RHSKnown = 819 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 820 821 // For those bits in RHS that are known, we can propagate them to known 822 // bits in V shifted to the right by C. 823 RHSKnown.Zero.lshrInPlace(C); 824 Known.Zero |= RHSKnown.Zero; 825 RHSKnown.One.lshrInPlace(C); 826 Known.One |= RHSKnown.One; 827 // assume(~(v << c) = a) 828 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), 829 m_Value(A))) && 830 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 831 KnownBits RHSKnown = 832 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 833 // For those bits in RHS that are known, we can propagate them inverted 834 // to known bits in V shifted to the right by C. 835 RHSKnown.One.lshrInPlace(C); 836 Known.Zero |= RHSKnown.One; 837 RHSKnown.Zero.lshrInPlace(C); 838 Known.One |= RHSKnown.Zero; 839 // assume(v >> c = a) 840 } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)), 841 m_Value(A))) && 842 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 843 KnownBits RHSKnown = 844 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 845 // For those bits in RHS that are known, we can propagate them to known 846 // bits in V shifted to the right by C. 847 Known.Zero |= RHSKnown.Zero << C; 848 Known.One |= RHSKnown.One << C; 849 // assume(~(v >> c) = a) 850 } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))), 851 m_Value(A))) && 852 isValidAssumeForContext(I, Q.CxtI, Q.DT) && C < BitWidth) { 853 KnownBits RHSKnown = 854 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 855 // For those bits in RHS that are known, we can propagate them inverted 856 // to known bits in V shifted to the right by C. 857 Known.Zero |= RHSKnown.One << C; 858 Known.One |= RHSKnown.Zero << C; 859 } 860 break; 861 case ICmpInst::ICMP_SGE: 862 // assume(v >=_s c) where c is non-negative 863 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 864 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 865 KnownBits RHSKnown = 866 computeKnownBits(A, Depth + 1, Query(Q, I)).anyextOrTrunc(BitWidth); 867 868 if (RHSKnown.isNonNegative()) { 869 // We know that the sign bit is zero. 870 Known.makeNonNegative(); 871 } 872 } 873 break; 874 case ICmpInst::ICMP_SGT: 875 // assume(v >_s c) where c is at least -1. 876 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 877 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 878 KnownBits RHSKnown = 879 computeKnownBits(A, Depth + 1, Query(Q, I)).anyextOrTrunc(BitWidth); 880 881 if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) { 882 // We know that the sign bit is zero. 883 Known.makeNonNegative(); 884 } 885 } 886 break; 887 case ICmpInst::ICMP_SLE: 888 // assume(v <=_s c) where c is negative 889 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 890 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 891 KnownBits RHSKnown = 892 computeKnownBits(A, Depth + 1, Query(Q, I)).anyextOrTrunc(BitWidth); 893 894 if (RHSKnown.isNegative()) { 895 // We know that the sign bit is one. 896 Known.makeNegative(); 897 } 898 } 899 break; 900 case ICmpInst::ICMP_SLT: 901 // assume(v <_s c) where c is non-positive 902 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 903 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 904 KnownBits RHSKnown = 905 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 906 907 if (RHSKnown.isZero() || RHSKnown.isNegative()) { 908 // We know that the sign bit is one. 909 Known.makeNegative(); 910 } 911 } 912 break; 913 case ICmpInst::ICMP_ULE: 914 // assume(v <=_u c) 915 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 916 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 917 KnownBits RHSKnown = 918 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 919 920 // Whatever high bits in c are zero are known to be zero. 921 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); 922 } 923 break; 924 case ICmpInst::ICMP_ULT: 925 // assume(v <_u c) 926 if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A))) && 927 isValidAssumeForContext(I, Q.CxtI, Q.DT)) { 928 KnownBits RHSKnown = 929 computeKnownBits(A, Depth+1, Query(Q, I)).anyextOrTrunc(BitWidth); 930 931 // If the RHS is known zero, then this assumption must be wrong (nothing 932 // is unsigned less than zero). Signal a conflict and get out of here. 933 if (RHSKnown.isZero()) { 934 Known.Zero.setAllBits(); 935 Known.One.setAllBits(); 936 break; 937 } 938 939 // Whatever high bits in c are zero are known to be zero (if c is a power 940 // of 2, then one more). 941 if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I))) 942 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1); 943 else 944 Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); 945 } 946 break; 947 } 948 } 949 950 // If assumptions conflict with each other or previous known bits, then we 951 // have a logical fallacy. It's possible that the assumption is not reachable, 952 // so this isn't a real bug. On the other hand, the program may have undefined 953 // behavior, or we might have a bug in the compiler. We can't assert/crash, so 954 // clear out the known bits, try to warn the user, and hope for the best. 955 if (Known.Zero.intersects(Known.One)) { 956 Known.resetAll(); 957 958 if (Q.ORE) 959 Q.ORE->emit([&]() { 960 auto *CxtI = const_cast<Instruction *>(Q.CxtI); 961 return OptimizationRemarkAnalysis("value-tracking", "BadAssumption", 962 CxtI) 963 << "Detected conflicting code assumptions. Program may " 964 "have undefined behavior, or compiler may have " 965 "internal error."; 966 }); 967 } 968 } 969 970 /// Compute known bits from a shift operator, including those with a 971 /// non-constant shift amount. Known is the output of this function. Known2 is a 972 /// pre-allocated temporary with the same bit width as Known and on return 973 /// contains the known bit of the shift value source. KF is an 974 /// operator-specific function that, given the known-bits and a shift amount, 975 /// compute the implied known-bits of the shift operator's result respectively 976 /// for that shift amount. The results from calling KF are conservatively 977 /// combined for all permitted shift amounts. 978 static void computeKnownBitsFromShiftOperator( 979 const Operator *I, const APInt &DemandedElts, KnownBits &Known, 980 KnownBits &Known2, unsigned Depth, const Query &Q, 981 function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) { 982 unsigned BitWidth = Known.getBitWidth(); 983 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 984 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 985 986 // Note: We cannot use Known.Zero.getLimitedValue() here, because if 987 // BitWidth > 64 and any upper bits are known, we'll end up returning the 988 // limit value (which implies all bits are known). 989 uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue(); 990 uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue(); 991 bool ShiftAmtIsConstant = Known.isConstant(); 992 bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth); 993 994 if (ShiftAmtIsConstant) { 995 Known = KF(Known2, Known); 996 997 // If the known bits conflict, this must be an overflowing left shift, so 998 // the shift result is poison. We can return anything we want. Choose 0 for 999 // the best folding opportunity. 1000 if (Known.hasConflict()) 1001 Known.setAllZero(); 1002 1003 return; 1004 } 1005 1006 // If the shift amount could be greater than or equal to the bit-width of the 1007 // LHS, the value could be poison, but bail out because the check below is 1008 // expensive. 1009 // TODO: Should we just carry on? 1010 if (MaxShiftAmtIsOutOfRange) { 1011 Known.resetAll(); 1012 return; 1013 } 1014 1015 // It would be more-clearly correct to use the two temporaries for this 1016 // calculation. Reusing the APInts here to prevent unnecessary allocations. 1017 Known.resetAll(); 1018 1019 // If we know the shifter operand is nonzero, we can sometimes infer more 1020 // known bits. However this is expensive to compute, so be lazy about it and 1021 // only compute it when absolutely necessary. 1022 Optional<bool> ShifterOperandIsNonZero; 1023 1024 // Early exit if we can't constrain any well-defined shift amount. 1025 if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) && 1026 !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) { 1027 ShifterOperandIsNonZero = 1028 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); 1029 if (!*ShifterOperandIsNonZero) 1030 return; 1031 } 1032 1033 Known.Zero.setAllBits(); 1034 Known.One.setAllBits(); 1035 for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { 1036 // Combine the shifted known input bits only for those shift amounts 1037 // compatible with its known constraints. 1038 if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) 1039 continue; 1040 if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) 1041 continue; 1042 // If we know the shifter is nonzero, we may be able to infer more known 1043 // bits. This check is sunk down as far as possible to avoid the expensive 1044 // call to isKnownNonZero if the cheaper checks above fail. 1045 if (ShiftAmt == 0) { 1046 if (!ShifterOperandIsNonZero.hasValue()) 1047 ShifterOperandIsNonZero = 1048 isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); 1049 if (*ShifterOperandIsNonZero) 1050 continue; 1051 } 1052 1053 Known = KnownBits::commonBits( 1054 Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt)))); 1055 } 1056 1057 // If the known bits conflict, the result is poison. Return a 0 and hope the 1058 // caller can further optimize that. 1059 if (Known.hasConflict()) 1060 Known.setAllZero(); 1061 } 1062 1063 static void computeKnownBitsFromOperator(const Operator *I, 1064 const APInt &DemandedElts, 1065 KnownBits &Known, unsigned Depth, 1066 const Query &Q) { 1067 unsigned BitWidth = Known.getBitWidth(); 1068 1069 KnownBits Known2(BitWidth); 1070 switch (I->getOpcode()) { 1071 default: break; 1072 case Instruction::Load: 1073 if (MDNode *MD = 1074 Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range)) 1075 computeKnownBitsFromRangeMetadata(*MD, Known); 1076 break; 1077 case Instruction::And: { 1078 // If either the LHS or the RHS are Zero, the result is zero. 1079 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1080 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1081 1082 Known &= Known2; 1083 1084 // and(x, add (x, -1)) is a common idiom that always clears the low bit; 1085 // here we handle the more general case of adding any odd number by 1086 // matching the form add(x, add(x, y)) where y is odd. 1087 // TODO: This could be generalized to clearing any bit set in y where the 1088 // following bit is known to be unset in y. 1089 Value *X = nullptr, *Y = nullptr; 1090 if (!Known.Zero[0] && !Known.One[0] && 1091 match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) { 1092 Known2.resetAll(); 1093 computeKnownBits(Y, DemandedElts, Known2, Depth + 1, Q); 1094 if (Known2.countMinTrailingOnes() > 0) 1095 Known.Zero.setBit(0); 1096 } 1097 break; 1098 } 1099 case Instruction::Or: 1100 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1101 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1102 1103 Known |= Known2; 1104 break; 1105 case Instruction::Xor: 1106 computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); 1107 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1108 1109 Known ^= Known2; 1110 break; 1111 case Instruction::Mul: { 1112 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1113 computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts, 1114 Known, Known2, Depth, Q); 1115 break; 1116 } 1117 case Instruction::UDiv: { 1118 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1119 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1120 Known = KnownBits::udiv(Known, Known2); 1121 break; 1122 } 1123 case Instruction::Select: { 1124 const Value *LHS = nullptr, *RHS = nullptr; 1125 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; 1126 if (SelectPatternResult::isMinOrMax(SPF)) { 1127 computeKnownBits(RHS, Known, Depth + 1, Q); 1128 computeKnownBits(LHS, Known2, Depth + 1, Q); 1129 switch (SPF) { 1130 default: 1131 llvm_unreachable("Unhandled select pattern flavor!"); 1132 case SPF_SMAX: 1133 Known = KnownBits::smax(Known, Known2); 1134 break; 1135 case SPF_SMIN: 1136 Known = KnownBits::smin(Known, Known2); 1137 break; 1138 case SPF_UMAX: 1139 Known = KnownBits::umax(Known, Known2); 1140 break; 1141 case SPF_UMIN: 1142 Known = KnownBits::umin(Known, Known2); 1143 break; 1144 } 1145 break; 1146 } 1147 1148 computeKnownBits(I->getOperand(2), Known, Depth + 1, Q); 1149 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1150 1151 // Only known if known in both the LHS and RHS. 1152 Known = KnownBits::commonBits(Known, Known2); 1153 1154 if (SPF == SPF_ABS) { 1155 // RHS from matchSelectPattern returns the negation part of abs pattern. 1156 // If the negate has an NSW flag we can assume the sign bit of the result 1157 // will be 0 because that makes abs(INT_MIN) undefined. 1158 if (match(RHS, m_Neg(m_Specific(LHS))) && 1159 Q.IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) 1160 Known.Zero.setSignBit(); 1161 } 1162 1163 break; 1164 } 1165 case Instruction::FPTrunc: 1166 case Instruction::FPExt: 1167 case Instruction::FPToUI: 1168 case Instruction::FPToSI: 1169 case Instruction::SIToFP: 1170 case Instruction::UIToFP: 1171 break; // Can't work with floating point. 1172 case Instruction::PtrToInt: 1173 case Instruction::IntToPtr: 1174 // Fall through and handle them the same as zext/trunc. 1175 LLVM_FALLTHROUGH; 1176 case Instruction::ZExt: 1177 case Instruction::Trunc: { 1178 Type *SrcTy = I->getOperand(0)->getType(); 1179 1180 unsigned SrcBitWidth; 1181 // Note that we handle pointer operands here because of inttoptr/ptrtoint 1182 // which fall through here. 1183 Type *ScalarTy = SrcTy->getScalarType(); 1184 SrcBitWidth = ScalarTy->isPointerTy() ? 1185 Q.DL.getPointerTypeSizeInBits(ScalarTy) : 1186 Q.DL.getTypeSizeInBits(ScalarTy); 1187 1188 assert(SrcBitWidth && "SrcBitWidth can't be zero"); 1189 Known = Known.anyextOrTrunc(SrcBitWidth); 1190 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1191 Known = Known.zextOrTrunc(BitWidth); 1192 break; 1193 } 1194 case Instruction::BitCast: { 1195 Type *SrcTy = I->getOperand(0)->getType(); 1196 if (SrcTy->isIntOrPtrTy() && 1197 // TODO: For now, not handling conversions like: 1198 // (bitcast i64 %x to <2 x i32>) 1199 !I->getType()->isVectorTy()) { 1200 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1201 break; 1202 } 1203 break; 1204 } 1205 case Instruction::SExt: { 1206 // Compute the bits in the result that are not present in the input. 1207 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 1208 1209 Known = Known.trunc(SrcBitWidth); 1210 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1211 // If the sign bit of the input is known set or clear, then we know the 1212 // top bits of the result. 1213 Known = Known.sext(BitWidth); 1214 break; 1215 } 1216 case Instruction::Shl: { 1217 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1218 auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1219 KnownBits Result = KnownBits::shl(KnownVal, KnownAmt); 1220 // If this shift has "nsw" keyword, then the result is either a poison 1221 // value or has the same sign bit as the first operand. 1222 if (NSW) { 1223 if (KnownVal.Zero.isSignBitSet()) 1224 Result.Zero.setSignBit(); 1225 if (KnownVal.One.isSignBitSet()) 1226 Result.One.setSignBit(); 1227 } 1228 return Result; 1229 }; 1230 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1231 KF); 1232 break; 1233 } 1234 case Instruction::LShr: { 1235 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1236 return KnownBits::lshr(KnownVal, KnownAmt); 1237 }; 1238 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1239 KF); 1240 break; 1241 } 1242 case Instruction::AShr: { 1243 auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { 1244 return KnownBits::ashr(KnownVal, KnownAmt); 1245 }; 1246 computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, 1247 KF); 1248 break; 1249 } 1250 case Instruction::Sub: { 1251 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1252 computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, 1253 DemandedElts, Known, Known2, Depth, Q); 1254 break; 1255 } 1256 case Instruction::Add: { 1257 bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); 1258 computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, 1259 DemandedElts, Known, Known2, Depth, Q); 1260 break; 1261 } 1262 case Instruction::SRem: 1263 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1264 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1265 Known = KnownBits::srem(Known, Known2); 1266 break; 1267 1268 case Instruction::URem: 1269 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1270 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1271 Known = KnownBits::urem(Known, Known2); 1272 break; 1273 case Instruction::Alloca: 1274 Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign())); 1275 break; 1276 case Instruction::GetElementPtr: { 1277 // Analyze all of the subscripts of this getelementptr instruction 1278 // to determine if we can prove known low zero bits. 1279 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1280 // Accumulate the constant indices in a separate variable 1281 // to minimize the number of calls to computeForAddSub. 1282 APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); 1283 1284 gep_type_iterator GTI = gep_type_begin(I); 1285 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { 1286 // TrailZ can only become smaller, short-circuit if we hit zero. 1287 if (Known.isUnknown()) 1288 break; 1289 1290 Value *Index = I->getOperand(i); 1291 1292 // Handle case when index is zero. 1293 Constant *CIndex = dyn_cast<Constant>(Index); 1294 if (CIndex && CIndex->isZeroValue()) 1295 continue; 1296 1297 if (StructType *STy = GTI.getStructTypeOrNull()) { 1298 // Handle struct member offset arithmetic. 1299 1300 assert(CIndex && 1301 "Access to structure field must be known at compile time"); 1302 1303 if (CIndex->getType()->isVectorTy()) 1304 Index = CIndex->getSplatValue(); 1305 1306 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); 1307 const StructLayout *SL = Q.DL.getStructLayout(STy); 1308 uint64_t Offset = SL->getElementOffset(Idx); 1309 AccConstIndices += Offset; 1310 continue; 1311 } 1312 1313 // Handle array index arithmetic. 1314 Type *IndexedTy = GTI.getIndexedType(); 1315 if (!IndexedTy->isSized()) { 1316 Known.resetAll(); 1317 break; 1318 } 1319 1320 unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); 1321 KnownBits IndexBits(IndexBitWidth); 1322 computeKnownBits(Index, IndexBits, Depth + 1, Q); 1323 TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy); 1324 uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinSize(); 1325 KnownBits ScalingFactor(IndexBitWidth); 1326 // Multiply by current sizeof type. 1327 // &A[i] == A + i * sizeof(*A[i]). 1328 if (IndexTypeSize.isScalable()) { 1329 // For scalable types the only thing we know about sizeof is 1330 // that this is a multiple of the minimum size. 1331 ScalingFactor.Zero.setLowBits(countTrailingZeros(TypeSizeInBytes)); 1332 } else if (IndexBits.isConstant()) { 1333 APInt IndexConst = IndexBits.getConstant(); 1334 APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); 1335 IndexConst *= ScalingFactor; 1336 AccConstIndices += IndexConst.sextOrTrunc(BitWidth); 1337 continue; 1338 } else { 1339 ScalingFactor = 1340 KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes)); 1341 } 1342 IndexBits = KnownBits::computeForMul(IndexBits, ScalingFactor); 1343 1344 // If the offsets have a different width from the pointer, according 1345 // to the language reference we need to sign-extend or truncate them 1346 // to the width of the pointer. 1347 IndexBits = IndexBits.sextOrTrunc(BitWidth); 1348 1349 // Note that inbounds does *not* guarantee nsw for the addition, as only 1350 // the offset is signed, while the base address is unsigned. 1351 Known = KnownBits::computeForAddSub( 1352 /*Add=*/true, /*NSW=*/false, Known, IndexBits); 1353 } 1354 if (!Known.isUnknown() && !AccConstIndices.isNullValue()) { 1355 KnownBits Index = KnownBits::makeConstant(AccConstIndices); 1356 Known = KnownBits::computeForAddSub( 1357 /*Add=*/true, /*NSW=*/false, Known, Index); 1358 } 1359 break; 1360 } 1361 case Instruction::PHI: { 1362 const PHINode *P = cast<PHINode>(I); 1363 // Handle the case of a simple two-predecessor recurrence PHI. 1364 // There's a lot more that could theoretically be done here, but 1365 // this is sufficient to catch some interesting cases. 1366 if (P->getNumIncomingValues() == 2) { 1367 for (unsigned i = 0; i != 2; ++i) { 1368 Value *L = P->getIncomingValue(i); 1369 Value *R = P->getIncomingValue(!i); 1370 Instruction *RInst = P->getIncomingBlock(!i)->getTerminator(); 1371 Instruction *LInst = P->getIncomingBlock(i)->getTerminator(); 1372 Operator *LU = dyn_cast<Operator>(L); 1373 if (!LU) 1374 continue; 1375 unsigned Opcode = LU->getOpcode(); 1376 // Check for operations that have the property that if 1377 // both their operands have low zero bits, the result 1378 // will have low zero bits. 1379 if (Opcode == Instruction::Add || 1380 Opcode == Instruction::Sub || 1381 Opcode == Instruction::And || 1382 Opcode == Instruction::Or || 1383 Opcode == Instruction::Mul) { 1384 Value *LL = LU->getOperand(0); 1385 Value *LR = LU->getOperand(1); 1386 // Find a recurrence. 1387 if (LL == I) 1388 L = LR; 1389 else if (LR == I) 1390 L = LL; 1391 else 1392 continue; // Check for recurrence with L and R flipped. 1393 1394 // Change the context instruction to the "edge" that flows into the 1395 // phi. This is important because that is where the value is actually 1396 // "evaluated" even though it is used later somewhere else. (see also 1397 // D69571). 1398 Query RecQ = Q; 1399 1400 // Ok, we have a PHI of the form L op= R. Check for low 1401 // zero bits. 1402 RecQ.CxtI = RInst; 1403 computeKnownBits(R, Known2, Depth + 1, RecQ); 1404 1405 // We need to take the minimum number of known bits 1406 KnownBits Known3(BitWidth); 1407 RecQ.CxtI = LInst; 1408 computeKnownBits(L, Known3, Depth + 1, RecQ); 1409 1410 Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(), 1411 Known3.countMinTrailingZeros())); 1412 1413 auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU); 1414 if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) { 1415 // If initial value of recurrence is nonnegative, and we are adding 1416 // a nonnegative number with nsw, the result can only be nonnegative 1417 // or poison value regardless of the number of times we execute the 1418 // add in phi recurrence. If initial value is negative and we are 1419 // adding a negative number with nsw, the result can only be 1420 // negative or poison value. Similar arguments apply to sub and mul. 1421 // 1422 // (add non-negative, non-negative) --> non-negative 1423 // (add negative, negative) --> negative 1424 if (Opcode == Instruction::Add) { 1425 if (Known2.isNonNegative() && Known3.isNonNegative()) 1426 Known.makeNonNegative(); 1427 else if (Known2.isNegative() && Known3.isNegative()) 1428 Known.makeNegative(); 1429 } 1430 1431 // (sub nsw non-negative, negative) --> non-negative 1432 // (sub nsw negative, non-negative) --> negative 1433 else if (Opcode == Instruction::Sub && LL == I) { 1434 if (Known2.isNonNegative() && Known3.isNegative()) 1435 Known.makeNonNegative(); 1436 else if (Known2.isNegative() && Known3.isNonNegative()) 1437 Known.makeNegative(); 1438 } 1439 1440 // (mul nsw non-negative, non-negative) --> non-negative 1441 else if (Opcode == Instruction::Mul && Known2.isNonNegative() && 1442 Known3.isNonNegative()) 1443 Known.makeNonNegative(); 1444 } 1445 1446 break; 1447 } 1448 } 1449 } 1450 1451 // Unreachable blocks may have zero-operand PHI nodes. 1452 if (P->getNumIncomingValues() == 0) 1453 break; 1454 1455 // Otherwise take the unions of the known bit sets of the operands, 1456 // taking conservative care to avoid excessive recursion. 1457 if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) { 1458 // Skip if every incoming value references to ourself. 1459 if (dyn_cast_or_null<UndefValue>(P->hasConstantValue())) 1460 break; 1461 1462 Known.Zero.setAllBits(); 1463 Known.One.setAllBits(); 1464 for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { 1465 Value *IncValue = P->getIncomingValue(u); 1466 // Skip direct self references. 1467 if (IncValue == P) continue; 1468 1469 // Change the context instruction to the "edge" that flows into the 1470 // phi. This is important because that is where the value is actually 1471 // "evaluated" even though it is used later somewhere else. (see also 1472 // D69571). 1473 Query RecQ = Q; 1474 RecQ.CxtI = P->getIncomingBlock(u)->getTerminator(); 1475 1476 Known2 = KnownBits(BitWidth); 1477 // Recurse, but cap the recursion to one level, because we don't 1478 // want to waste time spinning around in loops. 1479 computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ); 1480 Known = KnownBits::commonBits(Known, Known2); 1481 // If all bits have been ruled out, there's no need to check 1482 // more operands. 1483 if (Known.isUnknown()) 1484 break; 1485 } 1486 } 1487 break; 1488 } 1489 case Instruction::Call: 1490 case Instruction::Invoke: 1491 // If range metadata is attached to this call, set known bits from that, 1492 // and then intersect with known bits based on other properties of the 1493 // function. 1494 if (MDNode *MD = 1495 Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range)) 1496 computeKnownBitsFromRangeMetadata(*MD, Known); 1497 if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) { 1498 computeKnownBits(RV, Known2, Depth + 1, Q); 1499 Known.Zero |= Known2.Zero; 1500 Known.One |= Known2.One; 1501 } 1502 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1503 switch (II->getIntrinsicID()) { 1504 default: break; 1505 case Intrinsic::abs: { 1506 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1507 bool IntMinIsPoison = match(II->getArgOperand(1), m_One()); 1508 Known = Known2.abs(IntMinIsPoison); 1509 break; 1510 } 1511 case Intrinsic::bitreverse: 1512 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1513 Known.Zero |= Known2.Zero.reverseBits(); 1514 Known.One |= Known2.One.reverseBits(); 1515 break; 1516 case Intrinsic::bswap: 1517 computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); 1518 Known.Zero |= Known2.Zero.byteSwap(); 1519 Known.One |= Known2.One.byteSwap(); 1520 break; 1521 case Intrinsic::ctlz: { 1522 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1523 // If we have a known 1, its position is our upper bound. 1524 unsigned PossibleLZ = Known2.countMaxLeadingZeros(); 1525 // If this call is undefined for 0, the result will be less than 2^n. 1526 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1527 PossibleLZ = std::min(PossibleLZ, BitWidth - 1); 1528 unsigned LowBits = Log2_32(PossibleLZ)+1; 1529 Known.Zero.setBitsFrom(LowBits); 1530 break; 1531 } 1532 case Intrinsic::cttz: { 1533 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1534 // If we have a known 1, its position is our upper bound. 1535 unsigned PossibleTZ = Known2.countMaxTrailingZeros(); 1536 // If this call is undefined for 0, the result will be less than 2^n. 1537 if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) 1538 PossibleTZ = std::min(PossibleTZ, BitWidth - 1); 1539 unsigned LowBits = Log2_32(PossibleTZ)+1; 1540 Known.Zero.setBitsFrom(LowBits); 1541 break; 1542 } 1543 case Intrinsic::ctpop: { 1544 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1545 // We can bound the space the count needs. Also, bits known to be zero 1546 // can't contribute to the population. 1547 unsigned BitsPossiblySet = Known2.countMaxPopulation(); 1548 unsigned LowBits = Log2_32(BitsPossiblySet)+1; 1549 Known.Zero.setBitsFrom(LowBits); 1550 // TODO: we could bound KnownOne using the lower bound on the number 1551 // of bits which might be set provided by popcnt KnownOne2. 1552 break; 1553 } 1554 case Intrinsic::fshr: 1555 case Intrinsic::fshl: { 1556 const APInt *SA; 1557 if (!match(I->getOperand(2), m_APInt(SA))) 1558 break; 1559 1560 // Normalize to funnel shift left. 1561 uint64_t ShiftAmt = SA->urem(BitWidth); 1562 if (II->getIntrinsicID() == Intrinsic::fshr) 1563 ShiftAmt = BitWidth - ShiftAmt; 1564 1565 KnownBits Known3(BitWidth); 1566 computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); 1567 computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q); 1568 1569 Known.Zero = 1570 Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt); 1571 Known.One = 1572 Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt); 1573 break; 1574 } 1575 case Intrinsic::uadd_sat: 1576 case Intrinsic::usub_sat: { 1577 bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat; 1578 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1579 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1580 1581 // Add: Leading ones of either operand are preserved. 1582 // Sub: Leading zeros of LHS and leading ones of RHS are preserved 1583 // as leading zeros in the result. 1584 unsigned LeadingKnown; 1585 if (IsAdd) 1586 LeadingKnown = std::max(Known.countMinLeadingOnes(), 1587 Known2.countMinLeadingOnes()); 1588 else 1589 LeadingKnown = std::max(Known.countMinLeadingZeros(), 1590 Known2.countMinLeadingOnes()); 1591 1592 Known = KnownBits::computeForAddSub( 1593 IsAdd, /* NSW */ false, Known, Known2); 1594 1595 // We select between the operation result and all-ones/zero 1596 // respectively, so we can preserve known ones/zeros. 1597 if (IsAdd) { 1598 Known.One.setHighBits(LeadingKnown); 1599 Known.Zero.clearAllBits(); 1600 } else { 1601 Known.Zero.setHighBits(LeadingKnown); 1602 Known.One.clearAllBits(); 1603 } 1604 break; 1605 } 1606 case Intrinsic::umin: 1607 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1608 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1609 Known = KnownBits::umin(Known, Known2); 1610 break; 1611 case Intrinsic::umax: 1612 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1613 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1614 Known = KnownBits::umax(Known, Known2); 1615 break; 1616 case Intrinsic::smin: 1617 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1618 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1619 Known = KnownBits::smin(Known, Known2); 1620 break; 1621 case Intrinsic::smax: 1622 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1623 computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); 1624 Known = KnownBits::smax(Known, Known2); 1625 break; 1626 case Intrinsic::x86_sse42_crc32_64_64: 1627 Known.Zero.setBitsFrom(32); 1628 break; 1629 } 1630 } 1631 break; 1632 case Instruction::ShuffleVector: { 1633 auto *Shuf = dyn_cast<ShuffleVectorInst>(I); 1634 // FIXME: Do we need to handle ConstantExpr involving shufflevectors? 1635 if (!Shuf) { 1636 Known.resetAll(); 1637 return; 1638 } 1639 // For undef elements, we don't know anything about the common state of 1640 // the shuffle result. 1641 APInt DemandedLHS, DemandedRHS; 1642 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { 1643 Known.resetAll(); 1644 return; 1645 } 1646 Known.One.setAllBits(); 1647 Known.Zero.setAllBits(); 1648 if (!!DemandedLHS) { 1649 const Value *LHS = Shuf->getOperand(0); 1650 computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q); 1651 // If we don't know any bits, early out. 1652 if (Known.isUnknown()) 1653 break; 1654 } 1655 if (!!DemandedRHS) { 1656 const Value *RHS = Shuf->getOperand(1); 1657 computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q); 1658 Known = KnownBits::commonBits(Known, Known2); 1659 } 1660 break; 1661 } 1662 case Instruction::InsertElement: { 1663 const Value *Vec = I->getOperand(0); 1664 const Value *Elt = I->getOperand(1); 1665 auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2)); 1666 // Early out if the index is non-constant or out-of-range. 1667 unsigned NumElts = DemandedElts.getBitWidth(); 1668 if (!CIdx || CIdx->getValue().uge(NumElts)) { 1669 Known.resetAll(); 1670 return; 1671 } 1672 Known.One.setAllBits(); 1673 Known.Zero.setAllBits(); 1674 unsigned EltIdx = CIdx->getZExtValue(); 1675 // Do we demand the inserted element? 1676 if (DemandedElts[EltIdx]) { 1677 computeKnownBits(Elt, Known, Depth + 1, Q); 1678 // If we don't know any bits, early out. 1679 if (Known.isUnknown()) 1680 break; 1681 } 1682 // We don't need the base vector element that has been inserted. 1683 APInt DemandedVecElts = DemandedElts; 1684 DemandedVecElts.clearBit(EltIdx); 1685 if (!!DemandedVecElts) { 1686 computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q); 1687 Known = KnownBits::commonBits(Known, Known2); 1688 } 1689 break; 1690 } 1691 case Instruction::ExtractElement: { 1692 // Look through extract element. If the index is non-constant or 1693 // out-of-range demand all elements, otherwise just the extracted element. 1694 const Value *Vec = I->getOperand(0); 1695 const Value *Idx = I->getOperand(1); 1696 auto *CIdx = dyn_cast<ConstantInt>(Idx); 1697 if (isa<ScalableVectorType>(Vec->getType())) { 1698 // FIXME: there's probably *something* we can do with scalable vectors 1699 Known.resetAll(); 1700 break; 1701 } 1702 unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); 1703 APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); 1704 if (CIdx && CIdx->getValue().ult(NumElts)) 1705 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); 1706 computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q); 1707 break; 1708 } 1709 case Instruction::ExtractValue: 1710 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { 1711 const ExtractValueInst *EVI = cast<ExtractValueInst>(I); 1712 if (EVI->getNumIndices() != 1) break; 1713 if (EVI->getIndices()[0] == 0) { 1714 switch (II->getIntrinsicID()) { 1715 default: break; 1716 case Intrinsic::uadd_with_overflow: 1717 case Intrinsic::sadd_with_overflow: 1718 computeKnownBitsAddSub(true, II->getArgOperand(0), 1719 II->getArgOperand(1), false, DemandedElts, 1720 Known, Known2, Depth, Q); 1721 break; 1722 case Intrinsic::usub_with_overflow: 1723 case Intrinsic::ssub_with_overflow: 1724 computeKnownBitsAddSub(false, II->getArgOperand(0), 1725 II->getArgOperand(1), false, DemandedElts, 1726 Known, Known2, Depth, Q); 1727 break; 1728 case Intrinsic::umul_with_overflow: 1729 case Intrinsic::smul_with_overflow: 1730 computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, 1731 DemandedElts, Known, Known2, Depth, Q); 1732 break; 1733 } 1734 } 1735 } 1736 break; 1737 case Instruction::Freeze: 1738 if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, 1739 Depth + 1)) 1740 computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); 1741 break; 1742 } 1743 } 1744 1745 /// Determine which bits of V are known to be either zero or one and return 1746 /// them. 1747 KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, 1748 unsigned Depth, const Query &Q) { 1749 KnownBits Known(getBitWidth(V->getType(), Q.DL)); 1750 computeKnownBits(V, DemandedElts, Known, Depth, Q); 1751 return Known; 1752 } 1753 1754 /// Determine which bits of V are known to be either zero or one and return 1755 /// them. 1756 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) { 1757 KnownBits Known(getBitWidth(V->getType(), Q.DL)); 1758 computeKnownBits(V, Known, Depth, Q); 1759 return Known; 1760 } 1761 1762 /// Determine which bits of V are known to be either zero or one and return 1763 /// them in the Known bit set. 1764 /// 1765 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that 1766 /// we cannot optimize based on the assumption that it is zero without changing 1767 /// it to be an explicit zero. If we don't change it to zero, other code could 1768 /// optimized based on the contradictory assumption that it is non-zero. 1769 /// Because instcombine aggressively folds operations with undef args anyway, 1770 /// this won't lose us code quality. 1771 /// 1772 /// This function is defined on values with integer type, values with pointer 1773 /// type, and vectors of integers. In the case 1774 /// where V is a vector, known zero, and known one values are the 1775 /// same width as the vector element, and the bit is set only if it is true 1776 /// for all of the demanded elements in the vector specified by DemandedElts. 1777 void computeKnownBits(const Value *V, const APInt &DemandedElts, 1778 KnownBits &Known, unsigned Depth, const Query &Q) { 1779 if (!DemandedElts || isa<ScalableVectorType>(V->getType())) { 1780 // No demanded elts or V is a scalable vector, better to assume we don't 1781 // know anything. 1782 Known.resetAll(); 1783 return; 1784 } 1785 1786 assert(V && "No Value?"); 1787 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 1788 1789 #ifndef NDEBUG 1790 Type *Ty = V->getType(); 1791 unsigned BitWidth = Known.getBitWidth(); 1792 1793 assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && 1794 "Not integer or pointer type!"); 1795 1796 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { 1797 assert( 1798 FVTy->getNumElements() == DemandedElts.getBitWidth() && 1799 "DemandedElt width should equal the fixed vector number of elements"); 1800 } else { 1801 assert(DemandedElts == APInt(1, 1) && 1802 "DemandedElt width should be 1 for scalars"); 1803 } 1804 1805 Type *ScalarTy = Ty->getScalarType(); 1806 if (ScalarTy->isPointerTy()) { 1807 assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && 1808 "V and Known should have same BitWidth"); 1809 } else { 1810 assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && 1811 "V and Known should have same BitWidth"); 1812 } 1813 #endif 1814 1815 const APInt *C; 1816 if (match(V, m_APInt(C))) { 1817 // We know all of the bits for a scalar constant or a splat vector constant! 1818 Known = KnownBits::makeConstant(*C); 1819 return; 1820 } 1821 // Null and aggregate-zero are all-zeros. 1822 if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { 1823 Known.setAllZero(); 1824 return; 1825 } 1826 // Handle a constant vector by taking the intersection of the known bits of 1827 // each element. 1828 if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) { 1829 // We know that CDV must be a vector of integers. Take the intersection of 1830 // each element. 1831 Known.Zero.setAllBits(); Known.One.setAllBits(); 1832 for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { 1833 if (!DemandedElts[i]) 1834 continue; 1835 APInt Elt = CDV->getElementAsAPInt(i); 1836 Known.Zero &= ~Elt; 1837 Known.One &= Elt; 1838 } 1839 return; 1840 } 1841 1842 if (const auto *CV = dyn_cast<ConstantVector>(V)) { 1843 // We know that CV must be a vector of integers. Take the intersection of 1844 // each element. 1845 Known.Zero.setAllBits(); Known.One.setAllBits(); 1846 for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { 1847 if (!DemandedElts[i]) 1848 continue; 1849 Constant *Element = CV->getAggregateElement(i); 1850 auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); 1851 if (!ElementCI) { 1852 Known.resetAll(); 1853 return; 1854 } 1855 const APInt &Elt = ElementCI->getValue(); 1856 Known.Zero &= ~Elt; 1857 Known.One &= Elt; 1858 } 1859 return; 1860 } 1861 1862 // Start out not knowing anything. 1863 Known.resetAll(); 1864 1865 // We can't imply anything about undefs. 1866 if (isa<UndefValue>(V)) 1867 return; 1868 1869 // There's no point in looking through other users of ConstantData for 1870 // assumptions. Confirm that we've handled them all. 1871 assert(!isa<ConstantData>(V) && "Unhandled constant data!"); 1872 1873 // All recursive calls that increase depth must come after this. 1874 if (Depth == MaxAnalysisRecursionDepth) 1875 return; 1876 1877 // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has 1878 // the bits of its aliasee. 1879 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 1880 if (!GA->isInterposable()) 1881 computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q); 1882 return; 1883 } 1884 1885 if (const Operator *I = dyn_cast<Operator>(V)) 1886 computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); 1887 1888 // Aligned pointers have trailing zeros - refine Known.Zero set 1889 if (isa<PointerType>(V->getType())) { 1890 Align Alignment = V->getPointerAlignment(Q.DL); 1891 Known.Zero.setLowBits(Log2(Alignment)); 1892 } 1893 1894 // computeKnownBitsFromAssume strictly refines Known. 1895 // Therefore, we run them after computeKnownBitsFromOperator. 1896 1897 // Check whether a nearby assume intrinsic can determine some known bits. 1898 computeKnownBitsFromAssume(V, Known, Depth, Q); 1899 1900 assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?"); 1901 } 1902 1903 /// Return true if the given value is known to have exactly one 1904 /// bit set when defined. For vectors return true if every element is known to 1905 /// be a power of two when defined. Supports values with integer or pointer 1906 /// types and vectors of integers. 1907 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, 1908 const Query &Q) { 1909 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 1910 1911 // Attempt to match against constants. 1912 if (OrZero && match(V, m_Power2OrZero())) 1913 return true; 1914 if (match(V, m_Power2())) 1915 return true; 1916 1917 // 1 << X is clearly a power of two if the one is not shifted off the end. If 1918 // it is shifted off the end then the result is undefined. 1919 if (match(V, m_Shl(m_One(), m_Value()))) 1920 return true; 1921 1922 // (signmask) >>l X is clearly a power of two if the one is not shifted off 1923 // the bottom. If it is shifted off the bottom then the result is undefined. 1924 if (match(V, m_LShr(m_SignMask(), m_Value()))) 1925 return true; 1926 1927 // The remaining tests are all recursive, so bail out if we hit the limit. 1928 if (Depth++ == MaxAnalysisRecursionDepth) 1929 return false; 1930 1931 Value *X = nullptr, *Y = nullptr; 1932 // A shift left or a logical shift right of a power of two is a power of two 1933 // or zero. 1934 if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || 1935 match(V, m_LShr(m_Value(X), m_Value())))) 1936 return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); 1937 1938 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V)) 1939 return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); 1940 1941 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) 1942 return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && 1943 isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); 1944 1945 if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { 1946 // A power of two and'd with anything is a power of two or zero. 1947 if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || 1948 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) 1949 return true; 1950 // X & (-X) is always a power of two or zero. 1951 if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) 1952 return true; 1953 return false; 1954 } 1955 1956 // Adding a power-of-two or zero to the same power-of-two or zero yields 1957 // either the original power-of-two, a larger power-of-two or zero. 1958 if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 1959 const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); 1960 if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) || 1961 Q.IIQ.hasNoSignedWrap(VOBO)) { 1962 if (match(X, m_And(m_Specific(Y), m_Value())) || 1963 match(X, m_And(m_Value(), m_Specific(Y)))) 1964 if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) 1965 return true; 1966 if (match(Y, m_And(m_Specific(X), m_Value())) || 1967 match(Y, m_And(m_Value(), m_Specific(X)))) 1968 if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) 1969 return true; 1970 1971 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 1972 KnownBits LHSBits(BitWidth); 1973 computeKnownBits(X, LHSBits, Depth, Q); 1974 1975 KnownBits RHSBits(BitWidth); 1976 computeKnownBits(Y, RHSBits, Depth, Q); 1977 // If i8 V is a power of two or zero: 1978 // ZeroBits: 1 1 1 0 1 1 1 1 1979 // ~ZeroBits: 0 0 0 1 0 0 0 0 1980 if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) 1981 // If OrZero isn't set, we cannot give back a zero result. 1982 // Make sure either the LHS or RHS has a bit set. 1983 if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) 1984 return true; 1985 } 1986 } 1987 1988 // An exact divide or right shift can only shift off zero bits, so the result 1989 // is a power of two only if the first operand is a power of two and not 1990 // copying a sign bit (sdiv int_min, 2). 1991 if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || 1992 match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { 1993 return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, 1994 Depth, Q); 1995 } 1996 1997 return false; 1998 } 1999 2000 /// Test whether a GEP's result is known to be non-null. 2001 /// 2002 /// Uses properties inherent in a GEP to try to determine whether it is known 2003 /// to be non-null. 2004 /// 2005 /// Currently this routine does not support vector GEPs. 2006 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, 2007 const Query &Q) { 2008 const Function *F = nullptr; 2009 if (const Instruction *I = dyn_cast<Instruction>(GEP)) 2010 F = I->getFunction(); 2011 2012 if (!GEP->isInBounds() || 2013 NullPointerIsDefined(F, GEP->getPointerAddressSpace())) 2014 return false; 2015 2016 // FIXME: Support vector-GEPs. 2017 assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP"); 2018 2019 // If the base pointer is non-null, we cannot walk to a null address with an 2020 // inbounds GEP in address space zero. 2021 if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) 2022 return true; 2023 2024 // Walk the GEP operands and see if any operand introduces a non-zero offset. 2025 // If so, then the GEP cannot produce a null pointer, as doing so would 2026 // inherently violate the inbounds contract within address space zero. 2027 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 2028 GTI != GTE; ++GTI) { 2029 // Struct types are easy -- they must always be indexed by a constant. 2030 if (StructType *STy = GTI.getStructTypeOrNull()) { 2031 ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); 2032 unsigned ElementIdx = OpC->getZExtValue(); 2033 const StructLayout *SL = Q.DL.getStructLayout(STy); 2034 uint64_t ElementOffset = SL->getElementOffset(ElementIdx); 2035 if (ElementOffset > 0) 2036 return true; 2037 continue; 2038 } 2039 2040 // If we have a zero-sized type, the index doesn't matter. Keep looping. 2041 if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).getKnownMinSize() == 0) 2042 continue; 2043 2044 // Fast path the constant operand case both for efficiency and so we don't 2045 // increment Depth when just zipping down an all-constant GEP. 2046 if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { 2047 if (!OpC->isZero()) 2048 return true; 2049 continue; 2050 } 2051 2052 // We post-increment Depth here because while isKnownNonZero increments it 2053 // as well, when we pop back up that increment won't persist. We don't want 2054 // to recurse 10k times just because we have 10k GEP operands. We don't 2055 // bail completely out because we want to handle constant GEPs regardless 2056 // of depth. 2057 if (Depth++ >= MaxAnalysisRecursionDepth) 2058 continue; 2059 2060 if (isKnownNonZero(GTI.getOperand(), Depth, Q)) 2061 return true; 2062 } 2063 2064 return false; 2065 } 2066 2067 static bool isKnownNonNullFromDominatingCondition(const Value *V, 2068 const Instruction *CtxI, 2069 const DominatorTree *DT) { 2070 if (isa<Constant>(V)) 2071 return false; 2072 2073 if (!CtxI || !DT) 2074 return false; 2075 2076 unsigned NumUsesExplored = 0; 2077 for (auto *U : V->users()) { 2078 // Avoid massive lists 2079 if (NumUsesExplored >= DomConditionsMaxUses) 2080 break; 2081 NumUsesExplored++; 2082 2083 // If the value is used as an argument to a call or invoke, then argument 2084 // attributes may provide an answer about null-ness. 2085 if (const auto *CB = dyn_cast<CallBase>(U)) 2086 if (auto *CalledFunc = CB->getCalledFunction()) 2087 for (const Argument &Arg : CalledFunc->args()) 2088 if (CB->getArgOperand(Arg.getArgNo()) == V && 2089 Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && 2090 DT->dominates(CB, CtxI)) 2091 return true; 2092 2093 // If the value is used as a load/store, then the pointer must be non null. 2094 if (V == getLoadStorePointerOperand(U)) { 2095 const Instruction *I = cast<Instruction>(U); 2096 if (!NullPointerIsDefined(I->getFunction(), 2097 V->getType()->getPointerAddressSpace()) && 2098 DT->dominates(I, CtxI)) 2099 return true; 2100 } 2101 2102 // Consider only compare instructions uniquely controlling a branch 2103 Value *RHS; 2104 CmpInst::Predicate Pred; 2105 if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS)))) 2106 continue; 2107 2108 bool NonNullIfTrue; 2109 if (cmpExcludesZero(Pred, RHS)) 2110 NonNullIfTrue = true; 2111 else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS)) 2112 NonNullIfTrue = false; 2113 else 2114 continue; 2115 2116 SmallVector<const User *, 4> WorkList; 2117 SmallPtrSet<const User *, 4> Visited; 2118 for (auto *CmpU : U->users()) { 2119 assert(WorkList.empty() && "Should be!"); 2120 if (Visited.insert(CmpU).second) 2121 WorkList.push_back(CmpU); 2122 2123 while (!WorkList.empty()) { 2124 auto *Curr = WorkList.pop_back_val(); 2125 2126 // If a user is an AND, add all its users to the work list. We only 2127 // propagate "pred != null" condition through AND because it is only 2128 // correct to assume that all conditions of AND are met in true branch. 2129 // TODO: Support similar logic of OR and EQ predicate? 2130 if (NonNullIfTrue) 2131 if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) { 2132 for (auto *CurrU : Curr->users()) 2133 if (Visited.insert(CurrU).second) 2134 WorkList.push_back(CurrU); 2135 continue; 2136 } 2137 2138 if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) { 2139 assert(BI->isConditional() && "uses a comparison!"); 2140 2141 BasicBlock *NonNullSuccessor = 2142 BI->getSuccessor(NonNullIfTrue ? 0 : 1); 2143 BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); 2144 if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) 2145 return true; 2146 } else if (NonNullIfTrue && isGuard(Curr) && 2147 DT->dominates(cast<Instruction>(Curr), CtxI)) { 2148 return true; 2149 } 2150 } 2151 } 2152 } 2153 2154 return false; 2155 } 2156 2157 /// Does the 'Range' metadata (which must be a valid MD_range operand list) 2158 /// ensure that the value it's attached to is never Value? 'RangeType' is 2159 /// is the type of the value described by the range. 2160 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { 2161 const unsigned NumRanges = Ranges->getNumOperands() / 2; 2162 assert(NumRanges >= 1); 2163 for (unsigned i = 0; i < NumRanges; ++i) { 2164 ConstantInt *Lower = 2165 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); 2166 ConstantInt *Upper = 2167 mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); 2168 ConstantRange Range(Lower->getValue(), Upper->getValue()); 2169 if (Range.contains(Value)) 2170 return false; 2171 } 2172 return true; 2173 } 2174 2175 /// Return true if the given value is known to be non-zero when defined. For 2176 /// vectors, return true if every demanded element is known to be non-zero when 2177 /// defined. For pointers, if the context instruction and dominator tree are 2178 /// specified, perform context-sensitive analysis and return true if the 2179 /// pointer couldn't possibly be null at the specified instruction. 2180 /// Supports values with integer or pointer type and vectors of integers. 2181 bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth, 2182 const Query &Q) { 2183 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2184 // vector 2185 if (isa<ScalableVectorType>(V->getType())) 2186 return false; 2187 2188 if (auto *C = dyn_cast<Constant>(V)) { 2189 if (C->isNullValue()) 2190 return false; 2191 if (isa<ConstantInt>(C)) 2192 // Must be non-zero due to null test above. 2193 return true; 2194 2195 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 2196 // See the comment for IntToPtr/PtrToInt instructions below. 2197 if (CE->getOpcode() == Instruction::IntToPtr || 2198 CE->getOpcode() == Instruction::PtrToInt) 2199 if (Q.DL.getTypeSizeInBits(CE->getOperand(0)->getType()) 2200 .getFixedSize() <= 2201 Q.DL.getTypeSizeInBits(CE->getType()).getFixedSize()) 2202 return isKnownNonZero(CE->getOperand(0), Depth, Q); 2203 } 2204 2205 // For constant vectors, check that all elements are undefined or known 2206 // non-zero to determine that the whole vector is known non-zero. 2207 if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) { 2208 for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { 2209 if (!DemandedElts[i]) 2210 continue; 2211 Constant *Elt = C->getAggregateElement(i); 2212 if (!Elt || Elt->isNullValue()) 2213 return false; 2214 if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) 2215 return false; 2216 } 2217 return true; 2218 } 2219 2220 // A global variable in address space 0 is non null unless extern weak 2221 // or an absolute symbol reference. Other address spaces may have null as a 2222 // valid address for a global, so we can't assume anything. 2223 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 2224 if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && 2225 GV->getType()->getAddressSpace() == 0) 2226 return true; 2227 } else 2228 return false; 2229 } 2230 2231 if (auto *I = dyn_cast<Instruction>(V)) { 2232 if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) { 2233 // If the possible ranges don't contain zero, then the value is 2234 // definitely non-zero. 2235 if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { 2236 const APInt ZeroValue(Ty->getBitWidth(), 0); 2237 if (rangeMetadataExcludesValue(Ranges, ZeroValue)) 2238 return true; 2239 } 2240 } 2241 } 2242 2243 if (isKnownNonZeroFromAssume(V, Q)) 2244 return true; 2245 2246 // Some of the tests below are recursive, so bail out if we hit the limit. 2247 if (Depth++ >= MaxAnalysisRecursionDepth) 2248 return false; 2249 2250 // Check for pointer simplifications. 2251 2252 if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) { 2253 // Alloca never returns null, malloc might. 2254 if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0) 2255 return true; 2256 2257 // A byval, inalloca may not be null in a non-default addres space. A 2258 // nonnull argument is assumed never 0. 2259 if (const Argument *A = dyn_cast<Argument>(V)) { 2260 if (((A->hasPassPointeeByValueCopyAttr() && 2261 !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) || 2262 A->hasNonNullAttr())) 2263 return true; 2264 } 2265 2266 // A Load tagged with nonnull metadata is never null. 2267 if (const LoadInst *LI = dyn_cast<LoadInst>(V)) 2268 if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull)) 2269 return true; 2270 2271 if (const auto *Call = dyn_cast<CallBase>(V)) { 2272 if (Call->isReturnNonNull()) 2273 return true; 2274 if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true)) 2275 return isKnownNonZero(RP, Depth, Q); 2276 } 2277 } 2278 2279 if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT)) 2280 return true; 2281 2282 // Check for recursive pointer simplifications. 2283 if (V->getType()->isPointerTy()) { 2284 // Look through bitcast operations, GEPs, and int2ptr instructions as they 2285 // do not alter the value, or at least not the nullness property of the 2286 // value, e.g., int2ptr is allowed to zero/sign extend the value. 2287 // 2288 // Note that we have to take special care to avoid looking through 2289 // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well 2290 // as casts that can alter the value, e.g., AddrSpaceCasts. 2291 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 2292 return isGEPKnownNonNull(GEP, Depth, Q); 2293 2294 if (auto *BCO = dyn_cast<BitCastOperator>(V)) 2295 return isKnownNonZero(BCO->getOperand(0), Depth, Q); 2296 2297 if (auto *I2P = dyn_cast<IntToPtrInst>(V)) 2298 if (Q.DL.getTypeSizeInBits(I2P->getSrcTy()).getFixedSize() <= 2299 Q.DL.getTypeSizeInBits(I2P->getDestTy()).getFixedSize()) 2300 return isKnownNonZero(I2P->getOperand(0), Depth, Q); 2301 } 2302 2303 // Similar to int2ptr above, we can look through ptr2int here if the cast 2304 // is a no-op or an extend and not a truncate. 2305 if (auto *P2I = dyn_cast<PtrToIntInst>(V)) 2306 if (Q.DL.getTypeSizeInBits(P2I->getSrcTy()).getFixedSize() <= 2307 Q.DL.getTypeSizeInBits(P2I->getDestTy()).getFixedSize()) 2308 return isKnownNonZero(P2I->getOperand(0), Depth, Q); 2309 2310 unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); 2311 2312 // X | Y != 0 if X != 0 or Y != 0. 2313 Value *X = nullptr, *Y = nullptr; 2314 if (match(V, m_Or(m_Value(X), m_Value(Y)))) 2315 return isKnownNonZero(X, DemandedElts, Depth, Q) || 2316 isKnownNonZero(Y, DemandedElts, Depth, Q); 2317 2318 // ext X != 0 if X != 0. 2319 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) 2320 return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q); 2321 2322 // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined 2323 // if the lowest bit is shifted off the end. 2324 if (match(V, m_Shl(m_Value(X), m_Value(Y)))) { 2325 // shl nuw can't remove any non-zero bits. 2326 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 2327 if (Q.IIQ.hasNoUnsignedWrap(BO)) 2328 return isKnownNonZero(X, Depth, Q); 2329 2330 KnownBits Known(BitWidth); 2331 computeKnownBits(X, DemandedElts, Known, Depth, Q); 2332 if (Known.One[0]) 2333 return true; 2334 } 2335 // shr X, Y != 0 if X is negative. Note that the value of the shift is not 2336 // defined if the sign bit is shifted off the end. 2337 else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) { 2338 // shr exact can only shift out zero bits. 2339 const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); 2340 if (BO->isExact()) 2341 return isKnownNonZero(X, Depth, Q); 2342 2343 KnownBits Known = computeKnownBits(X, DemandedElts, Depth, Q); 2344 if (Known.isNegative()) 2345 return true; 2346 2347 // If the shifter operand is a constant, and all of the bits shifted 2348 // out are known to be zero, and X is known non-zero then at least one 2349 // non-zero bit must remain. 2350 if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) { 2351 auto ShiftVal = Shift->getLimitedValue(BitWidth - 1); 2352 // Is there a known one in the portion not shifted out? 2353 if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal) 2354 return true; 2355 // Are all the bits to be shifted out known zero? 2356 if (Known.countMinTrailingZeros() >= ShiftVal) 2357 return isKnownNonZero(X, DemandedElts, Depth, Q); 2358 } 2359 } 2360 // div exact can only produce a zero if the dividend is zero. 2361 else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) { 2362 return isKnownNonZero(X, DemandedElts, Depth, Q); 2363 } 2364 // X + Y. 2365 else if (match(V, m_Add(m_Value(X), m_Value(Y)))) { 2366 KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); 2367 KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); 2368 2369 // If X and Y are both non-negative (as signed values) then their sum is not 2370 // zero unless both X and Y are zero. 2371 if (XKnown.isNonNegative() && YKnown.isNonNegative()) 2372 if (isKnownNonZero(X, DemandedElts, Depth, Q) || 2373 isKnownNonZero(Y, DemandedElts, Depth, Q)) 2374 return true; 2375 2376 // If X and Y are both negative (as signed values) then their sum is not 2377 // zero unless both X and Y equal INT_MIN. 2378 if (XKnown.isNegative() && YKnown.isNegative()) { 2379 APInt Mask = APInt::getSignedMaxValue(BitWidth); 2380 // The sign bit of X is set. If some other bit is set then X is not equal 2381 // to INT_MIN. 2382 if (XKnown.One.intersects(Mask)) 2383 return true; 2384 // The sign bit of Y is set. If some other bit is set then Y is not equal 2385 // to INT_MIN. 2386 if (YKnown.One.intersects(Mask)) 2387 return true; 2388 } 2389 2390 // The sum of a non-negative number and a power of two is not zero. 2391 if (XKnown.isNonNegative() && 2392 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) 2393 return true; 2394 if (YKnown.isNonNegative() && 2395 isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) 2396 return true; 2397 } 2398 // X * Y. 2399 else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) { 2400 const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); 2401 // If X and Y are non-zero then so is X * Y as long as the multiplication 2402 // does not overflow. 2403 if ((Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) && 2404 isKnownNonZero(X, DemandedElts, Depth, Q) && 2405 isKnownNonZero(Y, DemandedElts, Depth, Q)) 2406 return true; 2407 } 2408 // (C ? X : Y) != 0 if X != 0 and Y != 0. 2409 else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 2410 if (isKnownNonZero(SI->getTrueValue(), DemandedElts, Depth, Q) && 2411 isKnownNonZero(SI->getFalseValue(), DemandedElts, Depth, Q)) 2412 return true; 2413 } 2414 // PHI 2415 else if (const PHINode *PN = dyn_cast<PHINode>(V)) { 2416 // Try and detect a recurrence that monotonically increases from a 2417 // starting value, as these are common as induction variables. 2418 if (PN->getNumIncomingValues() == 2) { 2419 Value *Start = PN->getIncomingValue(0); 2420 Value *Induction = PN->getIncomingValue(1); 2421 if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start)) 2422 std::swap(Start, Induction); 2423 if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) { 2424 if (!C->isZero() && !C->isNegative()) { 2425 ConstantInt *X; 2426 if (Q.IIQ.UseInstrInfo && 2427 (match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) || 2428 match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) && 2429 !X->isNegative()) 2430 return true; 2431 } 2432 } 2433 } 2434 // Check if all incoming values are non-zero using recursion. 2435 Query RecQ = Q; 2436 unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); 2437 return llvm::all_of(PN->operands(), [&](const Use &U) { 2438 if (U.get() == PN) 2439 return true; 2440 RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); 2441 return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ); 2442 }); 2443 } 2444 // ExtractElement 2445 else if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) { 2446 const Value *Vec = EEI->getVectorOperand(); 2447 const Value *Idx = EEI->getIndexOperand(); 2448 auto *CIdx = dyn_cast<ConstantInt>(Idx); 2449 if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { 2450 unsigned NumElts = VecTy->getNumElements(); 2451 APInt DemandedVecElts = APInt::getAllOnesValue(NumElts); 2452 if (CIdx && CIdx->getValue().ult(NumElts)) 2453 DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); 2454 return isKnownNonZero(Vec, DemandedVecElts, Depth, Q); 2455 } 2456 } 2457 // Freeze 2458 else if (const FreezeInst *FI = dyn_cast<FreezeInst>(V)) { 2459 auto *Op = FI->getOperand(0); 2460 if (isKnownNonZero(Op, Depth, Q) && 2461 isGuaranteedNotToBePoison(Op, Q.AC, Q.CxtI, Q.DT, Depth)) 2462 return true; 2463 } 2464 2465 KnownBits Known(BitWidth); 2466 computeKnownBits(V, DemandedElts, Known, Depth, Q); 2467 return Known.One != 0; 2468 } 2469 2470 bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) { 2471 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2472 // vector 2473 if (isa<ScalableVectorType>(V->getType())) 2474 return false; 2475 2476 auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); 2477 APInt DemandedElts = 2478 FVTy ? APInt::getAllOnesValue(FVTy->getNumElements()) : APInt(1, 1); 2479 return isKnownNonZero(V, DemandedElts, Depth, Q); 2480 } 2481 2482 /// Return true if V2 == V1 + X, where X is known non-zero. 2483 static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth, 2484 const Query &Q) { 2485 const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); 2486 if (!BO || BO->getOpcode() != Instruction::Add) 2487 return false; 2488 Value *Op = nullptr; 2489 if (V2 == BO->getOperand(0)) 2490 Op = BO->getOperand(1); 2491 else if (V2 == BO->getOperand(1)) 2492 Op = BO->getOperand(0); 2493 else 2494 return false; 2495 return isKnownNonZero(Op, Depth + 1, Q); 2496 } 2497 2498 2499 /// Return true if it is known that V1 != V2. 2500 static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, 2501 const Query &Q) { 2502 if (V1 == V2) 2503 return false; 2504 if (V1->getType() != V2->getType()) 2505 // We can't look through casts yet. 2506 return false; 2507 2508 if (Depth >= MaxAnalysisRecursionDepth) 2509 return false; 2510 2511 // See if we can recurse through (exactly one of) our operands. This 2512 // requires our operation be 1-to-1 and map every input value to exactly 2513 // one output value. Such an operation is invertible. 2514 auto *O1 = dyn_cast<Operator>(V1); 2515 auto *O2 = dyn_cast<Operator>(V2); 2516 if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { 2517 switch (O1->getOpcode()) { 2518 default: break; 2519 case Instruction::Add: 2520 case Instruction::Sub: 2521 // Assume operand order has been canonicalized 2522 if (O1->getOperand(0) == O2->getOperand(0)) 2523 return isKnownNonEqual(O1->getOperand(1), O2->getOperand(1), 2524 Depth + 1, Q); 2525 if (O1->getOperand(1) == O2->getOperand(1)) 2526 return isKnownNonEqual(O1->getOperand(0), O2->getOperand(0), 2527 Depth + 1, Q); 2528 break; 2529 case Instruction::Mul: { 2530 // invertible if A * B == (A * B) mod 2^N where A, and B are integers 2531 // and N is the bitwdith. The nsw case is non-obvious, but proven by 2532 // alive2: https://alive2.llvm.org/ce/z/Z6D5qK 2533 auto *OBO1 = cast<OverflowingBinaryOperator>(O1); 2534 auto *OBO2 = cast<OverflowingBinaryOperator>(O2); 2535 if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && 2536 (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) 2537 break; 2538 2539 // Assume operand order has been canonicalized 2540 if (O1->getOperand(1) == O2->getOperand(1) && 2541 isa<ConstantInt>(O1->getOperand(1)) && 2542 !cast<ConstantInt>(O1->getOperand(1))->isZero()) 2543 return isKnownNonEqual(O1->getOperand(0), O2->getOperand(0), 2544 Depth + 1, Q); 2545 break; 2546 } 2547 case Instruction::SExt: 2548 case Instruction::ZExt: 2549 if (O1->getOperand(0)->getType() == O2->getOperand(0)->getType()) 2550 return isKnownNonEqual(O1->getOperand(0), O2->getOperand(0), 2551 Depth + 1, Q); 2552 break; 2553 }; 2554 } 2555 2556 if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q)) 2557 return true; 2558 2559 if (V1->getType()->isIntOrIntVectorTy()) { 2560 // Are any known bits in V1 contradictory to known bits in V2? If V1 2561 // has a known zero where V2 has a known one, they must not be equal. 2562 KnownBits Known1 = computeKnownBits(V1, Depth, Q); 2563 KnownBits Known2 = computeKnownBits(V2, Depth, Q); 2564 2565 if (Known1.Zero.intersects(Known2.One) || 2566 Known2.Zero.intersects(Known1.One)) 2567 return true; 2568 } 2569 return false; 2570 } 2571 2572 /// Return true if 'V & Mask' is known to be zero. We use this predicate to 2573 /// simplify operations downstream. Mask is known to be zero for bits that V 2574 /// cannot have. 2575 /// 2576 /// This function is defined on values with integer type, values with pointer 2577 /// type, and vectors of integers. In the case 2578 /// where V is a vector, the mask, known zero, and known one values are the 2579 /// same width as the vector element, and the bit is set only if it is true 2580 /// for all of the elements in the vector. 2581 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, 2582 const Query &Q) { 2583 KnownBits Known(Mask.getBitWidth()); 2584 computeKnownBits(V, Known, Depth, Q); 2585 return Mask.isSubsetOf(Known.Zero); 2586 } 2587 2588 // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). 2589 // Returns the input and lower/upper bounds. 2590 static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, 2591 const APInt *&CLow, const APInt *&CHigh) { 2592 assert(isa<Operator>(Select) && 2593 cast<Operator>(Select)->getOpcode() == Instruction::Select && 2594 "Input should be a Select!"); 2595 2596 const Value *LHS = nullptr, *RHS = nullptr; 2597 SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor; 2598 if (SPF != SPF_SMAX && SPF != SPF_SMIN) 2599 return false; 2600 2601 if (!match(RHS, m_APInt(CLow))) 2602 return false; 2603 2604 const Value *LHS2 = nullptr, *RHS2 = nullptr; 2605 SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor; 2606 if (getInverseMinMaxFlavor(SPF) != SPF2) 2607 return false; 2608 2609 if (!match(RHS2, m_APInt(CHigh))) 2610 return false; 2611 2612 if (SPF == SPF_SMIN) 2613 std::swap(CLow, CHigh); 2614 2615 In = LHS2; 2616 return CLow->sle(*CHigh); 2617 } 2618 2619 /// For vector constants, loop over the elements and find the constant with the 2620 /// minimum number of sign bits. Return 0 if the value is not a vector constant 2621 /// or if any element was not analyzed; otherwise, return the count for the 2622 /// element with the minimum number of sign bits. 2623 static unsigned computeNumSignBitsVectorConstant(const Value *V, 2624 const APInt &DemandedElts, 2625 unsigned TyBits) { 2626 const auto *CV = dyn_cast<Constant>(V); 2627 if (!CV || !isa<FixedVectorType>(CV->getType())) 2628 return 0; 2629 2630 unsigned MinSignBits = TyBits; 2631 unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements(); 2632 for (unsigned i = 0; i != NumElts; ++i) { 2633 if (!DemandedElts[i]) 2634 continue; 2635 // If we find a non-ConstantInt, bail out. 2636 auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); 2637 if (!Elt) 2638 return 0; 2639 2640 MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits()); 2641 } 2642 2643 return MinSignBits; 2644 } 2645 2646 static unsigned ComputeNumSignBitsImpl(const Value *V, 2647 const APInt &DemandedElts, 2648 unsigned Depth, const Query &Q); 2649 2650 static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, 2651 unsigned Depth, const Query &Q) { 2652 unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); 2653 assert(Result > 0 && "At least one sign bit needs to be present!"); 2654 return Result; 2655 } 2656 2657 /// Return the number of times the sign bit of the register is replicated into 2658 /// the other bits. We know that at least 1 bit is always equal to the sign bit 2659 /// (itself), but other cases can give us information. For example, immediately 2660 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each 2661 /// other, so we return 3. For vectors, return the number of sign bits for the 2662 /// vector element with the minimum number of known sign bits of the demanded 2663 /// elements in the vector specified by DemandedElts. 2664 static unsigned ComputeNumSignBitsImpl(const Value *V, 2665 const APInt &DemandedElts, 2666 unsigned Depth, const Query &Q) { 2667 Type *Ty = V->getType(); 2668 2669 // FIXME: We currently have no way to represent the DemandedElts of a scalable 2670 // vector 2671 if (isa<ScalableVectorType>(Ty)) 2672 return 1; 2673 2674 #ifndef NDEBUG 2675 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 2676 2677 if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { 2678 assert( 2679 FVTy->getNumElements() == DemandedElts.getBitWidth() && 2680 "DemandedElt width should equal the fixed vector number of elements"); 2681 } else { 2682 assert(DemandedElts == APInt(1, 1) && 2683 "DemandedElt width should be 1 for scalars"); 2684 } 2685 #endif 2686 2687 // We return the minimum number of sign bits that are guaranteed to be present 2688 // in V, so for undef we have to conservatively return 1. We don't have the 2689 // same behavior for poison though -- that's a FIXME today. 2690 2691 Type *ScalarTy = Ty->getScalarType(); 2692 unsigned TyBits = ScalarTy->isPointerTy() ? 2693 Q.DL.getPointerTypeSizeInBits(ScalarTy) : 2694 Q.DL.getTypeSizeInBits(ScalarTy); 2695 2696 unsigned Tmp, Tmp2; 2697 unsigned FirstAnswer = 1; 2698 2699 // Note that ConstantInt is handled by the general computeKnownBits case 2700 // below. 2701 2702 if (Depth == MaxAnalysisRecursionDepth) 2703 return 1; 2704 2705 if (auto *U = dyn_cast<Operator>(V)) { 2706 switch (Operator::getOpcode(V)) { 2707 default: break; 2708 case Instruction::SExt: 2709 Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); 2710 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; 2711 2712 case Instruction::SDiv: { 2713 const APInt *Denominator; 2714 // sdiv X, C -> adds log(C) sign bits. 2715 if (match(U->getOperand(1), m_APInt(Denominator))) { 2716 2717 // Ignore non-positive denominator. 2718 if (!Denominator->isStrictlyPositive()) 2719 break; 2720 2721 // Calculate the incoming numerator bits. 2722 unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2723 2724 // Add floor(log(C)) bits to the numerator bits. 2725 return std::min(TyBits, NumBits + Denominator->logBase2()); 2726 } 2727 break; 2728 } 2729 2730 case Instruction::SRem: { 2731 const APInt *Denominator; 2732 // srem X, C -> we know that the result is within [-C+1,C) when C is a 2733 // positive constant. This let us put a lower bound on the number of sign 2734 // bits. 2735 if (match(U->getOperand(1), m_APInt(Denominator))) { 2736 2737 // Ignore non-positive denominator. 2738 if (!Denominator->isStrictlyPositive()) 2739 break; 2740 2741 // Calculate the incoming numerator bits. SRem by a positive constant 2742 // can't lower the number of sign bits. 2743 unsigned NumrBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2744 2745 // Calculate the leading sign bit constraints by examining the 2746 // denominator. Given that the denominator is positive, there are two 2747 // cases: 2748 // 2749 // 1. the numerator is positive. The result range is [0,C) and [0,C) u< 2750 // (1 << ceilLogBase2(C)). 2751 // 2752 // 2. the numerator is negative. Then the result range is (-C,0] and 2753 // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). 2754 // 2755 // Thus a lower bound on the number of sign bits is `TyBits - 2756 // ceilLogBase2(C)`. 2757 2758 unsigned ResBits = TyBits - Denominator->ceilLogBase2(); 2759 return std::max(NumrBits, ResBits); 2760 } 2761 break; 2762 } 2763 2764 case Instruction::AShr: { 2765 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2766 // ashr X, C -> adds C sign bits. Vectors too. 2767 const APInt *ShAmt; 2768 if (match(U->getOperand(1), m_APInt(ShAmt))) { 2769 if (ShAmt->uge(TyBits)) 2770 break; // Bad shift. 2771 unsigned ShAmtLimited = ShAmt->getZExtValue(); 2772 Tmp += ShAmtLimited; 2773 if (Tmp > TyBits) Tmp = TyBits; 2774 } 2775 return Tmp; 2776 } 2777 case Instruction::Shl: { 2778 const APInt *ShAmt; 2779 if (match(U->getOperand(1), m_APInt(ShAmt))) { 2780 // shl destroys sign bits. 2781 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2782 if (ShAmt->uge(TyBits) || // Bad shift. 2783 ShAmt->uge(Tmp)) break; // Shifted all sign bits out. 2784 Tmp2 = ShAmt->getZExtValue(); 2785 return Tmp - Tmp2; 2786 } 2787 break; 2788 } 2789 case Instruction::And: 2790 case Instruction::Or: 2791 case Instruction::Xor: // NOT is handled here. 2792 // Logical binary ops preserve the number of sign bits at the worst. 2793 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2794 if (Tmp != 1) { 2795 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2796 FirstAnswer = std::min(Tmp, Tmp2); 2797 // We computed what we know about the sign bits as our first 2798 // answer. Now proceed to the generic code that uses 2799 // computeKnownBits, and pick whichever answer is better. 2800 } 2801 break; 2802 2803 case Instruction::Select: { 2804 // If we have a clamp pattern, we know that the number of sign bits will 2805 // be the minimum of the clamp min/max range. 2806 const Value *X; 2807 const APInt *CLow, *CHigh; 2808 if (isSignedMinMaxClamp(U, X, CLow, CHigh)) 2809 return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); 2810 2811 Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2812 if (Tmp == 1) break; 2813 Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); 2814 return std::min(Tmp, Tmp2); 2815 } 2816 2817 case Instruction::Add: 2818 // Add can have at most one carry bit. Thus we know that the output 2819 // is, at worst, one more bit than the inputs. 2820 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2821 if (Tmp == 1) break; 2822 2823 // Special case decrementing a value (ADD X, -1): 2824 if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) 2825 if (CRHS->isAllOnesValue()) { 2826 KnownBits Known(TyBits); 2827 computeKnownBits(U->getOperand(0), Known, Depth + 1, Q); 2828 2829 // If the input is known to be 0 or 1, the output is 0/-1, which is 2830 // all sign bits set. 2831 if ((Known.Zero | 1).isAllOnesValue()) 2832 return TyBits; 2833 2834 // If we are subtracting one from a positive number, there is no carry 2835 // out of the result. 2836 if (Known.isNonNegative()) 2837 return Tmp; 2838 } 2839 2840 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2841 if (Tmp2 == 1) break; 2842 return std::min(Tmp, Tmp2) - 1; 2843 2844 case Instruction::Sub: 2845 Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2846 if (Tmp2 == 1) break; 2847 2848 // Handle NEG. 2849 if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) 2850 if (CLHS->isNullValue()) { 2851 KnownBits Known(TyBits); 2852 computeKnownBits(U->getOperand(1), Known, Depth + 1, Q); 2853 // If the input is known to be 0 or 1, the output is 0/-1, which is 2854 // all sign bits set. 2855 if ((Known.Zero | 1).isAllOnesValue()) 2856 return TyBits; 2857 2858 // If the input is known to be positive (the sign bit is known clear), 2859 // the output of the NEG has the same number of sign bits as the 2860 // input. 2861 if (Known.isNonNegative()) 2862 return Tmp2; 2863 2864 // Otherwise, we treat this like a SUB. 2865 } 2866 2867 // Sub can have at most one carry bit. Thus we know that the output 2868 // is, at worst, one more bit than the inputs. 2869 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2870 if (Tmp == 1) break; 2871 return std::min(Tmp, Tmp2) - 1; 2872 2873 case Instruction::Mul: { 2874 // The output of the Mul can be at most twice the valid bits in the 2875 // inputs. 2876 unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2877 if (SignBitsOp0 == 1) break; 2878 unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); 2879 if (SignBitsOp1 == 1) break; 2880 unsigned OutValidBits = 2881 (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); 2882 return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; 2883 } 2884 2885 case Instruction::PHI: { 2886 const PHINode *PN = cast<PHINode>(U); 2887 unsigned NumIncomingValues = PN->getNumIncomingValues(); 2888 // Don't analyze large in-degree PHIs. 2889 if (NumIncomingValues > 4) break; 2890 // Unreachable blocks may have zero-operand PHI nodes. 2891 if (NumIncomingValues == 0) break; 2892 2893 // Take the minimum of all incoming values. This can't infinitely loop 2894 // because of our depth threshold. 2895 Query RecQ = Q; 2896 Tmp = TyBits; 2897 for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { 2898 if (Tmp == 1) return Tmp; 2899 RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); 2900 Tmp = std::min( 2901 Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ)); 2902 } 2903 return Tmp; 2904 } 2905 2906 case Instruction::Trunc: 2907 // FIXME: it's tricky to do anything useful for this, but it is an 2908 // important case for targets like X86. 2909 break; 2910 2911 case Instruction::ExtractElement: 2912 // Look through extract element. At the moment we keep this simple and 2913 // skip tracking the specific element. But at least we might find 2914 // information valid for all elements of the vector (for example if vector 2915 // is sign extended, shifted, etc). 2916 return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2917 2918 case Instruction::ShuffleVector: { 2919 // Collect the minimum number of sign bits that are shared by every vector 2920 // element referenced by the shuffle. 2921 auto *Shuf = dyn_cast<ShuffleVectorInst>(U); 2922 if (!Shuf) { 2923 // FIXME: Add support for shufflevector constant expressions. 2924 return 1; 2925 } 2926 APInt DemandedLHS, DemandedRHS; 2927 // For undef elements, we don't know anything about the common state of 2928 // the shuffle result. 2929 if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) 2930 return 1; 2931 Tmp = std::numeric_limits<unsigned>::max(); 2932 if (!!DemandedLHS) { 2933 const Value *LHS = Shuf->getOperand(0); 2934 Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q); 2935 } 2936 // If we don't know anything, early out and try computeKnownBits 2937 // fall-back. 2938 if (Tmp == 1) 2939 break; 2940 if (!!DemandedRHS) { 2941 const Value *RHS = Shuf->getOperand(1); 2942 Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q); 2943 Tmp = std::min(Tmp, Tmp2); 2944 } 2945 // If we don't know anything, early out and try computeKnownBits 2946 // fall-back. 2947 if (Tmp == 1) 2948 break; 2949 assert(Tmp <= TyBits && "Failed to determine minimum sign bits"); 2950 return Tmp; 2951 } 2952 case Instruction::Call: { 2953 if (const auto *II = dyn_cast<IntrinsicInst>(U)) { 2954 switch (II->getIntrinsicID()) { 2955 default: break; 2956 case Intrinsic::abs: 2957 Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); 2958 if (Tmp == 1) break; 2959 2960 // Absolute value reduces number of sign bits by at most 1. 2961 return Tmp - 1; 2962 } 2963 } 2964 } 2965 } 2966 } 2967 2968 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2969 // use this information. 2970 2971 // If we can examine all elements of a vector constant successfully, we're 2972 // done (we can't do any better than that). If not, keep trying. 2973 if (unsigned VecSignBits = 2974 computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) 2975 return VecSignBits; 2976 2977 KnownBits Known(TyBits); 2978 computeKnownBits(V, DemandedElts, Known, Depth, Q); 2979 2980 // If we know that the sign bit is either zero or one, determine the number of 2981 // identical bits in the top of the input value. 2982 return std::max(FirstAnswer, Known.countMinSignBits()); 2983 } 2984 2985 /// This function computes the integer multiple of Base that equals V. 2986 /// If successful, it returns true and returns the multiple in 2987 /// Multiple. If unsuccessful, it returns false. It looks 2988 /// through SExt instructions only if LookThroughSExt is true. 2989 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, 2990 bool LookThroughSExt, unsigned Depth) { 2991 assert(V && "No Value?"); 2992 assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth"); 2993 assert(V->getType()->isIntegerTy() && "Not integer or pointer type!"); 2994 2995 Type *T = V->getType(); 2996 2997 ConstantInt *CI = dyn_cast<ConstantInt>(V); 2998 2999 if (Base == 0) 3000 return false; 3001 3002 if (Base == 1) { 3003 Multiple = V; 3004 return true; 3005 } 3006 3007 ConstantExpr *CO = dyn_cast<ConstantExpr>(V); 3008 Constant *BaseVal = ConstantInt::get(T, Base); 3009 if (CO && CO == BaseVal) { 3010 // Multiple is 1. 3011 Multiple = ConstantInt::get(T, 1); 3012 return true; 3013 } 3014 3015 if (CI && CI->getZExtValue() % Base == 0) { 3016 Multiple = ConstantInt::get(T, CI->getZExtValue() / Base); 3017 return true; 3018 } 3019 3020 if (Depth == MaxAnalysisRecursionDepth) return false; 3021 3022 Operator *I = dyn_cast<Operator>(V); 3023 if (!I) return false; 3024 3025 switch (I->getOpcode()) { 3026 default: break; 3027 case Instruction::SExt: 3028 if (!LookThroughSExt) return false; 3029 // otherwise fall through to ZExt 3030 LLVM_FALLTHROUGH; 3031 case Instruction::ZExt: 3032 return ComputeMultiple(I->getOperand(0), Base, Multiple, 3033 LookThroughSExt, Depth+1); 3034 case Instruction::Shl: 3035 case Instruction::Mul: { 3036 Value *Op0 = I->getOperand(0); 3037 Value *Op1 = I->getOperand(1); 3038 3039 if (I->getOpcode() == Instruction::Shl) { 3040 ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1); 3041 if (!Op1CI) return false; 3042 // Turn Op0 << Op1 into Op0 * 2^Op1 3043 APInt Op1Int = Op1CI->getValue(); 3044 uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1); 3045 APInt API(Op1Int.getBitWidth(), 0); 3046 API.setBit(BitToSet); 3047 Op1 = ConstantInt::get(V->getContext(), API); 3048 } 3049 3050 Value *Mul0 = nullptr; 3051 if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) { 3052 if (Constant *Op1C = dyn_cast<Constant>(Op1)) 3053 if (Constant *MulC = dyn_cast<Constant>(Mul0)) { 3054 if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() < 3055 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3056 Op1C = ConstantExpr::getZExt(Op1C, MulC->getType()); 3057 if (Op1C->getType()->getPrimitiveSizeInBits().getFixedSize() > 3058 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3059 MulC = ConstantExpr::getZExt(MulC, Op1C->getType()); 3060 3061 // V == Base * (Mul0 * Op1), so return (Mul0 * Op1) 3062 Multiple = ConstantExpr::getMul(MulC, Op1C); 3063 return true; 3064 } 3065 3066 if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0)) 3067 if (Mul0CI->getValue() == 1) { 3068 // V == Base * Op1, so return Op1 3069 Multiple = Op1; 3070 return true; 3071 } 3072 } 3073 3074 Value *Mul1 = nullptr; 3075 if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) { 3076 if (Constant *Op0C = dyn_cast<Constant>(Op0)) 3077 if (Constant *MulC = dyn_cast<Constant>(Mul1)) { 3078 if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() < 3079 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3080 Op0C = ConstantExpr::getZExt(Op0C, MulC->getType()); 3081 if (Op0C->getType()->getPrimitiveSizeInBits().getFixedSize() > 3082 MulC->getType()->getPrimitiveSizeInBits().getFixedSize()) 3083 MulC = ConstantExpr::getZExt(MulC, Op0C->getType()); 3084 3085 // V == Base * (Mul1 * Op0), so return (Mul1 * Op0) 3086 Multiple = ConstantExpr::getMul(MulC, Op0C); 3087 return true; 3088 } 3089 3090 if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1)) 3091 if (Mul1CI->getValue() == 1) { 3092 // V == Base * Op0, so return Op0 3093 Multiple = Op0; 3094 return true; 3095 } 3096 } 3097 } 3098 } 3099 3100 // We could not determine if V is a multiple of Base. 3101 return false; 3102 } 3103 3104 Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, 3105 const TargetLibraryInfo *TLI) { 3106 const Function *F = CB.getCalledFunction(); 3107 if (!F) 3108 return Intrinsic::not_intrinsic; 3109 3110 if (F->isIntrinsic()) 3111 return F->getIntrinsicID(); 3112 3113 // We are going to infer semantics of a library function based on mapping it 3114 // to an LLVM intrinsic. Check that the library function is available from 3115 // this callbase and in this environment. 3116 LibFunc Func; 3117 if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) || 3118 !CB.onlyReadsMemory()) 3119 return Intrinsic::not_intrinsic; 3120 3121 switch (Func) { 3122 default: 3123 break; 3124 case LibFunc_sin: 3125 case LibFunc_sinf: 3126 case LibFunc_sinl: 3127 return Intrinsic::sin; 3128 case LibFunc_cos: 3129 case LibFunc_cosf: 3130 case LibFunc_cosl: 3131 return Intrinsic::cos; 3132 case LibFunc_exp: 3133 case LibFunc_expf: 3134 case LibFunc_expl: 3135 return Intrinsic::exp; 3136 case LibFunc_exp2: 3137 case LibFunc_exp2f: 3138 case LibFunc_exp2l: 3139 return Intrinsic::exp2; 3140 case LibFunc_log: 3141 case LibFunc_logf: 3142 case LibFunc_logl: 3143 return Intrinsic::log; 3144 case LibFunc_log10: 3145 case LibFunc_log10f: 3146 case LibFunc_log10l: 3147 return Intrinsic::log10; 3148 case LibFunc_log2: 3149 case LibFunc_log2f: 3150 case LibFunc_log2l: 3151 return Intrinsic::log2; 3152 case LibFunc_fabs: 3153 case LibFunc_fabsf: 3154 case LibFunc_fabsl: 3155 return Intrinsic::fabs; 3156 case LibFunc_fmin: 3157 case LibFunc_fminf: 3158 case LibFunc_fminl: 3159 return Intrinsic::minnum; 3160 case LibFunc_fmax: 3161 case LibFunc_fmaxf: 3162 case LibFunc_fmaxl: 3163 return Intrinsic::maxnum; 3164 case LibFunc_copysign: 3165 case LibFunc_copysignf: 3166 case LibFunc_copysignl: 3167 return Intrinsic::copysign; 3168 case LibFunc_floor: 3169 case LibFunc_floorf: 3170 case LibFunc_floorl: 3171 return Intrinsic::floor; 3172 case LibFunc_ceil: 3173 case LibFunc_ceilf: 3174 case LibFunc_ceill: 3175 return Intrinsic::ceil; 3176 case LibFunc_trunc: 3177 case LibFunc_truncf: 3178 case LibFunc_truncl: 3179 return Intrinsic::trunc; 3180 case LibFunc_rint: 3181 case LibFunc_rintf: 3182 case LibFunc_rintl: 3183 return Intrinsic::rint; 3184 case LibFunc_nearbyint: 3185 case LibFunc_nearbyintf: 3186 case LibFunc_nearbyintl: 3187 return Intrinsic::nearbyint; 3188 case LibFunc_round: 3189 case LibFunc_roundf: 3190 case LibFunc_roundl: 3191 return Intrinsic::round; 3192 case LibFunc_roundeven: 3193 case LibFunc_roundevenf: 3194 case LibFunc_roundevenl: 3195 return Intrinsic::roundeven; 3196 case LibFunc_pow: 3197 case LibFunc_powf: 3198 case LibFunc_powl: 3199 return Intrinsic::pow; 3200 case LibFunc_sqrt: 3201 case LibFunc_sqrtf: 3202 case LibFunc_sqrtl: 3203 return Intrinsic::sqrt; 3204 } 3205 3206 return Intrinsic::not_intrinsic; 3207 } 3208 3209 /// Return true if we can prove that the specified FP value is never equal to 3210 /// -0.0. 3211 /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee 3212 /// that a value is not -0.0. It only guarantees that -0.0 may be treated 3213 /// the same as +0.0 in floating-point ops. 3214 /// 3215 /// NOTE: this function will need to be revisited when we support non-default 3216 /// rounding modes! 3217 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, 3218 unsigned Depth) { 3219 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3220 return !CFP->getValueAPF().isNegZero(); 3221 3222 if (Depth == MaxAnalysisRecursionDepth) 3223 return false; 3224 3225 auto *Op = dyn_cast<Operator>(V); 3226 if (!Op) 3227 return false; 3228 3229 // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. 3230 if (match(Op, m_FAdd(m_Value(), m_PosZeroFP()))) 3231 return true; 3232 3233 // sitofp and uitofp turn into +0.0 for zero. 3234 if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) 3235 return true; 3236 3237 if (auto *Call = dyn_cast<CallInst>(Op)) { 3238 Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI); 3239 switch (IID) { 3240 default: 3241 break; 3242 // sqrt(-0.0) = -0.0, no other negative results are possible. 3243 case Intrinsic::sqrt: 3244 case Intrinsic::canonicalize: 3245 return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); 3246 // fabs(x) != -0.0 3247 case Intrinsic::fabs: 3248 return true; 3249 } 3250 } 3251 3252 return false; 3253 } 3254 3255 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a 3256 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign 3257 /// bit despite comparing equal. 3258 static bool cannotBeOrderedLessThanZeroImpl(const Value *V, 3259 const TargetLibraryInfo *TLI, 3260 bool SignBitOnly, 3261 unsigned Depth) { 3262 // TODO: This function does not do the right thing when SignBitOnly is true 3263 // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform 3264 // which flips the sign bits of NaNs. See 3265 // https://llvm.org/bugs/show_bug.cgi?id=31702. 3266 3267 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 3268 return !CFP->getValueAPF().isNegative() || 3269 (!SignBitOnly && CFP->getValueAPF().isZero()); 3270 } 3271 3272 // Handle vector of constants. 3273 if (auto *CV = dyn_cast<Constant>(V)) { 3274 if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) { 3275 unsigned NumElts = CVFVTy->getNumElements(); 3276 for (unsigned i = 0; i != NumElts; ++i) { 3277 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i)); 3278 if (!CFP) 3279 return false; 3280 if (CFP->getValueAPF().isNegative() && 3281 (SignBitOnly || !CFP->getValueAPF().isZero())) 3282 return false; 3283 } 3284 3285 // All non-negative ConstantFPs. 3286 return true; 3287 } 3288 } 3289 3290 if (Depth == MaxAnalysisRecursionDepth) 3291 return false; 3292 3293 const Operator *I = dyn_cast<Operator>(V); 3294 if (!I) 3295 return false; 3296 3297 switch (I->getOpcode()) { 3298 default: 3299 break; 3300 // Unsigned integers are always nonnegative. 3301 case Instruction::UIToFP: 3302 return true; 3303 case Instruction::FMul: 3304 case Instruction::FDiv: 3305 // X * X is always non-negative or a NaN. 3306 // X / X is always exactly 1.0 or a NaN. 3307 if (I->getOperand(0) == I->getOperand(1) && 3308 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) 3309 return true; 3310 3311 LLVM_FALLTHROUGH; 3312 case Instruction::FAdd: 3313 case Instruction::FRem: 3314 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3315 Depth + 1) && 3316 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3317 Depth + 1); 3318 case Instruction::Select: 3319 return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3320 Depth + 1) && 3321 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, 3322 Depth + 1); 3323 case Instruction::FPExt: 3324 case Instruction::FPTrunc: 3325 // Widening/narrowing never change sign. 3326 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3327 Depth + 1); 3328 case Instruction::ExtractElement: 3329 // Look through extract element. At the moment we keep this simple and skip 3330 // tracking the specific element. But at least we might find information 3331 // valid for all elements of the vector. 3332 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3333 Depth + 1); 3334 case Instruction::Call: 3335 const auto *CI = cast<CallInst>(I); 3336 Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI); 3337 switch (IID) { 3338 default: 3339 break; 3340 case Intrinsic::maxnum: { 3341 Value *V0 = I->getOperand(0), *V1 = I->getOperand(1); 3342 auto isPositiveNum = [&](Value *V) { 3343 if (SignBitOnly) { 3344 // With SignBitOnly, this is tricky because the result of 3345 // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is 3346 // a constant strictly greater than 0.0. 3347 const APFloat *C; 3348 return match(V, m_APFloat(C)) && 3349 *C > APFloat::getZero(C->getSemantics()); 3350 } 3351 3352 // -0.0 compares equal to 0.0, so if this operand is at least -0.0, 3353 // maxnum can't be ordered-less-than-zero. 3354 return isKnownNeverNaN(V, TLI) && 3355 cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1); 3356 }; 3357 3358 // TODO: This could be improved. We could also check that neither operand 3359 // has its sign bit set (and at least 1 is not-NAN?). 3360 return isPositiveNum(V0) || isPositiveNum(V1); 3361 } 3362 3363 case Intrinsic::maximum: 3364 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3365 Depth + 1) || 3366 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3367 Depth + 1); 3368 case Intrinsic::minnum: 3369 case Intrinsic::minimum: 3370 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3371 Depth + 1) && 3372 cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, 3373 Depth + 1); 3374 case Intrinsic::exp: 3375 case Intrinsic::exp2: 3376 case Intrinsic::fabs: 3377 return true; 3378 3379 case Intrinsic::sqrt: 3380 // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0. 3381 if (!SignBitOnly) 3382 return true; 3383 return CI->hasNoNaNs() && (CI->hasNoSignedZeros() || 3384 CannotBeNegativeZero(CI->getOperand(0), TLI)); 3385 3386 case Intrinsic::powi: 3387 if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) { 3388 // powi(x,n) is non-negative if n is even. 3389 if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0) 3390 return true; 3391 } 3392 // TODO: This is not correct. Given that exp is an integer, here are the 3393 // ways that pow can return a negative value: 3394 // 3395 // pow(x, exp) --> negative if exp is odd and x is negative. 3396 // pow(-0, exp) --> -inf if exp is negative odd. 3397 // pow(-0, exp) --> -0 if exp is positive odd. 3398 // pow(-inf, exp) --> -0 if exp is negative odd. 3399 // pow(-inf, exp) --> -inf if exp is positive odd. 3400 // 3401 // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN, 3402 // but we must return false if x == -0. Unfortunately we do not currently 3403 // have a way of expressing this constraint. See details in 3404 // https://llvm.org/bugs/show_bug.cgi?id=31702. 3405 return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, 3406 Depth + 1); 3407 3408 case Intrinsic::fma: 3409 case Intrinsic::fmuladd: 3410 // x*x+y is non-negative if y is non-negative. 3411 return I->getOperand(0) == I->getOperand(1) && 3412 (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) && 3413 cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, 3414 Depth + 1); 3415 } 3416 break; 3417 } 3418 return false; 3419 } 3420 3421 bool llvm::CannotBeOrderedLessThanZero(const Value *V, 3422 const TargetLibraryInfo *TLI) { 3423 return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0); 3424 } 3425 3426 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) { 3427 return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0); 3428 } 3429 3430 bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI, 3431 unsigned Depth) { 3432 assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type"); 3433 3434 // If we're told that infinities won't happen, assume they won't. 3435 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) 3436 if (FPMathOp->hasNoInfs()) 3437 return true; 3438 3439 // Handle scalar constants. 3440 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3441 return !CFP->isInfinity(); 3442 3443 if (Depth == MaxAnalysisRecursionDepth) 3444 return false; 3445 3446 if (auto *Inst = dyn_cast<Instruction>(V)) { 3447 switch (Inst->getOpcode()) { 3448 case Instruction::Select: { 3449 return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) && 3450 isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1); 3451 } 3452 case Instruction::SIToFP: 3453 case Instruction::UIToFP: { 3454 // Get width of largest magnitude integer (remove a bit if signed). 3455 // This still works for a signed minimum value because the largest FP 3456 // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). 3457 int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits(); 3458 if (Inst->getOpcode() == Instruction::SIToFP) 3459 --IntSize; 3460 3461 // If the exponent of the largest finite FP value can hold the largest 3462 // integer, the result of the cast must be finite. 3463 Type *FPTy = Inst->getType()->getScalarType(); 3464 return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize; 3465 } 3466 default: 3467 break; 3468 } 3469 } 3470 3471 // try to handle fixed width vector constants 3472 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 3473 if (VFVTy && isa<Constant>(V)) { 3474 // For vectors, verify that each element is not infinity. 3475 unsigned NumElts = VFVTy->getNumElements(); 3476 for (unsigned i = 0; i != NumElts; ++i) { 3477 Constant *Elt = cast<Constant>(V)->getAggregateElement(i); 3478 if (!Elt) 3479 return false; 3480 if (isa<UndefValue>(Elt)) 3481 continue; 3482 auto *CElt = dyn_cast<ConstantFP>(Elt); 3483 if (!CElt || CElt->isInfinity()) 3484 return false; 3485 } 3486 // All elements were confirmed non-infinity or undefined. 3487 return true; 3488 } 3489 3490 // was not able to prove that V never contains infinity 3491 return false; 3492 } 3493 3494 bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI, 3495 unsigned Depth) { 3496 assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type"); 3497 3498 // If we're told that NaNs won't happen, assume they won't. 3499 if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) 3500 if (FPMathOp->hasNoNaNs()) 3501 return true; 3502 3503 // Handle scalar constants. 3504 if (auto *CFP = dyn_cast<ConstantFP>(V)) 3505 return !CFP->isNaN(); 3506 3507 if (Depth == MaxAnalysisRecursionDepth) 3508 return false; 3509 3510 if (auto *Inst = dyn_cast<Instruction>(V)) { 3511 switch (Inst->getOpcode()) { 3512 case Instruction::FAdd: 3513 case Instruction::FSub: 3514 // Adding positive and negative infinity produces NaN. 3515 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && 3516 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3517 (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) || 3518 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1)); 3519 3520 case Instruction::FMul: 3521 // Zero multiplied with infinity produces NaN. 3522 // FIXME: If neither side can be zero fmul never produces NaN. 3523 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && 3524 isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && 3525 isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3526 isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); 3527 3528 case Instruction::FDiv: 3529 case Instruction::FRem: 3530 // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN. 3531 return false; 3532 3533 case Instruction::Select: { 3534 return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && 3535 isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1); 3536 } 3537 case Instruction::SIToFP: 3538 case Instruction::UIToFP: 3539 return true; 3540 case Instruction::FPTrunc: 3541 case Instruction::FPExt: 3542 return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1); 3543 default: 3544 break; 3545 } 3546 } 3547 3548 if (const auto *II = dyn_cast<IntrinsicInst>(V)) { 3549 switch (II->getIntrinsicID()) { 3550 case Intrinsic::canonicalize: 3551 case Intrinsic::fabs: 3552 case Intrinsic::copysign: 3553 case Intrinsic::exp: 3554 case Intrinsic::exp2: 3555 case Intrinsic::floor: 3556 case Intrinsic::ceil: 3557 case Intrinsic::trunc: 3558 case Intrinsic::rint: 3559 case Intrinsic::nearbyint: 3560 case Intrinsic::round: 3561 case Intrinsic::roundeven: 3562 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1); 3563 case Intrinsic::sqrt: 3564 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) && 3565 CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI); 3566 case Intrinsic::minnum: 3567 case Intrinsic::maxnum: 3568 // If either operand is not NaN, the result is not NaN. 3569 return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) || 3570 isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1); 3571 default: 3572 return false; 3573 } 3574 } 3575 3576 // Try to handle fixed width vector constants 3577 auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); 3578 if (VFVTy && isa<Constant>(V)) { 3579 // For vectors, verify that each element is not NaN. 3580 unsigned NumElts = VFVTy->getNumElements(); 3581 for (unsigned i = 0; i != NumElts; ++i) { 3582 Constant *Elt = cast<Constant>(V)->getAggregateElement(i); 3583 if (!Elt) 3584 return false; 3585 if (isa<UndefValue>(Elt)) 3586 continue; 3587 auto *CElt = dyn_cast<ConstantFP>(Elt); 3588 if (!CElt || CElt->isNaN()) 3589 return false; 3590 } 3591 // All elements were confirmed not-NaN or undefined. 3592 return true; 3593 } 3594 3595 // Was not able to prove that V never contains NaN 3596 return false; 3597 } 3598 3599 Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { 3600 3601 // All byte-wide stores are splatable, even of arbitrary variables. 3602 if (V->getType()->isIntegerTy(8)) 3603 return V; 3604 3605 LLVMContext &Ctx = V->getContext(); 3606 3607 // Undef don't care. 3608 auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx)); 3609 if (isa<UndefValue>(V)) 3610 return UndefInt8; 3611 3612 // Return Undef for zero-sized type. 3613 if (!DL.getTypeStoreSize(V->getType()).isNonZero()) 3614 return UndefInt8; 3615 3616 Constant *C = dyn_cast<Constant>(V); 3617 if (!C) { 3618 // Conceptually, we could handle things like: 3619 // %a = zext i8 %X to i16 3620 // %b = shl i16 %a, 8 3621 // %c = or i16 %a, %b 3622 // but until there is an example that actually needs this, it doesn't seem 3623 // worth worrying about. 3624 return nullptr; 3625 } 3626 3627 // Handle 'null' ConstantArrayZero etc. 3628 if (C->isNullValue()) 3629 return Constant::getNullValue(Type::getInt8Ty(Ctx)); 3630 3631 // Constant floating-point values can be handled as integer values if the 3632 // corresponding integer value is "byteable". An important case is 0.0. 3633 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 3634 Type *Ty = nullptr; 3635 if (CFP->getType()->isHalfTy()) 3636 Ty = Type::getInt16Ty(Ctx); 3637 else if (CFP->getType()->isFloatTy()) 3638 Ty = Type::getInt32Ty(Ctx); 3639 else if (CFP->getType()->isDoubleTy()) 3640 Ty = Type::getInt64Ty(Ctx); 3641 // Don't handle long double formats, which have strange constraints. 3642 return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL) 3643 : nullptr; 3644 } 3645 3646 // We can handle constant integers that are multiple of 8 bits. 3647 if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { 3648 if (CI->getBitWidth() % 8 == 0) { 3649 assert(CI->getBitWidth() > 8 && "8 bits should be handled above!"); 3650 if (!CI->getValue().isSplat(8)) 3651 return nullptr; 3652 return ConstantInt::get(Ctx, CI->getValue().trunc(8)); 3653 } 3654 } 3655 3656 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 3657 if (CE->getOpcode() == Instruction::IntToPtr) { 3658 if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) { 3659 unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace()); 3660 return isBytewiseValue( 3661 ConstantExpr::getIntegerCast(CE->getOperand(0), 3662 Type::getIntNTy(Ctx, BitWidth), false), 3663 DL); 3664 } 3665 } 3666 } 3667 3668 auto Merge = [&](Value *LHS, Value *RHS) -> Value * { 3669 if (LHS == RHS) 3670 return LHS; 3671 if (!LHS || !RHS) 3672 return nullptr; 3673 if (LHS == UndefInt8) 3674 return RHS; 3675 if (RHS == UndefInt8) 3676 return LHS; 3677 return nullptr; 3678 }; 3679 3680 if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) { 3681 Value *Val = UndefInt8; 3682 for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) 3683 if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL)))) 3684 return nullptr; 3685 return Val; 3686 } 3687 3688 if (isa<ConstantAggregate>(C)) { 3689 Value *Val = UndefInt8; 3690 for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) 3691 if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL)))) 3692 return nullptr; 3693 return Val; 3694 } 3695 3696 // Don't try to handle the handful of other constants. 3697 return nullptr; 3698 } 3699 3700 // This is the recursive version of BuildSubAggregate. It takes a few different 3701 // arguments. Idxs is the index within the nested struct From that we are 3702 // looking at now (which is of type IndexedType). IdxSkip is the number of 3703 // indices from Idxs that should be left out when inserting into the resulting 3704 // struct. To is the result struct built so far, new insertvalue instructions 3705 // build on that. 3706 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, 3707 SmallVectorImpl<unsigned> &Idxs, 3708 unsigned IdxSkip, 3709 Instruction *InsertBefore) { 3710 StructType *STy = dyn_cast<StructType>(IndexedType); 3711 if (STy) { 3712 // Save the original To argument so we can modify it 3713 Value *OrigTo = To; 3714 // General case, the type indexed by Idxs is a struct 3715 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3716 // Process each struct element recursively 3717 Idxs.push_back(i); 3718 Value *PrevTo = To; 3719 To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, 3720 InsertBefore); 3721 Idxs.pop_back(); 3722 if (!To) { 3723 // Couldn't find any inserted value for this index? Cleanup 3724 while (PrevTo != OrigTo) { 3725 InsertValueInst* Del = cast<InsertValueInst>(PrevTo); 3726 PrevTo = Del->getAggregateOperand(); 3727 Del->eraseFromParent(); 3728 } 3729 // Stop processing elements 3730 break; 3731 } 3732 } 3733 // If we successfully found a value for each of our subaggregates 3734 if (To) 3735 return To; 3736 } 3737 // Base case, the type indexed by SourceIdxs is not a struct, or not all of 3738 // the struct's elements had a value that was inserted directly. In the latter 3739 // case, perhaps we can't determine each of the subelements individually, but 3740 // we might be able to find the complete struct somewhere. 3741 3742 // Find the value that is at that particular spot 3743 Value *V = FindInsertedValue(From, Idxs); 3744 3745 if (!V) 3746 return nullptr; 3747 3748 // Insert the value in the new (sub) aggregate 3749 return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip), 3750 "tmp", InsertBefore); 3751 } 3752 3753 // This helper takes a nested struct and extracts a part of it (which is again a 3754 // struct) into a new value. For example, given the struct: 3755 // { a, { b, { c, d }, e } } 3756 // and the indices "1, 1" this returns 3757 // { c, d }. 3758 // 3759 // It does this by inserting an insertvalue for each element in the resulting 3760 // struct, as opposed to just inserting a single struct. This will only work if 3761 // each of the elements of the substruct are known (ie, inserted into From by an 3762 // insertvalue instruction somewhere). 3763 // 3764 // All inserted insertvalue instructions are inserted before InsertBefore 3765 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, 3766 Instruction *InsertBefore) { 3767 assert(InsertBefore && "Must have someplace to insert!"); 3768 Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), 3769 idx_range); 3770 Value *To = UndefValue::get(IndexedType); 3771 SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); 3772 unsigned IdxSkip = Idxs.size(); 3773 3774 return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); 3775 } 3776 3777 /// Given an aggregate and a sequence of indices, see if the scalar value 3778 /// indexed is already around as a register, for example if it was inserted 3779 /// directly into the aggregate. 3780 /// 3781 /// If InsertBefore is not null, this function will duplicate (modified) 3782 /// insertvalues when a part of a nested struct is extracted. 3783 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, 3784 Instruction *InsertBefore) { 3785 // Nothing to index? Just return V then (this is useful at the end of our 3786 // recursion). 3787 if (idx_range.empty()) 3788 return V; 3789 // We have indices, so V should have an indexable type. 3790 assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && 3791 "Not looking at a struct or array?"); 3792 assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && 3793 "Invalid indices for type?"); 3794 3795 if (Constant *C = dyn_cast<Constant>(V)) { 3796 C = C->getAggregateElement(idx_range[0]); 3797 if (!C) return nullptr; 3798 return FindInsertedValue(C, idx_range.slice(1), InsertBefore); 3799 } 3800 3801 if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { 3802 // Loop the indices for the insertvalue instruction in parallel with the 3803 // requested indices 3804 const unsigned *req_idx = idx_range.begin(); 3805 for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); 3806 i != e; ++i, ++req_idx) { 3807 if (req_idx == idx_range.end()) { 3808 // We can't handle this without inserting insertvalues 3809 if (!InsertBefore) 3810 return nullptr; 3811 3812 // The requested index identifies a part of a nested aggregate. Handle 3813 // this specially. For example, 3814 // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 3815 // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 3816 // %C = extractvalue {i32, { i32, i32 } } %B, 1 3817 // This can be changed into 3818 // %A = insertvalue {i32, i32 } undef, i32 10, 0 3819 // %C = insertvalue {i32, i32 } %A, i32 11, 1 3820 // which allows the unused 0,0 element from the nested struct to be 3821 // removed. 3822 return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx), 3823 InsertBefore); 3824 } 3825 3826 // This insert value inserts something else than what we are looking for. 3827 // See if the (aggregate) value inserted into has the value we are 3828 // looking for, then. 3829 if (*req_idx != *i) 3830 return FindInsertedValue(I->getAggregateOperand(), idx_range, 3831 InsertBefore); 3832 } 3833 // If we end up here, the indices of the insertvalue match with those 3834 // requested (though possibly only partially). Now we recursively look at 3835 // the inserted value, passing any remaining indices. 3836 return FindInsertedValue(I->getInsertedValueOperand(), 3837 makeArrayRef(req_idx, idx_range.end()), 3838 InsertBefore); 3839 } 3840 3841 if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { 3842 // If we're extracting a value from an aggregate that was extracted from 3843 // something else, we can extract from that something else directly instead. 3844 // However, we will need to chain I's indices with the requested indices. 3845 3846 // Calculate the number of indices required 3847 unsigned size = I->getNumIndices() + idx_range.size(); 3848 // Allocate some space to put the new indices in 3849 SmallVector<unsigned, 5> Idxs; 3850 Idxs.reserve(size); 3851 // Add indices from the extract value instruction 3852 Idxs.append(I->idx_begin(), I->idx_end()); 3853 3854 // Add requested indices 3855 Idxs.append(idx_range.begin(), idx_range.end()); 3856 3857 assert(Idxs.size() == size 3858 && "Number of indices added not correct?"); 3859 3860 return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); 3861 } 3862 // Otherwise, we don't know (such as, extracting from a function return value 3863 // or load instruction) 3864 return nullptr; 3865 } 3866 3867 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, 3868 unsigned CharSize) { 3869 // Make sure the GEP has exactly three arguments. 3870 if (GEP->getNumOperands() != 3) 3871 return false; 3872 3873 // Make sure the index-ee is a pointer to array of \p CharSize integers. 3874 // CharSize. 3875 ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); 3876 if (!AT || !AT->getElementType()->isIntegerTy(CharSize)) 3877 return false; 3878 3879 // Check to make sure that the first operand of the GEP is an integer and 3880 // has value 0 so that we are sure we're indexing into the initializer. 3881 const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); 3882 if (!FirstIdx || !FirstIdx->isZero()) 3883 return false; 3884 3885 return true; 3886 } 3887 3888 bool llvm::getConstantDataArrayInfo(const Value *V, 3889 ConstantDataArraySlice &Slice, 3890 unsigned ElementSize, uint64_t Offset) { 3891 assert(V); 3892 3893 // Look through bitcast instructions and geps. 3894 V = V->stripPointerCasts(); 3895 3896 // If the value is a GEP instruction or constant expression, treat it as an 3897 // offset. 3898 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 3899 // The GEP operator should be based on a pointer to string constant, and is 3900 // indexing into the string constant. 3901 if (!isGEPBasedOnPointerToString(GEP, ElementSize)) 3902 return false; 3903 3904 // If the second index isn't a ConstantInt, then this is a variable index 3905 // into the array. If this occurs, we can't say anything meaningful about 3906 // the string. 3907 uint64_t StartIdx = 0; 3908 if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2))) 3909 StartIdx = CI->getZExtValue(); 3910 else 3911 return false; 3912 return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize, 3913 StartIdx + Offset); 3914 } 3915 3916 // The GEP instruction, constant or instruction, must reference a global 3917 // variable that is a constant and is initialized. The referenced constant 3918 // initializer is the array that we'll use for optimization. 3919 const GlobalVariable *GV = dyn_cast<GlobalVariable>(V); 3920 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) 3921 return false; 3922 3923 const ConstantDataArray *Array; 3924 ArrayType *ArrayTy; 3925 if (GV->getInitializer()->isNullValue()) { 3926 Type *GVTy = GV->getValueType(); 3927 if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) { 3928 // A zeroinitializer for the array; there is no ConstantDataArray. 3929 Array = nullptr; 3930 } else { 3931 const DataLayout &DL = GV->getParent()->getDataLayout(); 3932 uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedSize(); 3933 uint64_t Length = SizeInBytes / (ElementSize / 8); 3934 if (Length <= Offset) 3935 return false; 3936 3937 Slice.Array = nullptr; 3938 Slice.Offset = 0; 3939 Slice.Length = Length - Offset; 3940 return true; 3941 } 3942 } else { 3943 // This must be a ConstantDataArray. 3944 Array = dyn_cast<ConstantDataArray>(GV->getInitializer()); 3945 if (!Array) 3946 return false; 3947 ArrayTy = Array->getType(); 3948 } 3949 if (!ArrayTy->getElementType()->isIntegerTy(ElementSize)) 3950 return false; 3951 3952 uint64_t NumElts = ArrayTy->getArrayNumElements(); 3953 if (Offset > NumElts) 3954 return false; 3955 3956 Slice.Array = Array; 3957 Slice.Offset = Offset; 3958 Slice.Length = NumElts - Offset; 3959 return true; 3960 } 3961 3962 /// This function computes the length of a null-terminated C string pointed to 3963 /// by V. If successful, it returns true and returns the string in Str. 3964 /// If unsuccessful, it returns false. 3965 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, 3966 uint64_t Offset, bool TrimAtNul) { 3967 ConstantDataArraySlice Slice; 3968 if (!getConstantDataArrayInfo(V, Slice, 8, Offset)) 3969 return false; 3970 3971 if (Slice.Array == nullptr) { 3972 if (TrimAtNul) { 3973 Str = StringRef(); 3974 return true; 3975 } 3976 if (Slice.Length == 1) { 3977 Str = StringRef("", 1); 3978 return true; 3979 } 3980 // We cannot instantiate a StringRef as we do not have an appropriate string 3981 // of 0s at hand. 3982 return false; 3983 } 3984 3985 // Start out with the entire array in the StringRef. 3986 Str = Slice.Array->getAsString(); 3987 // Skip over 'offset' bytes. 3988 Str = Str.substr(Slice.Offset); 3989 3990 if (TrimAtNul) { 3991 // Trim off the \0 and anything after it. If the array is not nul 3992 // terminated, we just return the whole end of string. The client may know 3993 // some other way that the string is length-bound. 3994 Str = Str.substr(0, Str.find('\0')); 3995 } 3996 return true; 3997 } 3998 3999 // These next two are very similar to the above, but also look through PHI 4000 // nodes. 4001 // TODO: See if we can integrate these two together. 4002 4003 /// If we can compute the length of the string pointed to by 4004 /// the specified pointer, return 'len+1'. If we can't, return 0. 4005 static uint64_t GetStringLengthH(const Value *V, 4006 SmallPtrSetImpl<const PHINode*> &PHIs, 4007 unsigned CharSize) { 4008 // Look through noop bitcast instructions. 4009 V = V->stripPointerCasts(); 4010 4011 // If this is a PHI node, there are two cases: either we have already seen it 4012 // or we haven't. 4013 if (const PHINode *PN = dyn_cast<PHINode>(V)) { 4014 if (!PHIs.insert(PN).second) 4015 return ~0ULL; // already in the set. 4016 4017 // If it was new, see if all the input strings are the same length. 4018 uint64_t LenSoFar = ~0ULL; 4019 for (Value *IncValue : PN->incoming_values()) { 4020 uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize); 4021 if (Len == 0) return 0; // Unknown length -> unknown. 4022 4023 if (Len == ~0ULL) continue; 4024 4025 if (Len != LenSoFar && LenSoFar != ~0ULL) 4026 return 0; // Disagree -> unknown. 4027 LenSoFar = Len; 4028 } 4029 4030 // Success, all agree. 4031 return LenSoFar; 4032 } 4033 4034 // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) 4035 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 4036 uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize); 4037 if (Len1 == 0) return 0; 4038 uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize); 4039 if (Len2 == 0) return 0; 4040 if (Len1 == ~0ULL) return Len2; 4041 if (Len2 == ~0ULL) return Len1; 4042 if (Len1 != Len2) return 0; 4043 return Len1; 4044 } 4045 4046 // Otherwise, see if we can read the string. 4047 ConstantDataArraySlice Slice; 4048 if (!getConstantDataArrayInfo(V, Slice, CharSize)) 4049 return 0; 4050 4051 if (Slice.Array == nullptr) 4052 return 1; 4053 4054 // Search for nul characters 4055 unsigned NullIndex = 0; 4056 for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { 4057 if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0) 4058 break; 4059 } 4060 4061 return NullIndex + 1; 4062 } 4063 4064 /// If we can compute the length of the string pointed to by 4065 /// the specified pointer, return 'len+1'. If we can't, return 0. 4066 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { 4067 if (!V->getType()->isPointerTy()) 4068 return 0; 4069 4070 SmallPtrSet<const PHINode*, 32> PHIs; 4071 uint64_t Len = GetStringLengthH(V, PHIs, CharSize); 4072 // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return 4073 // an empty string as a length. 4074 return Len == ~0ULL ? 1 : Len; 4075 } 4076 4077 const Value * 4078 llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, 4079 bool MustPreserveNullness) { 4080 assert(Call && 4081 "getArgumentAliasingToReturnedPointer only works on nonnull calls"); 4082 if (const Value *RV = Call->getReturnedArgOperand()) 4083 return RV; 4084 // This can be used only as a aliasing property. 4085 if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( 4086 Call, MustPreserveNullness)) 4087 return Call->getArgOperand(0); 4088 return nullptr; 4089 } 4090 4091 bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( 4092 const CallBase *Call, bool MustPreserveNullness) { 4093 switch (Call->getIntrinsicID()) { 4094 case Intrinsic::launder_invariant_group: 4095 case Intrinsic::strip_invariant_group: 4096 case Intrinsic::aarch64_irg: 4097 case Intrinsic::aarch64_tagp: 4098 return true; 4099 case Intrinsic::ptrmask: 4100 return !MustPreserveNullness; 4101 default: 4102 return false; 4103 } 4104 } 4105 4106 /// \p PN defines a loop-variant pointer to an object. Check if the 4107 /// previous iteration of the loop was referring to the same object as \p PN. 4108 static bool isSameUnderlyingObjectInLoop(const PHINode *PN, 4109 const LoopInfo *LI) { 4110 // Find the loop-defined value. 4111 Loop *L = LI->getLoopFor(PN->getParent()); 4112 if (PN->getNumIncomingValues() != 2) 4113 return true; 4114 4115 // Find the value from previous iteration. 4116 auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); 4117 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 4118 PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); 4119 if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) 4120 return true; 4121 4122 // If a new pointer is loaded in the loop, the pointer references a different 4123 // object in every iteration. E.g.: 4124 // for (i) 4125 // int *p = a[i]; 4126 // ... 4127 if (auto *Load = dyn_cast<LoadInst>(PrevValue)) 4128 if (!L->isLoopInvariant(Load->getPointerOperand())) 4129 return false; 4130 return true; 4131 } 4132 4133 Value *llvm::getUnderlyingObject(Value *V, unsigned MaxLookup) { 4134 if (!V->getType()->isPointerTy()) 4135 return V; 4136 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { 4137 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 4138 V = GEP->getPointerOperand(); 4139 } else if (Operator::getOpcode(V) == Instruction::BitCast || 4140 Operator::getOpcode(V) == Instruction::AddrSpaceCast) { 4141 V = cast<Operator>(V)->getOperand(0); 4142 if (!V->getType()->isPointerTy()) 4143 return V; 4144 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 4145 if (GA->isInterposable()) 4146 return V; 4147 V = GA->getAliasee(); 4148 } else { 4149 if (auto *PHI = dyn_cast<PHINode>(V)) { 4150 // Look through single-arg phi nodes created by LCSSA. 4151 if (PHI->getNumIncomingValues() == 1) { 4152 V = PHI->getIncomingValue(0); 4153 continue; 4154 } 4155 } else if (auto *Call = dyn_cast<CallBase>(V)) { 4156 // CaptureTracking can know about special capturing properties of some 4157 // intrinsics like launder.invariant.group, that can't be expressed with 4158 // the attributes, but have properties like returning aliasing pointer. 4159 // Because some analysis may assume that nocaptured pointer is not 4160 // returned from some special intrinsic (because function would have to 4161 // be marked with returns attribute), it is crucial to use this function 4162 // because it should be in sync with CaptureTracking. Not using it may 4163 // cause weird miscompilations where 2 aliasing pointers are assumed to 4164 // noalias. 4165 if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) { 4166 V = RP; 4167 continue; 4168 } 4169 } 4170 4171 return V; 4172 } 4173 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 4174 } 4175 return V; 4176 } 4177 4178 void llvm::getUnderlyingObjects(const Value *V, 4179 SmallVectorImpl<const Value *> &Objects, 4180 LoopInfo *LI, unsigned MaxLookup) { 4181 SmallPtrSet<const Value *, 4> Visited; 4182 SmallVector<const Value *, 4> Worklist; 4183 Worklist.push_back(V); 4184 do { 4185 const Value *P = Worklist.pop_back_val(); 4186 P = getUnderlyingObject(P, MaxLookup); 4187 4188 if (!Visited.insert(P).second) 4189 continue; 4190 4191 if (auto *SI = dyn_cast<SelectInst>(P)) { 4192 Worklist.push_back(SI->getTrueValue()); 4193 Worklist.push_back(SI->getFalseValue()); 4194 continue; 4195 } 4196 4197 if (auto *PN = dyn_cast<PHINode>(P)) { 4198 // If this PHI changes the underlying object in every iteration of the 4199 // loop, don't look through it. Consider: 4200 // int **A; 4201 // for (i) { 4202 // Prev = Curr; // Prev = PHI (Prev_0, Curr) 4203 // Curr = A[i]; 4204 // *Prev, *Curr; 4205 // 4206 // Prev is tracking Curr one iteration behind so they refer to different 4207 // underlying objects. 4208 if (!LI || !LI->isLoopHeader(PN->getParent()) || 4209 isSameUnderlyingObjectInLoop(PN, LI)) 4210 append_range(Worklist, PN->incoming_values()); 4211 continue; 4212 } 4213 4214 Objects.push_back(P); 4215 } while (!Worklist.empty()); 4216 } 4217 4218 /// This is the function that does the work of looking through basic 4219 /// ptrtoint+arithmetic+inttoptr sequences. 4220 static const Value *getUnderlyingObjectFromInt(const Value *V) { 4221 do { 4222 if (const Operator *U = dyn_cast<Operator>(V)) { 4223 // If we find a ptrtoint, we can transfer control back to the 4224 // regular getUnderlyingObjectFromInt. 4225 if (U->getOpcode() == Instruction::PtrToInt) 4226 return U->getOperand(0); 4227 // If we find an add of a constant, a multiplied value, or a phi, it's 4228 // likely that the other operand will lead us to the base 4229 // object. We don't have to worry about the case where the 4230 // object address is somehow being computed by the multiply, 4231 // because our callers only care when the result is an 4232 // identifiable object. 4233 if (U->getOpcode() != Instruction::Add || 4234 (!isa<ConstantInt>(U->getOperand(1)) && 4235 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul && 4236 !isa<PHINode>(U->getOperand(1)))) 4237 return V; 4238 V = U->getOperand(0); 4239 } else { 4240 return V; 4241 } 4242 assert(V->getType()->isIntegerTy() && "Unexpected operand type!"); 4243 } while (true); 4244 } 4245 4246 /// This is a wrapper around getUnderlyingObjects and adds support for basic 4247 /// ptrtoint+arithmetic+inttoptr sequences. 4248 /// It returns false if unidentified object is found in getUnderlyingObjects. 4249 bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, 4250 SmallVectorImpl<Value *> &Objects) { 4251 SmallPtrSet<const Value *, 16> Visited; 4252 SmallVector<const Value *, 4> Working(1, V); 4253 do { 4254 V = Working.pop_back_val(); 4255 4256 SmallVector<const Value *, 4> Objs; 4257 getUnderlyingObjects(V, Objs); 4258 4259 for (const Value *V : Objs) { 4260 if (!Visited.insert(V).second) 4261 continue; 4262 if (Operator::getOpcode(V) == Instruction::IntToPtr) { 4263 const Value *O = 4264 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); 4265 if (O->getType()->isPointerTy()) { 4266 Working.push_back(O); 4267 continue; 4268 } 4269 } 4270 // If getUnderlyingObjects fails to find an identifiable object, 4271 // getUnderlyingObjectsForCodeGen also fails for safety. 4272 if (!isIdentifiedObject(V)) { 4273 Objects.clear(); 4274 return false; 4275 } 4276 Objects.push_back(const_cast<Value *>(V)); 4277 } 4278 } while (!Working.empty()); 4279 return true; 4280 } 4281 4282 AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { 4283 AllocaInst *Result = nullptr; 4284 SmallPtrSet<Value *, 4> Visited; 4285 SmallVector<Value *, 4> Worklist; 4286 4287 auto AddWork = [&](Value *V) { 4288 if (Visited.insert(V).second) 4289 Worklist.push_back(V); 4290 }; 4291 4292 AddWork(V); 4293 do { 4294 V = Worklist.pop_back_val(); 4295 assert(Visited.count(V)); 4296 4297 if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 4298 if (Result && Result != AI) 4299 return nullptr; 4300 Result = AI; 4301 } else if (CastInst *CI = dyn_cast<CastInst>(V)) { 4302 AddWork(CI->getOperand(0)); 4303 } else if (PHINode *PN = dyn_cast<PHINode>(V)) { 4304 for (Value *IncValue : PN->incoming_values()) 4305 AddWork(IncValue); 4306 } else if (auto *SI = dyn_cast<SelectInst>(V)) { 4307 AddWork(SI->getTrueValue()); 4308 AddWork(SI->getFalseValue()); 4309 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { 4310 if (OffsetZero && !GEP->hasAllZeroIndices()) 4311 return nullptr; 4312 AddWork(GEP->getPointerOperand()); 4313 } else { 4314 return nullptr; 4315 } 4316 } while (!Worklist.empty()); 4317 4318 return Result; 4319 } 4320 4321 static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4322 const Value *V, bool AllowLifetime, bool AllowDroppable) { 4323 for (const User *U : V->users()) { 4324 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 4325 if (!II) 4326 return false; 4327 4328 if (AllowLifetime && II->isLifetimeStartOrEnd()) 4329 continue; 4330 4331 if (AllowDroppable && II->isDroppable()) 4332 continue; 4333 4334 return false; 4335 } 4336 return true; 4337 } 4338 4339 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { 4340 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4341 V, /* AllowLifetime */ true, /* AllowDroppable */ false); 4342 } 4343 bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { 4344 return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( 4345 V, /* AllowLifetime */ true, /* AllowDroppable */ true); 4346 } 4347 4348 bool llvm::mustSuppressSpeculation(const LoadInst &LI) { 4349 if (!LI.isUnordered()) 4350 return true; 4351 const Function &F = *LI.getFunction(); 4352 // Speculative load may create a race that did not exist in the source. 4353 return F.hasFnAttribute(Attribute::SanitizeThread) || 4354 // Speculative load may load data from dirty regions. 4355 F.hasFnAttribute(Attribute::SanitizeAddress) || 4356 F.hasFnAttribute(Attribute::SanitizeHWAddress); 4357 } 4358 4359 4360 bool llvm::isSafeToSpeculativelyExecute(const Value *V, 4361 const Instruction *CtxI, 4362 const DominatorTree *DT) { 4363 const Operator *Inst = dyn_cast<Operator>(V); 4364 if (!Inst) 4365 return false; 4366 4367 for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i) 4368 if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i))) 4369 if (C->canTrap()) 4370 return false; 4371 4372 switch (Inst->getOpcode()) { 4373 default: 4374 return true; 4375 case Instruction::UDiv: 4376 case Instruction::URem: { 4377 // x / y is undefined if y == 0. 4378 const APInt *V; 4379 if (match(Inst->getOperand(1), m_APInt(V))) 4380 return *V != 0; 4381 return false; 4382 } 4383 case Instruction::SDiv: 4384 case Instruction::SRem: { 4385 // x / y is undefined if y == 0 or x == INT_MIN and y == -1 4386 const APInt *Numerator, *Denominator; 4387 if (!match(Inst->getOperand(1), m_APInt(Denominator))) 4388 return false; 4389 // We cannot hoist this division if the denominator is 0. 4390 if (*Denominator == 0) 4391 return false; 4392 // It's safe to hoist if the denominator is not 0 or -1. 4393 if (!Denominator->isAllOnesValue()) 4394 return true; 4395 // At this point we know that the denominator is -1. It is safe to hoist as 4396 // long we know that the numerator is not INT_MIN. 4397 if (match(Inst->getOperand(0), m_APInt(Numerator))) 4398 return !Numerator->isMinSignedValue(); 4399 // The numerator *might* be MinSignedValue. 4400 return false; 4401 } 4402 case Instruction::Load: { 4403 const LoadInst *LI = cast<LoadInst>(Inst); 4404 if (mustSuppressSpeculation(*LI)) 4405 return false; 4406 const DataLayout &DL = LI->getModule()->getDataLayout(); 4407 return isDereferenceableAndAlignedPointer( 4408 LI->getPointerOperand(), LI->getType(), MaybeAlign(LI->getAlignment()), 4409 DL, CtxI, DT); 4410 } 4411 case Instruction::Call: { 4412 auto *CI = cast<const CallInst>(Inst); 4413 const Function *Callee = CI->getCalledFunction(); 4414 4415 // The called function could have undefined behavior or side-effects, even 4416 // if marked readnone nounwind. 4417 return Callee && Callee->isSpeculatable(); 4418 } 4419 case Instruction::VAArg: 4420 case Instruction::Alloca: 4421 case Instruction::Invoke: 4422 case Instruction::CallBr: 4423 case Instruction::PHI: 4424 case Instruction::Store: 4425 case Instruction::Ret: 4426 case Instruction::Br: 4427 case Instruction::IndirectBr: 4428 case Instruction::Switch: 4429 case Instruction::Unreachable: 4430 case Instruction::Fence: 4431 case Instruction::AtomicRMW: 4432 case Instruction::AtomicCmpXchg: 4433 case Instruction::LandingPad: 4434 case Instruction::Resume: 4435 case Instruction::CatchSwitch: 4436 case Instruction::CatchPad: 4437 case Instruction::CatchRet: 4438 case Instruction::CleanupPad: 4439 case Instruction::CleanupRet: 4440 return false; // Misc instructions which have effects 4441 } 4442 } 4443 4444 bool llvm::mayBeMemoryDependent(const Instruction &I) { 4445 return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I); 4446 } 4447 4448 /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. 4449 static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { 4450 switch (OR) { 4451 case ConstantRange::OverflowResult::MayOverflow: 4452 return OverflowResult::MayOverflow; 4453 case ConstantRange::OverflowResult::AlwaysOverflowsLow: 4454 return OverflowResult::AlwaysOverflowsLow; 4455 case ConstantRange::OverflowResult::AlwaysOverflowsHigh: 4456 return OverflowResult::AlwaysOverflowsHigh; 4457 case ConstantRange::OverflowResult::NeverOverflows: 4458 return OverflowResult::NeverOverflows; 4459 } 4460 llvm_unreachable("Unknown OverflowResult"); 4461 } 4462 4463 /// Combine constant ranges from computeConstantRange() and computeKnownBits(). 4464 static ConstantRange computeConstantRangeIncludingKnownBits( 4465 const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth, 4466 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4467 OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) { 4468 KnownBits Known = computeKnownBits( 4469 V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo); 4470 ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned); 4471 ConstantRange CR2 = computeConstantRange(V, UseInstrInfo); 4472 ConstantRange::PreferredRangeType RangeType = 4473 ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; 4474 return CR1.intersectWith(CR2, RangeType); 4475 } 4476 4477 OverflowResult llvm::computeOverflowForUnsignedMul( 4478 const Value *LHS, const Value *RHS, const DataLayout &DL, 4479 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4480 bool UseInstrInfo) { 4481 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, 4482 nullptr, UseInstrInfo); 4483 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, 4484 nullptr, UseInstrInfo); 4485 ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false); 4486 ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false); 4487 return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange)); 4488 } 4489 4490 OverflowResult 4491 llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS, 4492 const DataLayout &DL, AssumptionCache *AC, 4493 const Instruction *CxtI, 4494 const DominatorTree *DT, bool UseInstrInfo) { 4495 // Multiplying n * m significant bits yields a result of n + m significant 4496 // bits. If the total number of significant bits does not exceed the 4497 // result bit width (minus 1), there is no overflow. 4498 // This means if we have enough leading sign bits in the operands 4499 // we can guarantee that the result does not overflow. 4500 // Ref: "Hacker's Delight" by Henry Warren 4501 unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); 4502 4503 // Note that underestimating the number of sign bits gives a more 4504 // conservative answer. 4505 unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) + 4506 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT); 4507 4508 // First handle the easy case: if we have enough sign bits there's 4509 // definitely no overflow. 4510 if (SignBits > BitWidth + 1) 4511 return OverflowResult::NeverOverflows; 4512 4513 // There are two ambiguous cases where there can be no overflow: 4514 // SignBits == BitWidth + 1 and 4515 // SignBits == BitWidth 4516 // The second case is difficult to check, therefore we only handle the 4517 // first case. 4518 if (SignBits == BitWidth + 1) { 4519 // It overflows only when both arguments are negative and the true 4520 // product is exactly the minimum negative number. 4521 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 4522 // For simplicity we just check if at least one side is not negative. 4523 KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, 4524 nullptr, UseInstrInfo); 4525 KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, 4526 nullptr, UseInstrInfo); 4527 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) 4528 return OverflowResult::NeverOverflows; 4529 } 4530 return OverflowResult::MayOverflow; 4531 } 4532 4533 OverflowResult llvm::computeOverflowForUnsignedAdd( 4534 const Value *LHS, const Value *RHS, const DataLayout &DL, 4535 AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, 4536 bool UseInstrInfo) { 4537 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4538 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, 4539 nullptr, UseInstrInfo); 4540 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4541 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, 4542 nullptr, UseInstrInfo); 4543 return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange)); 4544 } 4545 4546 static OverflowResult computeOverflowForSignedAdd(const Value *LHS, 4547 const Value *RHS, 4548 const AddOperator *Add, 4549 const DataLayout &DL, 4550 AssumptionCache *AC, 4551 const Instruction *CxtI, 4552 const DominatorTree *DT) { 4553 if (Add && Add->hasNoSignedWrap()) { 4554 return OverflowResult::NeverOverflows; 4555 } 4556 4557 // If LHS and RHS each have at least two sign bits, the addition will look 4558 // like 4559 // 4560 // XX..... + 4561 // YY..... 4562 // 4563 // If the carry into the most significant position is 0, X and Y can't both 4564 // be 1 and therefore the carry out of the addition is also 0. 4565 // 4566 // If the carry into the most significant position is 1, X and Y can't both 4567 // be 0 and therefore the carry out of the addition is also 1. 4568 // 4569 // Since the carry into the most significant position is always equal to 4570 // the carry out of the addition, there is no signed overflow. 4571 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && 4572 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) 4573 return OverflowResult::NeverOverflows; 4574 4575 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4576 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4577 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4578 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4579 OverflowResult OR = 4580 mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange)); 4581 if (OR != OverflowResult::MayOverflow) 4582 return OR; 4583 4584 // The remaining code needs Add to be available. Early returns if not so. 4585 if (!Add) 4586 return OverflowResult::MayOverflow; 4587 4588 // If the sign of Add is the same as at least one of the operands, this add 4589 // CANNOT overflow. If this can be determined from the known bits of the 4590 // operands the above signedAddMayOverflow() check will have already done so. 4591 // The only other way to improve on the known bits is from an assumption, so 4592 // call computeKnownBitsFromAssume() directly. 4593 bool LHSOrRHSKnownNonNegative = 4594 (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); 4595 bool LHSOrRHSKnownNegative = 4596 (LHSRange.isAllNegative() || RHSRange.isAllNegative()); 4597 if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { 4598 KnownBits AddKnown(LHSRange.getBitWidth()); 4599 computeKnownBitsFromAssume( 4600 Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true)); 4601 if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || 4602 (AddKnown.isNegative() && LHSOrRHSKnownNegative)) 4603 return OverflowResult::NeverOverflows; 4604 } 4605 4606 return OverflowResult::MayOverflow; 4607 } 4608 4609 OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, 4610 const Value *RHS, 4611 const DataLayout &DL, 4612 AssumptionCache *AC, 4613 const Instruction *CxtI, 4614 const DominatorTree *DT) { 4615 // Checking for conditions implied by dominating conditions may be expensive. 4616 // Limit it to usub_with_overflow calls for now. 4617 if (match(CxtI, 4618 m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) 4619 if (auto C = 4620 isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) { 4621 if (*C) 4622 return OverflowResult::NeverOverflows; 4623 return OverflowResult::AlwaysOverflowsLow; 4624 } 4625 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4626 LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); 4627 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4628 RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); 4629 return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange)); 4630 } 4631 4632 OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, 4633 const Value *RHS, 4634 const DataLayout &DL, 4635 AssumptionCache *AC, 4636 const Instruction *CxtI, 4637 const DominatorTree *DT) { 4638 // If LHS and RHS each have at least two sign bits, the subtraction 4639 // cannot overflow. 4640 if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && 4641 ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) 4642 return OverflowResult::NeverOverflows; 4643 4644 ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( 4645 LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4646 ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( 4647 RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); 4648 return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange)); 4649 } 4650 4651 bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, 4652 const DominatorTree &DT) { 4653 SmallVector<const BranchInst *, 2> GuardingBranches; 4654 SmallVector<const ExtractValueInst *, 2> Results; 4655 4656 for (const User *U : WO->users()) { 4657 if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) { 4658 assert(EVI->getNumIndices() == 1 && "Obvious from CI's type"); 4659 4660 if (EVI->getIndices()[0] == 0) 4661 Results.push_back(EVI); 4662 else { 4663 assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type"); 4664 4665 for (const auto *U : EVI->users()) 4666 if (const auto *B = dyn_cast<BranchInst>(U)) { 4667 assert(B->isConditional() && "How else is it using an i1?"); 4668 GuardingBranches.push_back(B); 4669 } 4670 } 4671 } else { 4672 // We are using the aggregate directly in a way we don't want to analyze 4673 // here (storing it to a global, say). 4674 return false; 4675 } 4676 } 4677 4678 auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { 4679 BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); 4680 if (!NoWrapEdge.isSingleEdge()) 4681 return false; 4682 4683 // Check if all users of the add are provably no-wrap. 4684 for (const auto *Result : Results) { 4685 // If the extractvalue itself is not executed on overflow, the we don't 4686 // need to check each use separately, since domination is transitive. 4687 if (DT.dominates(NoWrapEdge, Result->getParent())) 4688 continue; 4689 4690 for (auto &RU : Result->uses()) 4691 if (!DT.dominates(NoWrapEdge, RU)) 4692 return false; 4693 } 4694 4695 return true; 4696 }; 4697 4698 return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch); 4699 } 4700 4701 static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly) { 4702 // See whether I has flags that may create poison 4703 if (const auto *OvOp = dyn_cast<OverflowingBinaryOperator>(Op)) { 4704 if (OvOp->hasNoSignedWrap() || OvOp->hasNoUnsignedWrap()) 4705 return true; 4706 } 4707 if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(Op)) 4708 if (ExactOp->isExact()) 4709 return true; 4710 if (const auto *FP = dyn_cast<FPMathOperator>(Op)) { 4711 auto FMF = FP->getFastMathFlags(); 4712 if (FMF.noNaNs() || FMF.noInfs()) 4713 return true; 4714 } 4715 4716 unsigned Opcode = Op->getOpcode(); 4717 4718 // Check whether opcode is a poison/undef-generating operation 4719 switch (Opcode) { 4720 case Instruction::Shl: 4721 case Instruction::AShr: 4722 case Instruction::LShr: { 4723 // Shifts return poison if shiftwidth is larger than the bitwidth. 4724 if (auto *C = dyn_cast<Constant>(Op->getOperand(1))) { 4725 SmallVector<Constant *, 4> ShiftAmounts; 4726 if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) { 4727 unsigned NumElts = FVTy->getNumElements(); 4728 for (unsigned i = 0; i < NumElts; ++i) 4729 ShiftAmounts.push_back(C->getAggregateElement(i)); 4730 } else if (isa<ScalableVectorType>(C->getType())) 4731 return true; // Can't tell, just return true to be safe 4732 else 4733 ShiftAmounts.push_back(C); 4734 4735 bool Safe = llvm::all_of(ShiftAmounts, [](Constant *C) { 4736 auto *CI = dyn_cast_or_null<ConstantInt>(C); 4737 return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth()); 4738 }); 4739 return !Safe; 4740 } 4741 return true; 4742 } 4743 case Instruction::FPToSI: 4744 case Instruction::FPToUI: 4745 // fptosi/ui yields poison if the resulting value does not fit in the 4746 // destination type. 4747 return true; 4748 case Instruction::Call: 4749 case Instruction::CallBr: 4750 case Instruction::Invoke: { 4751 const auto *CB = cast<CallBase>(Op); 4752 return !CB->hasRetAttr(Attribute::NoUndef); 4753 } 4754 case Instruction::InsertElement: 4755 case Instruction::ExtractElement: { 4756 // If index exceeds the length of the vector, it returns poison 4757 auto *VTy = cast<VectorType>(Op->getOperand(0)->getType()); 4758 unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; 4759 auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp)); 4760 if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue())) 4761 return true; 4762 return false; 4763 } 4764 case Instruction::ShuffleVector: { 4765 // shufflevector may return undef. 4766 if (PoisonOnly) 4767 return false; 4768 ArrayRef<int> Mask = isa<ConstantExpr>(Op) 4769 ? cast<ConstantExpr>(Op)->getShuffleMask() 4770 : cast<ShuffleVectorInst>(Op)->getShuffleMask(); 4771 return is_contained(Mask, UndefMaskElem); 4772 } 4773 case Instruction::FNeg: 4774 case Instruction::PHI: 4775 case Instruction::Select: 4776 case Instruction::URem: 4777 case Instruction::SRem: 4778 case Instruction::ExtractValue: 4779 case Instruction::InsertValue: 4780 case Instruction::Freeze: 4781 case Instruction::ICmp: 4782 case Instruction::FCmp: 4783 return false; 4784 case Instruction::GetElementPtr: { 4785 const auto *GEP = cast<GEPOperator>(Op); 4786 return GEP->isInBounds(); 4787 } 4788 default: { 4789 const auto *CE = dyn_cast<ConstantExpr>(Op); 4790 if (isa<CastInst>(Op) || (CE && CE->isCast())) 4791 return false; 4792 else if (Instruction::isBinaryOp(Opcode)) 4793 return false; 4794 // Be conservative and return true. 4795 return true; 4796 } 4797 } 4798 } 4799 4800 bool llvm::canCreateUndefOrPoison(const Operator *Op) { 4801 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false); 4802 } 4803 4804 bool llvm::canCreatePoison(const Operator *Op) { 4805 return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true); 4806 } 4807 4808 static bool directlyImpliesPoison(const Value *ValAssumedPoison, 4809 const Value *V, unsigned Depth) { 4810 if (ValAssumedPoison == V) 4811 return true; 4812 4813 const unsigned MaxDepth = 2; 4814 if (Depth >= MaxDepth) 4815 return false; 4816 4817 const auto *I = dyn_cast<Instruction>(V); 4818 if (I && propagatesPoison(cast<Operator>(I))) { 4819 return any_of(I->operands(), [=](const Value *Op) { 4820 return directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1); 4821 }); 4822 } 4823 return false; 4824 } 4825 4826 static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, 4827 unsigned Depth) { 4828 if (isGuaranteedNotToBeUndefOrPoison(ValAssumedPoison)) 4829 return true; 4830 4831 if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) 4832 return true; 4833 4834 const unsigned MaxDepth = 2; 4835 if (Depth >= MaxDepth) 4836 return false; 4837 4838 const auto *I = dyn_cast<Instruction>(ValAssumedPoison); 4839 if (I && !canCreatePoison(cast<Operator>(I))) { 4840 return all_of(I->operands(), [=](const Value *Op) { 4841 return impliesPoison(Op, V, Depth + 1); 4842 }); 4843 } 4844 return false; 4845 } 4846 4847 bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { 4848 return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); 4849 } 4850 4851 static bool programUndefinedIfUndefOrPoison(const Value *V, 4852 bool PoisonOnly); 4853 4854 static bool isGuaranteedNotToBeUndefOrPoison(const Value *V, 4855 AssumptionCache *AC, 4856 const Instruction *CtxI, 4857 const DominatorTree *DT, 4858 unsigned Depth, bool PoisonOnly) { 4859 if (Depth >= MaxAnalysisRecursionDepth) 4860 return false; 4861 4862 if (isa<MetadataAsValue>(V)) 4863 return false; 4864 4865 if (const auto *A = dyn_cast<Argument>(V)) { 4866 if (A->hasAttribute(Attribute::NoUndef)) 4867 return true; 4868 } 4869 4870 if (auto *C = dyn_cast<Constant>(V)) { 4871 if (isa<UndefValue>(C)) 4872 return PoisonOnly && !isa<PoisonValue>(C); 4873 4874 if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) || 4875 isa<ConstantPointerNull>(C) || isa<Function>(C)) 4876 return true; 4877 4878 if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C)) 4879 return (PoisonOnly ? !C->containsPoisonElement() 4880 : !C->containsUndefOrPoisonElement()) && 4881 !C->containsConstantExpression(); 4882 } 4883 4884 // Strip cast operations from a pointer value. 4885 // Note that stripPointerCastsSameRepresentation can strip off getelementptr 4886 // inbounds with zero offset. To guarantee that the result isn't poison, the 4887 // stripped pointer is checked as it has to be pointing into an allocated 4888 // object or be null `null` to ensure `inbounds` getelement pointers with a 4889 // zero offset could not produce poison. 4890 // It can strip off addrspacecast that do not change bit representation as 4891 // well. We believe that such addrspacecast is equivalent to no-op. 4892 auto *StrippedV = V->stripPointerCastsSameRepresentation(); 4893 if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) || 4894 isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV)) 4895 return true; 4896 4897 auto OpCheck = [&](const Value *V) { 4898 return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1, 4899 PoisonOnly); 4900 }; 4901 4902 if (auto *Opr = dyn_cast<Operator>(V)) { 4903 // If the value is a freeze instruction, then it can never 4904 // be undef or poison. 4905 if (isa<FreezeInst>(V)) 4906 return true; 4907 4908 if (const auto *CB = dyn_cast<CallBase>(V)) { 4909 if (CB->hasRetAttr(Attribute::NoUndef)) 4910 return true; 4911 } 4912 4913 if (const auto *PN = dyn_cast<PHINode>(V)) { 4914 unsigned Num = PN->getNumIncomingValues(); 4915 bool IsWellDefined = true; 4916 for (unsigned i = 0; i < Num; ++i) { 4917 auto *TI = PN->getIncomingBlock(i)->getTerminator(); 4918 if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI, 4919 DT, Depth + 1, PoisonOnly)) { 4920 IsWellDefined = false; 4921 break; 4922 } 4923 } 4924 if (IsWellDefined) 4925 return true; 4926 } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck)) 4927 return true; 4928 } 4929 4930 if (auto *I = dyn_cast<LoadInst>(V)) 4931 if (I->getMetadata(LLVMContext::MD_noundef)) 4932 return true; 4933 4934 if (programUndefinedIfUndefOrPoison(V, PoisonOnly)) 4935 return true; 4936 4937 // CxtI may be null or a cloned instruction. 4938 if (!CtxI || !CtxI->getParent() || !DT) 4939 return false; 4940 4941 auto *DNode = DT->getNode(CtxI->getParent()); 4942 if (!DNode) 4943 // Unreachable block 4944 return false; 4945 4946 // If V is used as a branch condition before reaching CtxI, V cannot be 4947 // undef or poison. 4948 // br V, BB1, BB2 4949 // BB1: 4950 // CtxI ; V cannot be undef or poison here 4951 auto *Dominator = DNode->getIDom(); 4952 while (Dominator) { 4953 auto *TI = Dominator->getBlock()->getTerminator(); 4954 4955 Value *Cond = nullptr; 4956 if (auto BI = dyn_cast<BranchInst>(TI)) { 4957 if (BI->isConditional()) 4958 Cond = BI->getCondition(); 4959 } else if (auto SI = dyn_cast<SwitchInst>(TI)) { 4960 Cond = SI->getCondition(); 4961 } 4962 4963 if (Cond) { 4964 if (Cond == V) 4965 return true; 4966 else if (PoisonOnly && isa<Operator>(Cond)) { 4967 // For poison, we can analyze further 4968 auto *Opr = cast<Operator>(Cond); 4969 if (propagatesPoison(Opr) && is_contained(Opr->operand_values(), V)) 4970 return true; 4971 } 4972 } 4973 4974 Dominator = Dominator->getIDom(); 4975 } 4976 4977 SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NoUndef}; 4978 if (getKnowledgeValidInContext(V, AttrKinds, CtxI, DT, AC)) 4979 return true; 4980 4981 return false; 4982 } 4983 4984 bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, 4985 const Instruction *CtxI, 4986 const DominatorTree *DT, 4987 unsigned Depth) { 4988 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false); 4989 } 4990 4991 bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, 4992 const Instruction *CtxI, 4993 const DominatorTree *DT, unsigned Depth) { 4994 return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true); 4995 } 4996 4997 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, 4998 const DataLayout &DL, 4999 AssumptionCache *AC, 5000 const Instruction *CxtI, 5001 const DominatorTree *DT) { 5002 return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), 5003 Add, DL, AC, CxtI, DT); 5004 } 5005 5006 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS, 5007 const Value *RHS, 5008 const DataLayout &DL, 5009 AssumptionCache *AC, 5010 const Instruction *CxtI, 5011 const DominatorTree *DT) { 5012 return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); 5013 } 5014 5015 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { 5016 // Note: An atomic operation isn't guaranteed to return in a reasonable amount 5017 // of time because it's possible for another thread to interfere with it for an 5018 // arbitrary length of time, but programs aren't allowed to rely on that. 5019 5020 // If there is no successor, then execution can't transfer to it. 5021 if (isa<ReturnInst>(I)) 5022 return false; 5023 if (isa<UnreachableInst>(I)) 5024 return false; 5025 5026 // An instruction that returns without throwing must transfer control flow 5027 // to a successor. 5028 return !I->mayThrow() && I->willReturn(); 5029 } 5030 5031 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { 5032 // TODO: This is slightly conservative for invoke instruction since exiting 5033 // via an exception *is* normal control for them. 5034 for (auto I = BB->begin(), E = BB->end(); I != E; ++I) 5035 if (!isGuaranteedToTransferExecutionToSuccessor(&*I)) 5036 return false; 5037 return true; 5038 } 5039 5040 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, 5041 const Loop *L) { 5042 // The loop header is guaranteed to be executed for every iteration. 5043 // 5044 // FIXME: Relax this constraint to cover all basic blocks that are 5045 // guaranteed to be executed at every iteration. 5046 if (I->getParent() != L->getHeader()) return false; 5047 5048 for (const Instruction &LI : *L->getHeader()) { 5049 if (&LI == I) return true; 5050 if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; 5051 } 5052 llvm_unreachable("Instruction not contained in its own parent basic block."); 5053 } 5054 5055 bool llvm::propagatesPoison(const Operator *I) { 5056 switch (I->getOpcode()) { 5057 case Instruction::Freeze: 5058 case Instruction::Select: 5059 case Instruction::PHI: 5060 case Instruction::Call: 5061 case Instruction::Invoke: 5062 return false; 5063 case Instruction::ICmp: 5064 case Instruction::FCmp: 5065 case Instruction::GetElementPtr: 5066 return true; 5067 default: 5068 if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I)) 5069 return true; 5070 5071 // Be conservative and return false. 5072 return false; 5073 } 5074 } 5075 5076 void llvm::getGuaranteedNonPoisonOps(const Instruction *I, 5077 SmallPtrSetImpl<const Value *> &Operands) { 5078 switch (I->getOpcode()) { 5079 case Instruction::Store: 5080 Operands.insert(cast<StoreInst>(I)->getPointerOperand()); 5081 break; 5082 5083 case Instruction::Load: 5084 Operands.insert(cast<LoadInst>(I)->getPointerOperand()); 5085 break; 5086 5087 case Instruction::AtomicCmpXchg: 5088 Operands.insert(cast<AtomicCmpXchgInst>(I)->getPointerOperand()); 5089 break; 5090 5091 case Instruction::AtomicRMW: 5092 Operands.insert(cast<AtomicRMWInst>(I)->getPointerOperand()); 5093 break; 5094 5095 case Instruction::UDiv: 5096 case Instruction::SDiv: 5097 case Instruction::URem: 5098 case Instruction::SRem: 5099 Operands.insert(I->getOperand(1)); 5100 break; 5101 5102 case Instruction::Call: 5103 case Instruction::Invoke: { 5104 const CallBase *CB = cast<CallBase>(I); 5105 if (CB->isIndirectCall()) 5106 Operands.insert(CB->getCalledOperand()); 5107 for (unsigned i = 0; i < CB->arg_size(); ++i) { 5108 if (CB->paramHasAttr(i, Attribute::NoUndef)) 5109 Operands.insert(CB->getArgOperand(i)); 5110 } 5111 break; 5112 } 5113 5114 default: 5115 break; 5116 } 5117 } 5118 5119 bool llvm::mustTriggerUB(const Instruction *I, 5120 const SmallSet<const Value *, 16>& KnownPoison) { 5121 SmallPtrSet<const Value *, 4> NonPoisonOps; 5122 getGuaranteedNonPoisonOps(I, NonPoisonOps); 5123 5124 for (const auto *V : NonPoisonOps) 5125 if (KnownPoison.count(V)) 5126 return true; 5127 5128 return false; 5129 } 5130 5131 static bool programUndefinedIfUndefOrPoison(const Value *V, 5132 bool PoisonOnly) { 5133 // We currently only look for uses of values within the same basic 5134 // block, as that makes it easier to guarantee that the uses will be 5135 // executed given that Inst is executed. 5136 // 5137 // FIXME: Expand this to consider uses beyond the same basic block. To do 5138 // this, look out for the distinction between post-dominance and strong 5139 // post-dominance. 5140 const BasicBlock *BB = nullptr; 5141 BasicBlock::const_iterator Begin; 5142 if (const auto *Inst = dyn_cast<Instruction>(V)) { 5143 BB = Inst->getParent(); 5144 Begin = Inst->getIterator(); 5145 Begin++; 5146 } else if (const auto *Arg = dyn_cast<Argument>(V)) { 5147 BB = &Arg->getParent()->getEntryBlock(); 5148 Begin = BB->begin(); 5149 } else { 5150 return false; 5151 } 5152 5153 // Limit number of instructions we look at, to avoid scanning through large 5154 // blocks. The current limit is chosen arbitrarily. 5155 unsigned ScanLimit = 32; 5156 BasicBlock::const_iterator End = BB->end(); 5157 5158 if (!PoisonOnly) { 5159 // Be conservative & just check whether a value is passed to a noundef 5160 // argument. 5161 // Instructions that raise UB with a poison operand are well-defined 5162 // or have unclear semantics when the input is partially undef. 5163 // For example, 'udiv x, (undef | 1)' isn't UB. 5164 5165 for (auto &I : make_range(Begin, End)) { 5166 if (isa<DbgInfoIntrinsic>(I)) 5167 continue; 5168 if (--ScanLimit == 0) 5169 break; 5170 5171 if (const auto *CB = dyn_cast<CallBase>(&I)) { 5172 for (unsigned i = 0; i < CB->arg_size(); ++i) { 5173 if (CB->paramHasAttr(i, Attribute::NoUndef) && 5174 CB->getArgOperand(i) == V) 5175 return true; 5176 } 5177 } 5178 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 5179 break; 5180 } 5181 return false; 5182 } 5183 5184 // Set of instructions that we have proved will yield poison if Inst 5185 // does. 5186 SmallSet<const Value *, 16> YieldsPoison; 5187 SmallSet<const BasicBlock *, 4> Visited; 5188 5189 YieldsPoison.insert(V); 5190 auto Propagate = [&](const User *User) { 5191 if (propagatesPoison(cast<Operator>(User))) 5192 YieldsPoison.insert(User); 5193 }; 5194 for_each(V->users(), Propagate); 5195 Visited.insert(BB); 5196 5197 while (true) { 5198 for (auto &I : make_range(Begin, End)) { 5199 if (isa<DbgInfoIntrinsic>(I)) 5200 continue; 5201 if (--ScanLimit == 0) 5202 return false; 5203 if (mustTriggerUB(&I, YieldsPoison)) 5204 return true; 5205 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) 5206 return false; 5207 5208 // Mark poison that propagates from I through uses of I. 5209 if (YieldsPoison.count(&I)) 5210 for_each(I.users(), Propagate); 5211 } 5212 5213 if (auto *NextBB = BB->getSingleSuccessor()) { 5214 if (Visited.insert(NextBB).second) { 5215 BB = NextBB; 5216 Begin = BB->getFirstNonPHI()->getIterator(); 5217 End = BB->end(); 5218 continue; 5219 } 5220 } 5221 5222 break; 5223 } 5224 return false; 5225 } 5226 5227 bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { 5228 return ::programUndefinedIfUndefOrPoison(Inst, false); 5229 } 5230 5231 bool llvm::programUndefinedIfPoison(const Instruction *Inst) { 5232 return ::programUndefinedIfUndefOrPoison(Inst, true); 5233 } 5234 5235 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { 5236 if (FMF.noNaNs()) 5237 return true; 5238 5239 if (auto *C = dyn_cast<ConstantFP>(V)) 5240 return !C->isNaN(); 5241 5242 if (auto *C = dyn_cast<ConstantDataVector>(V)) { 5243 if (!C->getElementType()->isFloatingPointTy()) 5244 return false; 5245 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { 5246 if (C->getElementAsAPFloat(I).isNaN()) 5247 return false; 5248 } 5249 return true; 5250 } 5251 5252 if (isa<ConstantAggregateZero>(V)) 5253 return true; 5254 5255 return false; 5256 } 5257 5258 static bool isKnownNonZero(const Value *V) { 5259 if (auto *C = dyn_cast<ConstantFP>(V)) 5260 return !C->isZero(); 5261 5262 if (auto *C = dyn_cast<ConstantDataVector>(V)) { 5263 if (!C->getElementType()->isFloatingPointTy()) 5264 return false; 5265 for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { 5266 if (C->getElementAsAPFloat(I).isZero()) 5267 return false; 5268 } 5269 return true; 5270 } 5271 5272 return false; 5273 } 5274 5275 /// Match clamp pattern for float types without care about NaNs or signed zeros. 5276 /// Given non-min/max outer cmp/select from the clamp pattern this 5277 /// function recognizes if it can be substitued by a "canonical" min/max 5278 /// pattern. 5279 static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, 5280 Value *CmpLHS, Value *CmpRHS, 5281 Value *TrueVal, Value *FalseVal, 5282 Value *&LHS, Value *&RHS) { 5283 // Try to match 5284 // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) 5285 // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) 5286 // and return description of the outer Max/Min. 5287 5288 // First, check if select has inverse order: 5289 if (CmpRHS == FalseVal) { 5290 std::swap(TrueVal, FalseVal); 5291 Pred = CmpInst::getInversePredicate(Pred); 5292 } 5293 5294 // Assume success now. If there's no match, callers should not use these anyway. 5295 LHS = TrueVal; 5296 RHS = FalseVal; 5297 5298 const APFloat *FC1; 5299 if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite()) 5300 return {SPF_UNKNOWN, SPNB_NA, false}; 5301 5302 const APFloat *FC2; 5303 switch (Pred) { 5304 case CmpInst::FCMP_OLT: 5305 case CmpInst::FCMP_OLE: 5306 case CmpInst::FCMP_ULT: 5307 case CmpInst::FCMP_ULE: 5308 if (match(FalseVal, 5309 m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)), 5310 m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) && 5311 *FC1 < *FC2) 5312 return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false}; 5313 break; 5314 case CmpInst::FCMP_OGT: 5315 case CmpInst::FCMP_OGE: 5316 case CmpInst::FCMP_UGT: 5317 case CmpInst::FCMP_UGE: 5318 if (match(FalseVal, 5319 m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)), 5320 m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) && 5321 *FC1 > *FC2) 5322 return {SPF_FMINNUM, SPNB_RETURNS_ANY, false}; 5323 break; 5324 default: 5325 break; 5326 } 5327 5328 return {SPF_UNKNOWN, SPNB_NA, false}; 5329 } 5330 5331 /// Recognize variations of: 5332 /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) 5333 static SelectPatternResult matchClamp(CmpInst::Predicate Pred, 5334 Value *CmpLHS, Value *CmpRHS, 5335 Value *TrueVal, Value *FalseVal) { 5336 // Swap the select operands and predicate to match the patterns below. 5337 if (CmpRHS != TrueVal) { 5338 Pred = ICmpInst::getSwappedPredicate(Pred); 5339 std::swap(TrueVal, FalseVal); 5340 } 5341 const APInt *C1; 5342 if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) { 5343 const APInt *C2; 5344 // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) 5345 if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) && 5346 C1->slt(*C2) && Pred == CmpInst::ICMP_SLT) 5347 return {SPF_SMAX, SPNB_NA, false}; 5348 5349 // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) 5350 if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) && 5351 C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT) 5352 return {SPF_SMIN, SPNB_NA, false}; 5353 5354 // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) 5355 if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) && 5356 C1->ult(*C2) && Pred == CmpInst::ICMP_ULT) 5357 return {SPF_UMAX, SPNB_NA, false}; 5358 5359 // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) 5360 if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) && 5361 C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT) 5362 return {SPF_UMIN, SPNB_NA, false}; 5363 } 5364 return {SPF_UNKNOWN, SPNB_NA, false}; 5365 } 5366 5367 /// Recognize variations of: 5368 /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) 5369 static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, 5370 Value *CmpLHS, Value *CmpRHS, 5371 Value *TVal, Value *FVal, 5372 unsigned Depth) { 5373 // TODO: Allow FP min/max with nnan/nsz. 5374 assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison"); 5375 5376 Value *A = nullptr, *B = nullptr; 5377 SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1); 5378 if (!SelectPatternResult::isMinOrMax(L.Flavor)) 5379 return {SPF_UNKNOWN, SPNB_NA, false}; 5380 5381 Value *C = nullptr, *D = nullptr; 5382 SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1); 5383 if (L.Flavor != R.Flavor) 5384 return {SPF_UNKNOWN, SPNB_NA, false}; 5385 5386 // We have something like: x Pred y ? min(a, b) : min(c, d). 5387 // Try to match the compare to the min/max operations of the select operands. 5388 // First, make sure we have the right compare predicate. 5389 switch (L.Flavor) { 5390 case SPF_SMIN: 5391 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { 5392 Pred = ICmpInst::getSwappedPredicate(Pred); 5393 std::swap(CmpLHS, CmpRHS); 5394 } 5395 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) 5396 break; 5397 return {SPF_UNKNOWN, SPNB_NA, false}; 5398 case SPF_SMAX: 5399 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { 5400 Pred = ICmpInst::getSwappedPredicate(Pred); 5401 std::swap(CmpLHS, CmpRHS); 5402 } 5403 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) 5404 break; 5405 return {SPF_UNKNOWN, SPNB_NA, false}; 5406 case SPF_UMIN: 5407 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 5408 Pred = ICmpInst::getSwappedPredicate(Pred); 5409 std::swap(CmpLHS, CmpRHS); 5410 } 5411 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) 5412 break; 5413 return {SPF_UNKNOWN, SPNB_NA, false}; 5414 case SPF_UMAX: 5415 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { 5416 Pred = ICmpInst::getSwappedPredicate(Pred); 5417 std::swap(CmpLHS, CmpRHS); 5418 } 5419 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) 5420 break; 5421 return {SPF_UNKNOWN, SPNB_NA, false}; 5422 default: 5423 return {SPF_UNKNOWN, SPNB_NA, false}; 5424 } 5425 5426 // If there is a common operand in the already matched min/max and the other 5427 // min/max operands match the compare operands (either directly or inverted), 5428 // then this is min/max of the same flavor. 5429 5430 // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) 5431 // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) 5432 if (D == B) { 5433 if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && 5434 match(A, m_Not(m_Specific(CmpRHS))))) 5435 return {L.Flavor, SPNB_NA, false}; 5436 } 5437 // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) 5438 // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) 5439 if (C == B) { 5440 if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && 5441 match(A, m_Not(m_Specific(CmpRHS))))) 5442 return {L.Flavor, SPNB_NA, false}; 5443 } 5444 // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) 5445 // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) 5446 if (D == A) { 5447 if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && 5448 match(B, m_Not(m_Specific(CmpRHS))))) 5449 return {L.Flavor, SPNB_NA, false}; 5450 } 5451 // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) 5452 // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) 5453 if (C == A) { 5454 if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && 5455 match(B, m_Not(m_Specific(CmpRHS))))) 5456 return {L.Flavor, SPNB_NA, false}; 5457 } 5458 5459 return {SPF_UNKNOWN, SPNB_NA, false}; 5460 } 5461 5462 /// If the input value is the result of a 'not' op, constant integer, or vector 5463 /// splat of a constant integer, return the bitwise-not source value. 5464 /// TODO: This could be extended to handle non-splat vector integer constants. 5465 static Value *getNotValue(Value *V) { 5466 Value *NotV; 5467 if (match(V, m_Not(m_Value(NotV)))) 5468 return NotV; 5469 5470 const APInt *C; 5471 if (match(V, m_APInt(C))) 5472 return ConstantInt::get(V->getType(), ~(*C)); 5473 5474 return nullptr; 5475 } 5476 5477 /// Match non-obvious integer minimum and maximum sequences. 5478 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, 5479 Value *CmpLHS, Value *CmpRHS, 5480 Value *TrueVal, Value *FalseVal, 5481 Value *&LHS, Value *&RHS, 5482 unsigned Depth) { 5483 // Assume success. If there's no match, callers should not use these anyway. 5484 LHS = TrueVal; 5485 RHS = FalseVal; 5486 5487 SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); 5488 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) 5489 return SPR; 5490 5491 SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth); 5492 if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) 5493 return SPR; 5494 5495 // Look through 'not' ops to find disguised min/max. 5496 // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) 5497 // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) 5498 if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) { 5499 switch (Pred) { 5500 case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false}; 5501 case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false}; 5502 case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false}; 5503 case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false}; 5504 default: break; 5505 } 5506 } 5507 5508 // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) 5509 // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) 5510 if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) { 5511 switch (Pred) { 5512 case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false}; 5513 case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false}; 5514 case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false}; 5515 case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false}; 5516 default: break; 5517 } 5518 } 5519 5520 if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) 5521 return {SPF_UNKNOWN, SPNB_NA, false}; 5522 5523 // Z = X -nsw Y 5524 // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0) 5525 // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0) 5526 if (match(TrueVal, m_Zero()) && 5527 match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) 5528 return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false}; 5529 5530 // Z = X -nsw Y 5531 // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0) 5532 // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0) 5533 if (match(FalseVal, m_Zero()) && 5534 match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS)))) 5535 return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false}; 5536 5537 const APInt *C1; 5538 if (!match(CmpRHS, m_APInt(C1))) 5539 return {SPF_UNKNOWN, SPNB_NA, false}; 5540 5541 // An unsigned min/max can be written with a signed compare. 5542 const APInt *C2; 5543 if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) || 5544 (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) { 5545 // Is the sign bit set? 5546 // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX 5547 // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN 5548 if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() && 5549 C2->isMaxSignedValue()) 5550 return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; 5551 5552 // Is the sign bit clear? 5553 // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX 5554 // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN 5555 if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() && 5556 C2->isMinSignedValue()) 5557 return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; 5558 } 5559 5560 return {SPF_UNKNOWN, SPNB_NA, false}; 5561 } 5562 5563 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) { 5564 assert(X && Y && "Invalid operand"); 5565 5566 // X = sub (0, Y) || X = sub nsw (0, Y) 5567 if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) || 5568 (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y))))) 5569 return true; 5570 5571 // Y = sub (0, X) || Y = sub nsw (0, X) 5572 if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) || 5573 (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X))))) 5574 return true; 5575 5576 // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) 5577 Value *A, *B; 5578 return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) && 5579 match(Y, m_Sub(m_Specific(B), m_Specific(A))))) || 5580 (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) && 5581 match(Y, m_NSWSub(m_Specific(B), m_Specific(A))))); 5582 } 5583 5584 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, 5585 FastMathFlags FMF, 5586 Value *CmpLHS, Value *CmpRHS, 5587 Value *TrueVal, Value *FalseVal, 5588 Value *&LHS, Value *&RHS, 5589 unsigned Depth) { 5590 if (CmpInst::isFPPredicate(Pred)) { 5591 // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one 5592 // 0.0 operand, set the compare's 0.0 operands to that same value for the 5593 // purpose of identifying min/max. Disregard vector constants with undefined 5594 // elements because those can not be back-propagated for analysis. 5595 Value *OutputZeroVal = nullptr; 5596 if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) && 5597 !cast<Constant>(TrueVal)->containsUndefOrPoisonElement()) 5598 OutputZeroVal = TrueVal; 5599 else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) && 5600 !cast<Constant>(FalseVal)->containsUndefOrPoisonElement()) 5601 OutputZeroVal = FalseVal; 5602 5603 if (OutputZeroVal) { 5604 if (match(CmpLHS, m_AnyZeroFP())) 5605 CmpLHS = OutputZeroVal; 5606 if (match(CmpRHS, m_AnyZeroFP())) 5607 CmpRHS = OutputZeroVal; 5608 } 5609 } 5610 5611 LHS = CmpLHS; 5612 RHS = CmpRHS; 5613 5614 // Signed zero may return inconsistent results between implementations. 5615 // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 5616 // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) 5617 // Therefore, we behave conservatively and only proceed if at least one of the 5618 // operands is known to not be zero or if we don't care about signed zero. 5619 switch (Pred) { 5620 default: break; 5621 // FIXME: Include OGT/OLT/UGT/ULT. 5622 case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: 5623 case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: 5624 if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && 5625 !isKnownNonZero(CmpRHS)) 5626 return {SPF_UNKNOWN, SPNB_NA, false}; 5627 } 5628 5629 SelectPatternNaNBehavior NaNBehavior = SPNB_NA; 5630 bool Ordered = false; 5631 5632 // When given one NaN and one non-NaN input: 5633 // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. 5634 // - A simple C99 (a < b ? a : b) construction will return 'b' (as the 5635 // ordered comparison fails), which could be NaN or non-NaN. 5636 // so here we discover exactly what NaN behavior is required/accepted. 5637 if (CmpInst::isFPPredicate(Pred)) { 5638 bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); 5639 bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); 5640 5641 if (LHSSafe && RHSSafe) { 5642 // Both operands are known non-NaN. 5643 NaNBehavior = SPNB_RETURNS_ANY; 5644 } else if (CmpInst::isOrdered(Pred)) { 5645 // An ordered comparison will return false when given a NaN, so it 5646 // returns the RHS. 5647 Ordered = true; 5648 if (LHSSafe) 5649 // LHS is non-NaN, so if RHS is NaN then NaN will be returned. 5650 NaNBehavior = SPNB_RETURNS_NAN; 5651 else if (RHSSafe) 5652 NaNBehavior = SPNB_RETURNS_OTHER; 5653 else 5654 // Completely unsafe. 5655 return {SPF_UNKNOWN, SPNB_NA, false}; 5656 } else { 5657 Ordered = false; 5658 // An unordered comparison will return true when given a NaN, so it 5659 // returns the LHS. 5660 if (LHSSafe) 5661 // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. 5662 NaNBehavior = SPNB_RETURNS_OTHER; 5663 else if (RHSSafe) 5664 NaNBehavior = SPNB_RETURNS_NAN; 5665 else 5666 // Completely unsafe. 5667 return {SPF_UNKNOWN, SPNB_NA, false}; 5668 } 5669 } 5670 5671 if (TrueVal == CmpRHS && FalseVal == CmpLHS) { 5672 std::swap(CmpLHS, CmpRHS); 5673 Pred = CmpInst::getSwappedPredicate(Pred); 5674 if (NaNBehavior == SPNB_RETURNS_NAN) 5675 NaNBehavior = SPNB_RETURNS_OTHER; 5676 else if (NaNBehavior == SPNB_RETURNS_OTHER) 5677 NaNBehavior = SPNB_RETURNS_NAN; 5678 Ordered = !Ordered; 5679 } 5680 5681 // ([if]cmp X, Y) ? X : Y 5682 if (TrueVal == CmpLHS && FalseVal == CmpRHS) { 5683 switch (Pred) { 5684 default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. 5685 case ICmpInst::ICMP_UGT: 5686 case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; 5687 case ICmpInst::ICMP_SGT: 5688 case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; 5689 case ICmpInst::ICMP_ULT: 5690 case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; 5691 case ICmpInst::ICMP_SLT: 5692 case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; 5693 case FCmpInst::FCMP_UGT: 5694 case FCmpInst::FCMP_UGE: 5695 case FCmpInst::FCMP_OGT: 5696 case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; 5697 case FCmpInst::FCMP_ULT: 5698 case FCmpInst::FCMP_ULE: 5699 case FCmpInst::FCMP_OLT: 5700 case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; 5701 } 5702 } 5703 5704 if (isKnownNegation(TrueVal, FalseVal)) { 5705 // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can 5706 // match against either LHS or sext(LHS). 5707 auto MaybeSExtCmpLHS = 5708 m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS))); 5709 auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes()); 5710 auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One()); 5711 if (match(TrueVal, MaybeSExtCmpLHS)) { 5712 // Set the return values. If the compare uses the negated value (-X >s 0), 5713 // swap the return values because the negated value is always 'RHS'. 5714 LHS = TrueVal; 5715 RHS = FalseVal; 5716 if (match(CmpLHS, m_Neg(m_Specific(FalseVal)))) 5717 std::swap(LHS, RHS); 5718 5719 // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) 5720 // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) 5721 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) 5722 return {SPF_ABS, SPNB_NA, false}; 5723 5724 // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) 5725 if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne)) 5726 return {SPF_ABS, SPNB_NA, false}; 5727 5728 // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) 5729 // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) 5730 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) 5731 return {SPF_NABS, SPNB_NA, false}; 5732 } 5733 else if (match(FalseVal, MaybeSExtCmpLHS)) { 5734 // Set the return values. If the compare uses the negated value (-X >s 0), 5735 // swap the return values because the negated value is always 'RHS'. 5736 LHS = FalseVal; 5737 RHS = TrueVal; 5738 if (match(CmpLHS, m_Neg(m_Specific(TrueVal)))) 5739 std::swap(LHS, RHS); 5740 5741 // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) 5742 // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) 5743 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) 5744 return {SPF_NABS, SPNB_NA, false}; 5745 5746 // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) 5747 // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) 5748 if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) 5749 return {SPF_ABS, SPNB_NA, false}; 5750 } 5751 } 5752 5753 if (CmpInst::isIntPredicate(Pred)) 5754 return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); 5755 5756 // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar 5757 // may return either -0.0 or 0.0, so fcmp/select pair has stricter 5758 // semantics than minNum. Be conservative in such case. 5759 if (NaNBehavior != SPNB_RETURNS_ANY || 5760 (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && 5761 !isKnownNonZero(CmpRHS))) 5762 return {SPF_UNKNOWN, SPNB_NA, false}; 5763 5764 return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); 5765 } 5766 5767 /// Helps to match a select pattern in case of a type mismatch. 5768 /// 5769 /// The function processes the case when type of true and false values of a 5770 /// select instruction differs from type of the cmp instruction operands because 5771 /// of a cast instruction. The function checks if it is legal to move the cast 5772 /// operation after "select". If yes, it returns the new second value of 5773 /// "select" (with the assumption that cast is moved): 5774 /// 1. As operand of cast instruction when both values of "select" are same cast 5775 /// instructions. 5776 /// 2. As restored constant (by applying reverse cast operation) when the first 5777 /// value of the "select" is a cast operation and the second value is a 5778 /// constant. 5779 /// NOTE: We return only the new second value because the first value could be 5780 /// accessed as operand of cast instruction. 5781 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, 5782 Instruction::CastOps *CastOp) { 5783 auto *Cast1 = dyn_cast<CastInst>(V1); 5784 if (!Cast1) 5785 return nullptr; 5786 5787 *CastOp = Cast1->getOpcode(); 5788 Type *SrcTy = Cast1->getSrcTy(); 5789 if (auto *Cast2 = dyn_cast<CastInst>(V2)) { 5790 // If V1 and V2 are both the same cast from the same type, look through V1. 5791 if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) 5792 return Cast2->getOperand(0); 5793 return nullptr; 5794 } 5795 5796 auto *C = dyn_cast<Constant>(V2); 5797 if (!C) 5798 return nullptr; 5799 5800 Constant *CastedTo = nullptr; 5801 switch (*CastOp) { 5802 case Instruction::ZExt: 5803 if (CmpI->isUnsigned()) 5804 CastedTo = ConstantExpr::getTrunc(C, SrcTy); 5805 break; 5806 case Instruction::SExt: 5807 if (CmpI->isSigned()) 5808 CastedTo = ConstantExpr::getTrunc(C, SrcTy, true); 5809 break; 5810 case Instruction::Trunc: 5811 Constant *CmpConst; 5812 if (match(CmpI->getOperand(1), m_Constant(CmpConst)) && 5813 CmpConst->getType() == SrcTy) { 5814 // Here we have the following case: 5815 // 5816 // %cond = cmp iN %x, CmpConst 5817 // %tr = trunc iN %x to iK 5818 // %narrowsel = select i1 %cond, iK %t, iK C 5819 // 5820 // We can always move trunc after select operation: 5821 // 5822 // %cond = cmp iN %x, CmpConst 5823 // %widesel = select i1 %cond, iN %x, iN CmpConst 5824 // %tr = trunc iN %widesel to iK 5825 // 5826 // Note that C could be extended in any way because we don't care about 5827 // upper bits after truncation. It can't be abs pattern, because it would 5828 // look like: 5829 // 5830 // select i1 %cond, x, -x. 5831 // 5832 // So only min/max pattern could be matched. Such match requires widened C 5833 // == CmpConst. That is why set widened C = CmpConst, condition trunc 5834 // CmpConst == C is checked below. 5835 CastedTo = CmpConst; 5836 } else { 5837 CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned()); 5838 } 5839 break; 5840 case Instruction::FPTrunc: 5841 CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true); 5842 break; 5843 case Instruction::FPExt: 5844 CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true); 5845 break; 5846 case Instruction::FPToUI: 5847 CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true); 5848 break; 5849 case Instruction::FPToSI: 5850 CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true); 5851 break; 5852 case Instruction::UIToFP: 5853 CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true); 5854 break; 5855 case Instruction::SIToFP: 5856 CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true); 5857 break; 5858 default: 5859 break; 5860 } 5861 5862 if (!CastedTo) 5863 return nullptr; 5864 5865 // Make sure the cast doesn't lose any information. 5866 Constant *CastedBack = 5867 ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true); 5868 if (CastedBack != C) 5869 return nullptr; 5870 5871 return CastedTo; 5872 } 5873 5874 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, 5875 Instruction::CastOps *CastOp, 5876 unsigned Depth) { 5877 if (Depth >= MaxAnalysisRecursionDepth) 5878 return {SPF_UNKNOWN, SPNB_NA, false}; 5879 5880 SelectInst *SI = dyn_cast<SelectInst>(V); 5881 if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; 5882 5883 CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); 5884 if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; 5885 5886 Value *TrueVal = SI->getTrueValue(); 5887 Value *FalseVal = SI->getFalseValue(); 5888 5889 return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, 5890 CastOp, Depth); 5891 } 5892 5893 SelectPatternResult llvm::matchDecomposedSelectPattern( 5894 CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, 5895 Instruction::CastOps *CastOp, unsigned Depth) { 5896 CmpInst::Predicate Pred = CmpI->getPredicate(); 5897 Value *CmpLHS = CmpI->getOperand(0); 5898 Value *CmpRHS = CmpI->getOperand(1); 5899 FastMathFlags FMF; 5900 if (isa<FPMathOperator>(CmpI)) 5901 FMF = CmpI->getFastMathFlags(); 5902 5903 // Bail out early. 5904 if (CmpI->isEquality()) 5905 return {SPF_UNKNOWN, SPNB_NA, false}; 5906 5907 // Deal with type mismatches. 5908 if (CastOp && CmpLHS->getType() != TrueVal->getType()) { 5909 if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) { 5910 // If this is a potential fmin/fmax with a cast to integer, then ignore 5911 // -0.0 because there is no corresponding integer value. 5912 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) 5913 FMF.setNoSignedZeros(); 5914 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 5915 cast<CastInst>(TrueVal)->getOperand(0), C, 5916 LHS, RHS, Depth); 5917 } 5918 if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) { 5919 // If this is a potential fmin/fmax with a cast to integer, then ignore 5920 // -0.0 because there is no corresponding integer value. 5921 if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) 5922 FMF.setNoSignedZeros(); 5923 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, 5924 C, cast<CastInst>(FalseVal)->getOperand(0), 5925 LHS, RHS, Depth); 5926 } 5927 } 5928 return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, 5929 LHS, RHS, Depth); 5930 } 5931 5932 CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { 5933 if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; 5934 if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; 5935 if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; 5936 if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; 5937 if (SPF == SPF_FMINNUM) 5938 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; 5939 if (SPF == SPF_FMAXNUM) 5940 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; 5941 llvm_unreachable("unhandled!"); 5942 } 5943 5944 SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { 5945 if (SPF == SPF_SMIN) return SPF_SMAX; 5946 if (SPF == SPF_UMIN) return SPF_UMAX; 5947 if (SPF == SPF_SMAX) return SPF_SMIN; 5948 if (SPF == SPF_UMAX) return SPF_UMIN; 5949 llvm_unreachable("unhandled!"); 5950 } 5951 5952 CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) { 5953 return getMinMaxPred(getInverseMinMaxFlavor(SPF)); 5954 } 5955 5956 std::pair<Intrinsic::ID, bool> 5957 llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { 5958 // Check if VL contains select instructions that can be folded into a min/max 5959 // vector intrinsic and return the intrinsic if it is possible. 5960 // TODO: Support floating point min/max. 5961 bool AllCmpSingleUse = true; 5962 SelectPatternResult SelectPattern; 5963 SelectPattern.Flavor = SPF_UNKNOWN; 5964 if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) { 5965 Value *LHS, *RHS; 5966 auto CurrentPattern = matchSelectPattern(I, LHS, RHS); 5967 if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) || 5968 CurrentPattern.Flavor == SPF_FMINNUM || 5969 CurrentPattern.Flavor == SPF_FMAXNUM || 5970 !I->getType()->isIntOrIntVectorTy()) 5971 return false; 5972 if (SelectPattern.Flavor != SPF_UNKNOWN && 5973 SelectPattern.Flavor != CurrentPattern.Flavor) 5974 return false; 5975 SelectPattern = CurrentPattern; 5976 AllCmpSingleUse &= 5977 match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value())); 5978 return true; 5979 })) { 5980 switch (SelectPattern.Flavor) { 5981 case SPF_SMIN: 5982 return {Intrinsic::smin, AllCmpSingleUse}; 5983 case SPF_UMIN: 5984 return {Intrinsic::umin, AllCmpSingleUse}; 5985 case SPF_SMAX: 5986 return {Intrinsic::smax, AllCmpSingleUse}; 5987 case SPF_UMAX: 5988 return {Intrinsic::umax, AllCmpSingleUse}; 5989 default: 5990 llvm_unreachable("unexpected select pattern flavor"); 5991 } 5992 } 5993 return {Intrinsic::not_intrinsic, false}; 5994 } 5995 5996 /// Return true if "icmp Pred LHS RHS" is always true. 5997 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, 5998 const Value *RHS, const DataLayout &DL, 5999 unsigned Depth) { 6000 assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!"); 6001 if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) 6002 return true; 6003 6004 switch (Pred) { 6005 default: 6006 return false; 6007 6008 case CmpInst::ICMP_SLE: { 6009 const APInt *C; 6010 6011 // LHS s<= LHS +_{nsw} C if C >= 0 6012 if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) 6013 return !C->isNegative(); 6014 return false; 6015 } 6016 6017 case CmpInst::ICMP_ULE: { 6018 const APInt *C; 6019 6020 // LHS u<= LHS +_{nuw} C for any C 6021 if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) 6022 return true; 6023 6024 // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) 6025 auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B, 6026 const Value *&X, 6027 const APInt *&CA, const APInt *&CB) { 6028 if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && 6029 match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) 6030 return true; 6031 6032 // If X & C == 0 then (X | C) == X +_{nuw} C 6033 if (match(A, m_Or(m_Value(X), m_APInt(CA))) && 6034 match(B, m_Or(m_Specific(X), m_APInt(CB)))) { 6035 KnownBits Known(CA->getBitWidth()); 6036 computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr, 6037 /*CxtI*/ nullptr, /*DT*/ nullptr); 6038 if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero)) 6039 return true; 6040 } 6041 6042 return false; 6043 }; 6044 6045 const Value *X; 6046 const APInt *CLHS, *CRHS; 6047 if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) 6048 return CLHS->ule(*CRHS); 6049 6050 return false; 6051 } 6052 } 6053 } 6054 6055 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred 6056 /// ALHS ARHS" is true. Otherwise, return None. 6057 static Optional<bool> 6058 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, 6059 const Value *ARHS, const Value *BLHS, const Value *BRHS, 6060 const DataLayout &DL, unsigned Depth) { 6061 switch (Pred) { 6062 default: 6063 return None; 6064 6065 case CmpInst::ICMP_SLT: 6066 case CmpInst::ICMP_SLE: 6067 if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) && 6068 isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth)) 6069 return true; 6070 return None; 6071 6072 case CmpInst::ICMP_ULT: 6073 case CmpInst::ICMP_ULE: 6074 if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) && 6075 isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth)) 6076 return true; 6077 return None; 6078 } 6079 } 6080 6081 /// Return true if the operands of the two compares match. IsSwappedOps is true 6082 /// when the operands match, but are swapped. 6083 static bool isMatchingOps(const Value *ALHS, const Value *ARHS, 6084 const Value *BLHS, const Value *BRHS, 6085 bool &IsSwappedOps) { 6086 6087 bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS); 6088 IsSwappedOps = (ALHS == BRHS && ARHS == BLHS); 6089 return IsMatchingOps || IsSwappedOps; 6090 } 6091 6092 /// Return true if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is true. 6093 /// Return false if "icmp1 APred X, Y" implies "icmp2 BPred X, Y" is false. 6094 /// Otherwise, return None if we can't infer anything. 6095 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred, 6096 CmpInst::Predicate BPred, 6097 bool AreSwappedOps) { 6098 // Canonicalize the predicate as if the operands were not commuted. 6099 if (AreSwappedOps) 6100 BPred = ICmpInst::getSwappedPredicate(BPred); 6101 6102 if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred)) 6103 return true; 6104 if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred)) 6105 return false; 6106 6107 return None; 6108 } 6109 6110 /// Return true if "icmp APred X, C1" implies "icmp BPred X, C2" is true. 6111 /// Return false if "icmp APred X, C1" implies "icmp BPred X, C2" is false. 6112 /// Otherwise, return None if we can't infer anything. 6113 static Optional<bool> 6114 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, 6115 const ConstantInt *C1, 6116 CmpInst::Predicate BPred, 6117 const ConstantInt *C2) { 6118 ConstantRange DomCR = 6119 ConstantRange::makeExactICmpRegion(APred, C1->getValue()); 6120 ConstantRange CR = 6121 ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue()); 6122 ConstantRange Intersection = DomCR.intersectWith(CR); 6123 ConstantRange Difference = DomCR.difference(CR); 6124 if (Intersection.isEmptySet()) 6125 return false; 6126 if (Difference.isEmptySet()) 6127 return true; 6128 return None; 6129 } 6130 6131 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is 6132 /// false. Otherwise, return None if we can't infer anything. 6133 static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS, 6134 CmpInst::Predicate BPred, 6135 const Value *BLHS, const Value *BRHS, 6136 const DataLayout &DL, bool LHSIsTrue, 6137 unsigned Depth) { 6138 Value *ALHS = LHS->getOperand(0); 6139 Value *ARHS = LHS->getOperand(1); 6140 6141 // The rest of the logic assumes the LHS condition is true. If that's not the 6142 // case, invert the predicate to make it so. 6143 CmpInst::Predicate APred = 6144 LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); 6145 6146 // Can we infer anything when the two compares have matching operands? 6147 bool AreSwappedOps; 6148 if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, AreSwappedOps)) { 6149 if (Optional<bool> Implication = isImpliedCondMatchingOperands( 6150 APred, BPred, AreSwappedOps)) 6151 return Implication; 6152 // No amount of additional analysis will infer the second condition, so 6153 // early exit. 6154 return None; 6155 } 6156 6157 // Can we infer anything when the LHS operands match and the RHS operands are 6158 // constants (not necessarily matching)? 6159 if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) { 6160 if (Optional<bool> Implication = isImpliedCondMatchingImmOperands( 6161 APred, cast<ConstantInt>(ARHS), BPred, cast<ConstantInt>(BRHS))) 6162 return Implication; 6163 // No amount of additional analysis will infer the second condition, so 6164 // early exit. 6165 return None; 6166 } 6167 6168 if (APred == BPred) 6169 return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth); 6170 return None; 6171 } 6172 6173 /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is 6174 /// false. Otherwise, return None if we can't infer anything. We expect the 6175 /// RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' instruction. 6176 static Optional<bool> 6177 isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, 6178 const Value *RHSOp0, const Value *RHSOp1, 6179 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { 6180 // The LHS must be an 'or', 'and', or a 'select' instruction. 6181 assert((LHS->getOpcode() == Instruction::And || 6182 LHS->getOpcode() == Instruction::Or || 6183 LHS->getOpcode() == Instruction::Select) && 6184 "Expected LHS to be 'and', 'or', or 'select'."); 6185 6186 assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit"); 6187 6188 // If the result of an 'or' is false, then we know both legs of the 'or' are 6189 // false. Similarly, if the result of an 'and' is true, then we know both 6190 // legs of the 'and' are true. 6191 const Value *ALHS, *ARHS; 6192 if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) || 6193 (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) { 6194 // FIXME: Make this non-recursion. 6195 if (Optional<bool> Implication = isImpliedCondition( 6196 ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) 6197 return Implication; 6198 if (Optional<bool> Implication = isImpliedCondition( 6199 ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) 6200 return Implication; 6201 return None; 6202 } 6203 return None; 6204 } 6205 6206 Optional<bool> 6207 llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, 6208 const Value *RHSOp0, const Value *RHSOp1, 6209 const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { 6210 // Bail out when we hit the limit. 6211 if (Depth == MaxAnalysisRecursionDepth) 6212 return None; 6213 6214 // A mismatch occurs when we compare a scalar cmp to a vector cmp, for 6215 // example. 6216 if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) 6217 return None; 6218 6219 Type *OpTy = LHS->getType(); 6220 assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!"); 6221 6222 // FIXME: Extending the code below to handle vectors. 6223 if (OpTy->isVectorTy()) 6224 return None; 6225 6226 assert(OpTy->isIntegerTy(1) && "implied by above"); 6227 6228 // Both LHS and RHS are icmps. 6229 const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS); 6230 if (LHSCmp) 6231 return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, 6232 Depth); 6233 6234 /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect 6235 /// the RHS to be an icmp. 6236 /// FIXME: Add support for and/or/select on the RHS. 6237 if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) { 6238 if ((LHSI->getOpcode() == Instruction::And || 6239 LHSI->getOpcode() == Instruction::Or || 6240 LHSI->getOpcode() == Instruction::Select)) 6241 return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, 6242 Depth); 6243 } 6244 return None; 6245 } 6246 6247 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, 6248 const DataLayout &DL, bool LHSIsTrue, 6249 unsigned Depth) { 6250 // LHS ==> RHS by definition 6251 if (LHS == RHS) 6252 return LHSIsTrue; 6253 6254 const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS); 6255 if (RHSCmp) 6256 return isImpliedCondition(LHS, RHSCmp->getPredicate(), 6257 RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL, 6258 LHSIsTrue, Depth); 6259 return None; 6260 } 6261 6262 // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch 6263 // condition dominating ContextI or nullptr, if no condition is found. 6264 static std::pair<Value *, bool> 6265 getDomPredecessorCondition(const Instruction *ContextI) { 6266 if (!ContextI || !ContextI->getParent()) 6267 return {nullptr, false}; 6268 6269 // TODO: This is a poor/cheap way to determine dominance. Should we use a 6270 // dominator tree (eg, from a SimplifyQuery) instead? 6271 const BasicBlock *ContextBB = ContextI->getParent(); 6272 const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); 6273 if (!PredBB) 6274 return {nullptr, false}; 6275 6276 // We need a conditional branch in the predecessor. 6277 Value *PredCond; 6278 BasicBlock *TrueBB, *FalseBB; 6279 if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB))) 6280 return {nullptr, false}; 6281 6282 // The branch should get simplified. Don't bother simplifying this condition. 6283 if (TrueBB == FalseBB) 6284 return {nullptr, false}; 6285 6286 assert((TrueBB == ContextBB || FalseBB == ContextBB) && 6287 "Predecessor block does not point to successor?"); 6288 6289 // Is this condition implied by the predecessor condition? 6290 return {PredCond, TrueBB == ContextBB}; 6291 } 6292 6293 Optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, 6294 const Instruction *ContextI, 6295 const DataLayout &DL) { 6296 assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool"); 6297 auto PredCond = getDomPredecessorCondition(ContextI); 6298 if (PredCond.first) 6299 return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second); 6300 return None; 6301 } 6302 6303 Optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, 6304 const Value *LHS, const Value *RHS, 6305 const Instruction *ContextI, 6306 const DataLayout &DL) { 6307 auto PredCond = getDomPredecessorCondition(ContextI); 6308 if (PredCond.first) 6309 return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL, 6310 PredCond.second); 6311 return None; 6312 } 6313 6314 static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, 6315 APInt &Upper, const InstrInfoQuery &IIQ) { 6316 unsigned Width = Lower.getBitWidth(); 6317 const APInt *C; 6318 switch (BO.getOpcode()) { 6319 case Instruction::Add: 6320 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { 6321 // FIXME: If we have both nuw and nsw, we should reduce the range further. 6322 if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) { 6323 // 'add nuw x, C' produces [C, UINT_MAX]. 6324 Lower = *C; 6325 } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) { 6326 if (C->isNegative()) { 6327 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. 6328 Lower = APInt::getSignedMinValue(Width); 6329 Upper = APInt::getSignedMaxValue(Width) + *C + 1; 6330 } else { 6331 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. 6332 Lower = APInt::getSignedMinValue(Width) + *C; 6333 Upper = APInt::getSignedMaxValue(Width) + 1; 6334 } 6335 } 6336 } 6337 break; 6338 6339 case Instruction::And: 6340 if (match(BO.getOperand(1), m_APInt(C))) 6341 // 'and x, C' produces [0, C]. 6342 Upper = *C + 1; 6343 break; 6344 6345 case Instruction::Or: 6346 if (match(BO.getOperand(1), m_APInt(C))) 6347 // 'or x, C' produces [C, UINT_MAX]. 6348 Lower = *C; 6349 break; 6350 6351 case Instruction::AShr: 6352 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 6353 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. 6354 Lower = APInt::getSignedMinValue(Width).ashr(*C); 6355 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; 6356 } else if (match(BO.getOperand(0), m_APInt(C))) { 6357 unsigned ShiftAmount = Width - 1; 6358 if (!C->isNullValue() && IIQ.isExact(&BO)) 6359 ShiftAmount = C->countTrailingZeros(); 6360 if (C->isNegative()) { 6361 // 'ashr C, x' produces [C, C >> (Width-1)] 6362 Lower = *C; 6363 Upper = C->ashr(ShiftAmount) + 1; 6364 } else { 6365 // 'ashr C, x' produces [C >> (Width-1), C] 6366 Lower = C->ashr(ShiftAmount); 6367 Upper = *C + 1; 6368 } 6369 } 6370 break; 6371 6372 case Instruction::LShr: 6373 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { 6374 // 'lshr x, C' produces [0, UINT_MAX >> C]. 6375 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1; 6376 } else if (match(BO.getOperand(0), m_APInt(C))) { 6377 // 'lshr C, x' produces [C >> (Width-1), C]. 6378 unsigned ShiftAmount = Width - 1; 6379 if (!C->isNullValue() && IIQ.isExact(&BO)) 6380 ShiftAmount = C->countTrailingZeros(); 6381 Lower = C->lshr(ShiftAmount); 6382 Upper = *C + 1; 6383 } 6384 break; 6385 6386 case Instruction::Shl: 6387 if (match(BO.getOperand(0), m_APInt(C))) { 6388 if (IIQ.hasNoUnsignedWrap(&BO)) { 6389 // 'shl nuw C, x' produces [C, C << CLZ(C)] 6390 Lower = *C; 6391 Upper = Lower.shl(Lower.countLeadingZeros()) + 1; 6392 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? 6393 if (C->isNegative()) { 6394 // 'shl nsw C, x' produces [C << CLO(C)-1, C] 6395 unsigned ShiftAmount = C->countLeadingOnes() - 1; 6396 Lower = C->shl(ShiftAmount); 6397 Upper = *C + 1; 6398 } else { 6399 // 'shl nsw C, x' produces [C, C << CLZ(C)-1] 6400 unsigned ShiftAmount = C->countLeadingZeros() - 1; 6401 Lower = *C; 6402 Upper = C->shl(ShiftAmount) + 1; 6403 } 6404 } 6405 } 6406 break; 6407 6408 case Instruction::SDiv: 6409 if (match(BO.getOperand(1), m_APInt(C))) { 6410 APInt IntMin = APInt::getSignedMinValue(Width); 6411 APInt IntMax = APInt::getSignedMaxValue(Width); 6412 if (C->isAllOnesValue()) { 6413 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] 6414 // where C != -1 and C != 0 and C != 1 6415 Lower = IntMin + 1; 6416 Upper = IntMax + 1; 6417 } else if (C->countLeadingZeros() < Width - 1) { 6418 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] 6419 // where C != -1 and C != 0 and C != 1 6420 Lower = IntMin.sdiv(*C); 6421 Upper = IntMax.sdiv(*C); 6422 if (Lower.sgt(Upper)) 6423 std::swap(Lower, Upper); 6424 Upper = Upper + 1; 6425 assert(Upper != Lower && "Upper part of range has wrapped!"); 6426 } 6427 } else if (match(BO.getOperand(0), m_APInt(C))) { 6428 if (C->isMinSignedValue()) { 6429 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. 6430 Lower = *C; 6431 Upper = Lower.lshr(1) + 1; 6432 } else { 6433 // 'sdiv C, x' produces [-|C|, |C|]. 6434 Upper = C->abs() + 1; 6435 Lower = (-Upper) + 1; 6436 } 6437 } 6438 break; 6439 6440 case Instruction::UDiv: 6441 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) { 6442 // 'udiv x, C' produces [0, UINT_MAX / C]. 6443 Upper = APInt::getMaxValue(Width).udiv(*C) + 1; 6444 } else if (match(BO.getOperand(0), m_APInt(C))) { 6445 // 'udiv C, x' produces [0, C]. 6446 Upper = *C + 1; 6447 } 6448 break; 6449 6450 case Instruction::SRem: 6451 if (match(BO.getOperand(1), m_APInt(C))) { 6452 // 'srem x, C' produces (-|C|, |C|). 6453 Upper = C->abs(); 6454 Lower = (-Upper) + 1; 6455 } 6456 break; 6457 6458 case Instruction::URem: 6459 if (match(BO.getOperand(1), m_APInt(C))) 6460 // 'urem x, C' produces [0, C). 6461 Upper = *C; 6462 break; 6463 6464 default: 6465 break; 6466 } 6467 } 6468 6469 static void setLimitsForIntrinsic(const IntrinsicInst &II, APInt &Lower, 6470 APInt &Upper) { 6471 unsigned Width = Lower.getBitWidth(); 6472 const APInt *C; 6473 switch (II.getIntrinsicID()) { 6474 case Intrinsic::ctpop: 6475 case Intrinsic::ctlz: 6476 case Intrinsic::cttz: 6477 // Maximum of set/clear bits is the bit width. 6478 assert(Lower == 0 && "Expected lower bound to be zero"); 6479 Upper = Width + 1; 6480 break; 6481 case Intrinsic::uadd_sat: 6482 // uadd.sat(x, C) produces [C, UINT_MAX]. 6483 if (match(II.getOperand(0), m_APInt(C)) || 6484 match(II.getOperand(1), m_APInt(C))) 6485 Lower = *C; 6486 break; 6487 case Intrinsic::sadd_sat: 6488 if (match(II.getOperand(0), m_APInt(C)) || 6489 match(II.getOperand(1), m_APInt(C))) { 6490 if (C->isNegative()) { 6491 // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. 6492 Lower = APInt::getSignedMinValue(Width); 6493 Upper = APInt::getSignedMaxValue(Width) + *C + 1; 6494 } else { 6495 // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. 6496 Lower = APInt::getSignedMinValue(Width) + *C; 6497 Upper = APInt::getSignedMaxValue(Width) + 1; 6498 } 6499 } 6500 break; 6501 case Intrinsic::usub_sat: 6502 // usub.sat(C, x) produces [0, C]. 6503 if (match(II.getOperand(0), m_APInt(C))) 6504 Upper = *C + 1; 6505 // usub.sat(x, C) produces [0, UINT_MAX - C]. 6506 else if (match(II.getOperand(1), m_APInt(C))) 6507 Upper = APInt::getMaxValue(Width) - *C + 1; 6508 break; 6509 case Intrinsic::ssub_sat: 6510 if (match(II.getOperand(0), m_APInt(C))) { 6511 if (C->isNegative()) { 6512 // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. 6513 Lower = APInt::getSignedMinValue(Width); 6514 Upper = *C - APInt::getSignedMinValue(Width) + 1; 6515 } else { 6516 // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. 6517 Lower = *C - APInt::getSignedMaxValue(Width); 6518 Upper = APInt::getSignedMaxValue(Width) + 1; 6519 } 6520 } else if (match(II.getOperand(1), m_APInt(C))) { 6521 if (C->isNegative()) { 6522 // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: 6523 Lower = APInt::getSignedMinValue(Width) - *C; 6524 Upper = APInt::getSignedMaxValue(Width) + 1; 6525 } else { 6526 // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. 6527 Lower = APInt::getSignedMinValue(Width); 6528 Upper = APInt::getSignedMaxValue(Width) - *C + 1; 6529 } 6530 } 6531 break; 6532 case Intrinsic::umin: 6533 case Intrinsic::umax: 6534 case Intrinsic::smin: 6535 case Intrinsic::smax: 6536 if (!match(II.getOperand(0), m_APInt(C)) && 6537 !match(II.getOperand(1), m_APInt(C))) 6538 break; 6539 6540 switch (II.getIntrinsicID()) { 6541 case Intrinsic::umin: 6542 Upper = *C + 1; 6543 break; 6544 case Intrinsic::umax: 6545 Lower = *C; 6546 break; 6547 case Intrinsic::smin: 6548 Lower = APInt::getSignedMinValue(Width); 6549 Upper = *C + 1; 6550 break; 6551 case Intrinsic::smax: 6552 Lower = *C; 6553 Upper = APInt::getSignedMaxValue(Width) + 1; 6554 break; 6555 default: 6556 llvm_unreachable("Must be min/max intrinsic"); 6557 } 6558 break; 6559 case Intrinsic::abs: 6560 // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], 6561 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. 6562 if (match(II.getOperand(1), m_One())) 6563 Upper = APInt::getSignedMaxValue(Width) + 1; 6564 else 6565 Upper = APInt::getSignedMinValue(Width) + 1; 6566 break; 6567 default: 6568 break; 6569 } 6570 } 6571 6572 static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower, 6573 APInt &Upper, const InstrInfoQuery &IIQ) { 6574 const Value *LHS = nullptr, *RHS = nullptr; 6575 SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS); 6576 if (R.Flavor == SPF_UNKNOWN) 6577 return; 6578 6579 unsigned BitWidth = SI.getType()->getScalarSizeInBits(); 6580 6581 if (R.Flavor == SelectPatternFlavor::SPF_ABS) { 6582 // If the negation part of the abs (in RHS) has the NSW flag, 6583 // then the result of abs(X) is [0..SIGNED_MAX], 6584 // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. 6585 Lower = APInt::getNullValue(BitWidth); 6586 if (match(RHS, m_Neg(m_Specific(LHS))) && 6587 IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) 6588 Upper = APInt::getSignedMaxValue(BitWidth) + 1; 6589 else 6590 Upper = APInt::getSignedMinValue(BitWidth) + 1; 6591 return; 6592 } 6593 6594 if (R.Flavor == SelectPatternFlavor::SPF_NABS) { 6595 // The result of -abs(X) is <= 0. 6596 Lower = APInt::getSignedMinValue(BitWidth); 6597 Upper = APInt(BitWidth, 1); 6598 return; 6599 } 6600 6601 const APInt *C; 6602 if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C))) 6603 return; 6604 6605 switch (R.Flavor) { 6606 case SPF_UMIN: 6607 Upper = *C + 1; 6608 break; 6609 case SPF_UMAX: 6610 Lower = *C; 6611 break; 6612 case SPF_SMIN: 6613 Lower = APInt::getSignedMinValue(BitWidth); 6614 Upper = *C + 1; 6615 break; 6616 case SPF_SMAX: 6617 Lower = *C; 6618 Upper = APInt::getSignedMaxValue(BitWidth) + 1; 6619 break; 6620 default: 6621 break; 6622 } 6623 } 6624 6625 ConstantRange llvm::computeConstantRange(const Value *V, bool UseInstrInfo, 6626 AssumptionCache *AC, 6627 const Instruction *CtxI, 6628 unsigned Depth) { 6629 assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction"); 6630 6631 if (Depth == MaxAnalysisRecursionDepth) 6632 return ConstantRange::getFull(V->getType()->getScalarSizeInBits()); 6633 6634 const APInt *C; 6635 if (match(V, m_APInt(C))) 6636 return ConstantRange(*C); 6637 6638 InstrInfoQuery IIQ(UseInstrInfo); 6639 unsigned BitWidth = V->getType()->getScalarSizeInBits(); 6640 APInt Lower = APInt(BitWidth, 0); 6641 APInt Upper = APInt(BitWidth, 0); 6642 if (auto *BO = dyn_cast<BinaryOperator>(V)) 6643 setLimitsForBinOp(*BO, Lower, Upper, IIQ); 6644 else if (auto *II = dyn_cast<IntrinsicInst>(V)) 6645 setLimitsForIntrinsic(*II, Lower, Upper); 6646 else if (auto *SI = dyn_cast<SelectInst>(V)) 6647 setLimitsForSelectPattern(*SI, Lower, Upper, IIQ); 6648 6649 ConstantRange CR = ConstantRange::getNonEmpty(Lower, Upper); 6650 6651 if (auto *I = dyn_cast<Instruction>(V)) 6652 if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range)) 6653 CR = CR.intersectWith(getConstantRangeFromMetadata(*Range)); 6654 6655 if (CtxI && AC) { 6656 // Try to restrict the range based on information from assumptions. 6657 for (auto &AssumeVH : AC->assumptionsFor(V)) { 6658 if (!AssumeVH) 6659 continue; 6660 CallInst *I = cast<CallInst>(AssumeVH); 6661 assert(I->getParent()->getParent() == CtxI->getParent()->getParent() && 6662 "Got assumption for the wrong function!"); 6663 assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && 6664 "must be an assume intrinsic"); 6665 6666 if (!isValidAssumeForContext(I, CtxI, nullptr)) 6667 continue; 6668 Value *Arg = I->getArgOperand(0); 6669 ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); 6670 // Currently we just use information from comparisons. 6671 if (!Cmp || Cmp->getOperand(0) != V) 6672 continue; 6673 ConstantRange RHS = computeConstantRange(Cmp->getOperand(1), UseInstrInfo, 6674 AC, I, Depth + 1); 6675 CR = CR.intersectWith( 6676 ConstantRange::makeSatisfyingICmpRegion(Cmp->getPredicate(), RHS)); 6677 } 6678 } 6679 6680 return CR; 6681 } 6682 6683 static Optional<int64_t> 6684 getOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, const DataLayout &DL) { 6685 // Skip over the first indices. 6686 gep_type_iterator GTI = gep_type_begin(GEP); 6687 for (unsigned i = 1; i != Idx; ++i, ++GTI) 6688 /*skip along*/; 6689 6690 // Compute the offset implied by the rest of the indices. 6691 int64_t Offset = 0; 6692 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 6693 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 6694 if (!OpC) 6695 return None; 6696 if (OpC->isZero()) 6697 continue; // No offset. 6698 6699 // Handle struct indices, which add their field offset to the pointer. 6700 if (StructType *STy = GTI.getStructTypeOrNull()) { 6701 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 6702 continue; 6703 } 6704 6705 // Otherwise, we have a sequential type like an array or fixed-length 6706 // vector. Multiply the index by the ElementSize. 6707 TypeSize Size = DL.getTypeAllocSize(GTI.getIndexedType()); 6708 if (Size.isScalable()) 6709 return None; 6710 Offset += Size.getFixedSize() * OpC->getSExtValue(); 6711 } 6712 6713 return Offset; 6714 } 6715 6716 Optional<int64_t> llvm::isPointerOffset(const Value *Ptr1, const Value *Ptr2, 6717 const DataLayout &DL) { 6718 Ptr1 = Ptr1->stripPointerCasts(); 6719 Ptr2 = Ptr2->stripPointerCasts(); 6720 6721 // Handle the trivial case first. 6722 if (Ptr1 == Ptr2) { 6723 return 0; 6724 } 6725 6726 const GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); 6727 const GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); 6728 6729 // If one pointer is a GEP see if the GEP is a constant offset from the base, 6730 // as in "P" and "gep P, 1". 6731 // Also do this iteratively to handle the the following case: 6732 // Ptr_t1 = GEP Ptr1, c1 6733 // Ptr_t2 = GEP Ptr_t1, c2 6734 // Ptr2 = GEP Ptr_t2, c3 6735 // where we will return c1+c2+c3. 6736 // TODO: Handle the case when both Ptr1 and Ptr2 are GEPs of some common base 6737 // -- replace getOffsetFromBase with getOffsetAndBase, check that the bases 6738 // are the same, and return the difference between offsets. 6739 auto getOffsetFromBase = [&DL](const GEPOperator *GEP, 6740 const Value *Ptr) -> Optional<int64_t> { 6741 const GEPOperator *GEP_T = GEP; 6742 int64_t OffsetVal = 0; 6743 bool HasSameBase = false; 6744 while (GEP_T) { 6745 auto Offset = getOffsetFromIndex(GEP_T, 1, DL); 6746 if (!Offset) 6747 return None; 6748 OffsetVal += *Offset; 6749 auto Op0 = GEP_T->getOperand(0)->stripPointerCasts(); 6750 if (Op0 == Ptr) { 6751 HasSameBase = true; 6752 break; 6753 } 6754 GEP_T = dyn_cast<GEPOperator>(Op0); 6755 } 6756 if (!HasSameBase) 6757 return None; 6758 return OffsetVal; 6759 }; 6760 6761 if (GEP1) { 6762 auto Offset = getOffsetFromBase(GEP1, Ptr2); 6763 if (Offset) 6764 return -*Offset; 6765 } 6766 if (GEP2) { 6767 auto Offset = getOffsetFromBase(GEP2, Ptr1); 6768 if (Offset) 6769 return Offset; 6770 } 6771 6772 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 6773 // base. After that base, they may have some number of common (and 6774 // potentially variable) indices. After that they handle some constant 6775 // offset, which determines their offset from each other. At this point, we 6776 // handle no other case. 6777 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 6778 return None; 6779 6780 // Skip any common indices and track the GEP types. 6781 unsigned Idx = 1; 6782 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 6783 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 6784 break; 6785 6786 auto Offset1 = getOffsetFromIndex(GEP1, Idx, DL); 6787 auto Offset2 = getOffsetFromIndex(GEP2, Idx, DL); 6788 if (!Offset1 || !Offset2) 6789 return None; 6790 return *Offset2 - *Offset1; 6791 } 6792